Datasets:

AlgorithmicResearchGroup

/

arxiv_deep_learning_python_research_code

like 6

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/tools/merge_qrels.py

from utils import load_from_trec import argparse import torch import csv if __name__ == "__main__": parser = argparse.ArgumentParser() parser.add_argument("--scores_path", type=str) parser.add_argument("--qrels_path", type=str) parser.add_argument("--save_path", type=str) parser.add_argument("run") args = parser.parse_args() scores = torch.load(args.scores_path) print(scores.size()) run = load_from_trec(args.run, as_list=True) g = open(args.save_path, "w") qrels = {} with open(args.qrels_path, encoding="utf8") as f: tsvreader = csv.reader(f, delimiter="\t") for [qid, _, docid, rel] in tsvreader: assert rel == "1" if qid in qrels: qrels[qid].append(docid) else: qrels[qid] = [docid] id = 0 sum, overlap = 0, 0 for qid, rank_list in run.items(): docids = [] for doc_rank, (docid, _) in enumerate(rank_list): docids.append(docid) if len(docids) == 10: break sort_scores, sort_index = torch.sort(scores[id], descending=True) for docid in qrels[qid]: # pass g.write(f"{qid}\t0\t{docid}\t1\n") sum += len(qrels[qid]) for i in sort_index[:2]: if docids[i] not in qrels[qid]: # pass g.write(f"{qid}\t0\t{docids[i]}\t1\n") else: overlap += 1 id += 1 if id >= scores.size(0): break print(overlap, sum, overlap / sum)

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/tools/transform.py

# coding:utf-8 import torch import argparse import os import tqdm import copy def transform_new_model(model_hf, layer_num): model_new = {} cnt = 0 for i in range(layer_num): # encoder target_k = "encoder.blocks.{}.self_attn.self_attn.project.weight".format(i) source = [ "encoder.block.{}.layer.0.SelfAttention.q.weight".format(i), "encoder.block.{}.layer.0.SelfAttention.k.weight".format(i), "encoder.block.{}.layer.0.SelfAttention.v.weight".format(i), ] # qkv model_new[target_k] = torch.cat([model_hf[x] for x in source], 0) cnt += 3 target_k = "encoder.blocks.{}.self_attn.self_attn.dense.weight".format(i) source = "encoder.block.{}.layer.0.SelfAttention.o.weight".format(i) model_new[target_k] = model_hf[source] / 100 cnt += 1 target_k = "encoder.blocks.{}.self_attn.layer_norm.weight".format(i) source = "encoder.block.{}.layer.0.layer_norm.weight".format(i) model_new[target_k] = model_hf[source] cnt += 1 target_k = "encoder.blocks.{}.ff.dense_relu_dense.wi_0.weight".format(i) source = "encoder.block.{}.layer.1.DenseReluDense.wi_0.weight".format(i) model_new[target_k] = model_hf[source] cnt += 1 target_k = "encoder.blocks.{}.ff.dense_relu_dense.wi_1.weight".format(i) source = "encoder.block.{}.layer.1.DenseReluDense.wi_1.weight".format(i) model_new[target_k] = model_hf[source] / 10 cnt += 1 target_k = "encoder.blocks.{}.ff.dense_relu_dense.wo.weight".format(i) source = "encoder.block.{}.layer.1.DenseReluDense.wo.weight".format(i) model_new[target_k] = model_hf[source] / 10 cnt += 1 target_k = "encoder.blocks.{}.ff.layer_norm.weight".format(i) source = "encoder.block.{}.layer.1.layer_norm.weight".format(i) model_new[target_k] = model_hf[source] cnt += 1 # decoder target_k = "decoder.blocks.{}.self_attn.self_attn.project.weight".format(i) source = [ "decoder.block.{}.layer.0.SelfAttention.q.weight".format(i), "decoder.block.{}.layer.0.SelfAttention.k.weight".format(i), "decoder.block.{}.layer.0.SelfAttention.v.weight".format(i), ] # qkv model_new[target_k] = torch.cat([model_hf[x] for x in source], 0) cnt += 3 target_k = "decoder.blocks.{}.cross_attn.cross_attn.project_kv.weight".format(i) source = [ "decoder.block.{}.layer.1.EncDecAttention.k.weight".format(i), "decoder.block.{}.layer.1.EncDecAttention.v.weight".format(i), ] # kv model_new[target_k] = torch.cat([model_hf[x] for x in source], 0) cnt += 2 target_k = "decoder.blocks.{}.cross_attn.cross_attn.project_q.weight".format(i) source = "decoder.block.{}.layer.1.EncDecAttention.q.weight".format(i) model_new[target_k] = model_hf[source] cnt += 1 target_k = "decoder.blocks.{}.cross_attn.cross_attn.dense.weight".format(i) source = "decoder.block.{}.layer.1.EncDecAttention.o.weight".format(i) model_new[target_k] = model_hf[source] / 100 cnt += 1 target_k = "decoder.blocks.{}.cross_attn.layer_norm.weight".format(i) source = "decoder.block.{}.layer.1.layer_norm.weight".format(i) model_new[target_k] = model_hf[source] cnt += 1 target_k = "decoder.blocks.{}.self_attn.self_attn.dense.weight".format(i) source = "decoder.block.{}.layer.0.SelfAttention.o.weight".format(i) model_new[target_k] = model_hf[source] / 100 cnt += 1 target_k = "decoder.blocks.{}.self_attn.layer_norm.weight".format(i) source = "decoder.block.{}.layer.0.layer_norm.weight".format(i) model_new[target_k] = model_hf[source] cnt += 1 target_k = "decoder.blocks.{}.ff.dense_relu_dense.wi_0.weight".format(i) source = "decoder.block.{}.layer.2.DenseReluDense.wi_0.weight".format(i) model_new[target_k] = model_hf[source] cnt += 1 target_k = "decoder.blocks.{}.ff.dense_relu_dense.wi_1.weight".format(i) source = "decoder.block.{}.layer.2.DenseReluDense.wi_1.weight".format(i) model_new[target_k] = model_hf[source] / 10 cnt += 1 target_k = "decoder.blocks.{}.ff.dense_relu_dense.wo.weight".format(i) source = "decoder.block.{}.layer.2.DenseReluDense.wo.weight".format(i) model_new[target_k] = model_hf[source] / 10 cnt += 1 target_k = "decoder.blocks.{}.ff.layer_norm.weight".format(i) source = "decoder.block.{}.layer.2.layer_norm.weight".format(i) model_new[target_k] = model_hf[source] cnt += 1 source = "shared.weight" target_k = "word_embeds.weight" embeds = model_hf[source] model_new[target_k] = embeds / 100 target_k = "encoder.word_embeds.weight" model_new[target_k] = embeds / 100 target_k = "decoder.word_embeds.weight" model_new[target_k] = embeds / 100 cnt += 3 source = "encoder.block.0.layer.0.SelfAttention.relative_attention_bias.weight" target_k = "encoder.blocks.0.self_attn.self_attn.relative_attention_bias.weight" model_new[target_k] = model_hf[source] cnt += 1 source = "decoder.block.0.layer.0.SelfAttention.relative_attention_bias.weight" target_k = "decoder.blocks.0.self_attn.self_attn.relative_attention_bias.weight" model_new[target_k] = model_hf[source] cnt += 1 source = "lm_head.weight" target_k = "lm_head.weight" embeds = model_hf[source] model_new[target_k] = embeds cnt += 1 source = "encoder.final_layer_norm.weight" target_k = "encoder.final_layernorm.weight" model_new[target_k] = model_hf[source] cnt += 1 source = "decoder.final_layer_norm.weight" target_k = "decoder.final_layernorm.weight" model_new[target_k] = model_hf[source] cnt += 1 print("new module number:", cnt, "origin module number:", len(model_hf)) return {"module": model_new} def change_mp(d, output_dir, mp_size, half=False): os.makedirs(output_dir, exist_ok=True) os.makedirs(os.path.join(output_dir, "1"), exist_ok=True) with open(os.path.join(output_dir, "latest_checkpointed_iteration.txt"), "w") as f: f.write(str(1) + "\n") preserve_keys = [ "lr_scheduler", "skipped_steps", "global_steps", "global_samples", "dp_world_size", "iteration", ] dd = {} dd["lr_scheduler"] = {} dd["lr_scheduler"]["num_iters"] = 1 dd["lr_scheduler"]["start_lr"] = 0.001 dd["lr_scheduler"]["warmup_iter"] = 10000 dd["skipped_steps"] = 0 dd["global_steps"] = 1 dd["global_samples"] = 100 dd["iteration"] = 1 dd["dp_world_size"] = 1 print("Increase MP size.") ratio = mp_size start = 0 end = ratio for j in tqdm.tqdm(range(start, end)): d_new = {} shift = j - start for k, v in dd.items(): if k != "module": if k in preserve_keys: d_new[k] = copy.deepcopy(dd[k]) elif k == "mp_world_size": d_new[k] = ratio else: d_new[k] = None d_new["module"] = {} for k, v in d["module"].items(): assert len(v.shape) < 3 if len(v.shape) == 2: if "project.weight" in k: part = v.shape[0] // ratio // 3 d_new["module"][k] = torch.cat( [ v[shift * part : (shift + 1) * part, :], v[(shift + ratio) * part : (shift + 1 + ratio) * part, :], v[ (shift + 2 * ratio) * part : (shift + 1 + 2 * ratio) * part, :, ], ], 0, ) elif "project_q.weight" in k: part = v.shape[0] // ratio d_new["module"][k] = v[shift * part : (shift + 1) * part, :] elif "project_kv.weight" in k: part = v.shape[0] // ratio // 2 d_new["module"][k] = torch.cat( [ v[shift * part : (shift + 1) * part, :], v[(shift + ratio) * part : (shift + 1 + ratio) * part, :], ], 0, ) elif ( "word_embeds.weight" in k or "dense_relu_dense.wi_1.weight" in k or "dense_relu_dense.wi_0.weight" in k or "lm_head.weight" in k ): part = v.shape[0] // ratio d_new["module"][k] = v[shift * part : (shift + 1) * part, :] else: part = v.shape[1] // ratio d_new["module"][k] = v[:, shift * part : (shift + 1) * part] else: d_new["module"][k] = v if half: d_new["module"][k] = d_new["module"][k].half() filename = os.path.join( output_dir, "1", "mp_rank_0{}_model_states.pt".format(j) ) torch.save(d_new, filename) def main(): parser = argparse.ArgumentParser( "Transform huggingface checkpoints to megatron+deepspeed checkpoints" ) parser.add_argument("--hf_path", type=str) parser.add_argument("--ext_path", type=str, default="") parser.add_argument("--mp_size", type=int, default=1) parser.add_argument("--save_path", type=str) parser.add_argument("--half", action="store_true") args = parser.parse_args() model_hf = torch.load(args.hf_path, map_location="cpu") if args.ext_path: model_ext = torch.load(args.ext_path, map_location="cpu") model_hf.update(model_ext) print(len(model_hf)) new_model = transform_new_model(model_hf, 12 if "base" in args.save_path else 24) change_mp(new_model, args.save_path, args.mp_size, half=args.half) if __name__ == "__main__": main()

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/tools/utils.py

# Adapted from Tevatron (https://github.com/texttron/tevatron) import csv import json import warnings from dataclasses import dataclass from typing import Dict, List import datasets import torch from transformers import PreTrainedTokenizer @dataclass class SimpleTrainPreProcessor: query_file: str collection_file: str tokenizer: PreTrainedTokenizer doc_max_len: int = 128 query_max_len: int = 32 columns = ['text_id', 'title', 'text'] title_field = 'title' text_field = 'text' query_field = 'text' doc_template: str = None query_template: str = None allow_not_found: bool = False def __post_init__(self): self.queries = self.read_queries(self.query_file) self.collection = datasets.load_dataset( 'csv', data_files=self.collection_file, column_names=self.columns, delimiter='\t', )['train'] @staticmethod def read_queries(queries): qmap = {} with open(queries) as f: for l in f: qid, qry = l.strip().split('\t') qmap[qid] = qry return qmap @staticmethod def read_qrel(relevance_file): qrel = {} with open(relevance_file, encoding='utf8') as f: tsvreader = csv.reader(f, delimiter="\t") for [topicid, _, docid, rel] in tsvreader: assert rel == "1" if topicid in qrel: qrel[topicid].append(docid) else: qrel[topicid] = [docid] return qrel def get_query(self, q): if self.query_template is None: query = self.queries[q] else: query = fill_template(self.query_template, data={self.query_field: self.queries[q]}, allow_not_found=self.allow_not_found) query_encoded = self.tokenizer.encode( query, add_special_tokens=False, max_length=self.query_max_len, truncation=True ) return query_encoded def get_passage(self, p): entry = self.collection[int(p)] title = entry[self.title_field] title = "" if title is None else title body = entry[self.text_field] if self.doc_template is None: content = title + self.tokenizer.sep_token + body else: content = fill_template(self.doc_template, data=entry, allow_not_found=self.allow_not_found) passage_encoded = self.tokenizer.encode( content, add_special_tokens=False, max_length=self.doc_max_len, truncation=True ) return passage_encoded def process_one(self, train): q, pp, nn = train train_example = { 'query': self.get_query(q), 'positives': [self.get_passage(p) for p in pp], 'negatives': [self.get_passage(n) for n in nn], } return json.dumps(train_example) @dataclass class SimpleCollectionPreProcessor: tokenizer: PreTrainedTokenizer separator: str = '\t' max_length: int = 128 def process_line(self, line: str): xx = line.strip().split(self.separator) text_id, text = xx[0], xx[1:] text_encoded = self.tokenizer.encode( self.tokenizer.sep_token.join(text), add_special_tokens=False, max_length=self.max_length, truncation=True ) encoded = { 'text_id': text_id, 'text': text_encoded } return json.dumps(encoded) def save_as_trec(rank_result: Dict[str, Dict[str, float]], output_path: str, run_id: str = "OpenMatch"): """ Save the rank result as TREC format: <query_id> Q0 <doc_id> <rank> <score> <run_id> """ with open(output_path, "w") as f: for qid in rank_result: # sort the results by score sorted_results = sorted(rank_result[qid].items(), key=lambda x: x[1], reverse=True) for i, (doc_id, score) in enumerate(sorted_results): f.write("{} Q0 {} {} {} {}\n".format(qid, doc_id, i + 1, score, run_id)) def load_from_trec(input_path: str, as_list: bool = False, max_len_per_q: int = None): """ Load the rank result from TREC format: <query_id> Q0 <doc_id> <rank> <score> <run_id> or <query_id> <doc_id> <score> """ rank_result = {} cnt = 0 with open(input_path, "r") as f: for line in f: content = line.strip().split() if len(content) == 6: qid, _, doc_id, _, score, _ = content elif len(content) == 3: qid, doc_id, score = content else: raise ValueError("Invalid run format") if not as_list: if qid not in rank_result: rank_result[qid] = {} cnt = 0 if max_len_per_q is None or cnt < max_len_per_q: rank_result[qid][doc_id] = float(score) else: if qid not in rank_result: rank_result[qid] = [] cnt = 0 if max_len_per_q is None or cnt < max_len_per_q: rank_result[qid].append((doc_id, float(score))) cnt += 1 return rank_result def find_all_markers(template: str): """ Find all markers' names (quoted in "<>") in a template. """ markers = [] start = 0 while True: start = template.find("<", start) if start == -1: break end = template.find(">", start) if end == -1: break markers.append(template[start + 1:end]) start = end + 1 return markers def fill_template(template: str, data: Dict, markers: List[str] = None, allow_not_found: bool = False): """ Fill a template with data. """ if markers is None: markers = find_all_markers(template) for marker in markers: marker_hierarchy = marker.split(".") found = True content = data for marker_level in marker_hierarchy: content = content.get(marker_level, None) if content is None: found = False break if not found: if allow_not_found: warnings.warn("Marker '{}' not found in data. Replacing it with an empty string.".format(marker), RuntimeWarning) content = "" else: raise ValueError("Cannot find the marker '{}' in the data".format(marker)) template = template.replace("<{}>".format(marker), str(content)) return template def merge_retrieval_results_by_score(results: List[Dict[str, Dict[str, float]]], topk: int = 100): """ Merge retrieval results from multiple partitions of document embeddings and keep topk. """ merged_results = {} for result in results: for qid in result: if qid not in merged_results: merged_results[qid] = {} for doc_id in result[qid]: if doc_id not in merged_results[qid]: merged_results[qid][doc_id] = result[qid][doc_id] for qid in merged_results: merged_results[qid] = {k: v for k, v in sorted(merged_results[qid].items(), key=lambda x: x[1], reverse=True)[:topk]} return merged_results # Mean Pooling - Take attention mask into account for correct averaging def mean_pooling(token_embeddings, attention_mask): input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float() return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/tools/ds_fix/engine.py

''' Copyright 2019 The Microsoft DeepSpeed Team ''' import os import time import torch import warnings import torch.distributed as dist from torch.nn.modules import Module from torch.distributed.distributed_c10d import _get_global_rank from tensorboardX import SummaryWriter from deepspeed.runtime.zero.stage2 import FP16_DeepSpeedZeroOptimizer from deepspeed.runtime.zero.stage1 import FP16_DeepSpeedZeroOptimizer_Stage1 from deepspeed.runtime.zero.utils import is_zero_supported_optimizer from deepspeed.runtime.activation_checkpointing import checkpointing as activation_checkpointing from deepspeed.runtime.fp16.fused_optimizer import FP16_Optimizer from deepspeed.runtime.fp16.unfused_optimizer import FP16_UnfusedOptimizer from deepspeed.runtime.config import DeepSpeedConfig, DEEPSPEED_OPTIMIZERS, \ ADAM_OPTIMIZER, LAMB_OPTIMIZER, ONEBIT_ADAM_OPTIMIZER, \ TORCH_ADAM_PARAM, ADAM_W_MODE_PARAM from deepspeed.runtime.dataloader import DeepSpeedDataLoader from deepspeed.runtime.constants import \ ROUTE_TRAIN, ROUTE_PREDICT, ROUTE_EVAL, \ PLD_THETA, PLD_GAMMA from deepspeed.runtime.zero.constants import \ ZERO_OPTIMIZATION_OPTIMIZER_STATES, ZERO_OPTIMIZATION_GRADIENTS from deepspeed.runtime.csr_tensor import CSRTensor import deepspeed.runtime.lr_schedules as lr_schedules from deepspeed.utils import logger, log_dist, init_distributed from deepspeed.utils.timer import ThroughputTimer, SynchronizedWallClockTimer from deepspeed.runtime.progressive_layer_drop import ProgressiveLayerDrop from .pipe.module import PipelineModule from .utils import ensure_directory_exists from ..ops.op_builder import UtilsBuilder from ..ops.adam import DeepSpeedCPUAdam from ..ops.adam import FusedAdam MEMORY_OPT_ALLREDUCE_SIZE = 500000000 try: from apex import amp except ImportError: # Fail silently so we don't spam logs unnecessarily if user isn't using amp pass def split_half_float_double_csr(tensors): dtypes = [ "torch.cuda.HalfTensor", "torch.cuda.FloatTensor", "torch.cuda.DoubleTensor", CSRTensor.type() ] buckets = [] for i, dtype in enumerate(dtypes): bucket = [t for t in tensors if t.type() == dtype] if bucket: buckets.append((dtype, bucket)) return buckets def _initialize_parameter_parallel_groups(parameter_parallel_size=None): data_parallel_size = int(dist.get_world_size()) if parameter_parallel_size is None: parameter_parallel_size = int(data_parallel_size) logger.info("data_parallel_size: %s, parameter_parallel_size: %s", data_parallel_size, parameter_parallel_size) assert data_parallel_size % parameter_parallel_size == 0, \ 'world size should be divisible by parameter parallel size' rank = dist.get_rank() my_group = None for i in range(dist.get_world_size() // parameter_parallel_size): ranks = range(i * parameter_parallel_size, (i + 1) * parameter_parallel_size) group = torch.distributed.new_group(ranks) if rank in ranks: my_group = group return my_group def print_configuration(args, name): logger.info('{}:'.format(name)) for arg in sorted(vars(args)): dots = '.' * (29 - len(arg)) logger.info(' {} {} {}'.format(arg, dots, getattr(args, arg))) class DeepSpeedEngine(Module): r"""DeepSpeed engine for training. """ def __init__(self, args, model, optimizer=None, model_parameters=None, training_data=None, lr_scheduler=None, mpu=None, dist_init_required=None, collate_fn=None, config_params=None): super(DeepSpeedEngine, self).__init__() self.client_optimizer = optimizer self.client_model_parameters = model_parameters self.client_lr_scheduler = lr_scheduler self.training_data = training_data self.collate_fn = collate_fn self.mpu = mpu self.data_parallel_group = None self.global_steps = 0 self.global_samples = 0 self.micro_steps = 0 self.skipped_steps = 0 self.gradient_average = True self.warn_unscaled_loss = True self.config_params = config_params self.loaded_checkpoint_mp_world_size = None self.loaded_checkpoint_dp_world_size = None self.enable_backward_allreduce = True self.progressive_layer_drop = None self.dist_backend = "nccl" if dist_init_required is None: dist_init_required = not dist.is_initialized() if dist_init_required is False: assert (dist.is_initialized()==True), "Torch distributed not initialized. Please set dist_init_required to True or initialize before calling deepspeed.initialize()" # Initialize torch distributed if needed init_distributed(dist_backend=self.dist_backend) self._do_args_sanity_check(args) self._configure_with_arguments(args, mpu) self._do_sanity_check() if mpu is not None: assert not self.elasticity_enabled(), "Elasticity is not currently supported" \ " with model parallelism." self._set_distributed_vars() if self.tensorboard_enabled() and self.global_rank == 0: self.summary_writer = self.get_summary_writer() # Configure distributed model self._configure_distributed_model(model) # Configure wall clock timer self.timers = SynchronizedWallClockTimer() # Throughput timer self.tput_timer = ThroughputTimer( batch_size=self.train_micro_batch_size_per_gpu(), num_workers=self.dp_world_size, steps_per_output=self.steps_per_print(), monitor_memory=False) if training_data: self.training_dataloader = self.deepspeed_io(training_data) else: self.training_dataloader = None # Configure optimizer and scheduler self.optimizer = None self.lr_scheduler = None if model_parameters or optimizer: self._configure_optimizer(optimizer, model_parameters) self._configure_lr_scheduler(lr_scheduler) self._report_progress(0) # Bookkeeping for csr support self.csr_tensor_module_names = set() if self.sparse_gradients_enabled(): for name, module in self.module.named_modules(): if isinstance(module, torch.nn.Embedding): self.csr_tensor_module_names.add(name + ".weight") logger.info("Will convert {} to sparse (csr) " "tensor during training".format(name)) self.save_non_zero_checkpoint = False self.save_zero_checkpoint = False self._configure_checkpointing(dist_init_required) if self.pld_enabled(): self.progressive_layer_drop = self._configure_progressive_layer_drop() if self.global_rank == 0: self._config.print('DeepSpeedEngine configuration') if self.dump_state(): print_configuration(self, 'DeepSpeedEngine') # Load pre-installed or JIT compile (un)flatten ops util_ops = UtilsBuilder().load() self.flatten = util_ops.flatten self.unflatten = util_ops.unflatten def get_batch_info(self): """ Get all training batch related settings. Returns: train_batch_size (int): The effective training batch size. This is the amount of data samples that leads to one step of model update. train_micro_batch_size_per_gpu (int): Batch size to be processed by one GPU in one step (without gradient accumulation). gradient_accumulation_steps (int): Number of training steps to accumulate gradients before averaging and applying them. """ return self.train_batch_size, self.train_micro_batch_size_per_gpu, self.gradient_accumulation_steps def elasticity_enabled(self): return self._config.elasticity_enabled def pld_enabled(self): return self._config.pld_enabled def pld_params(self): return self._config.pld_params def pld_theta(self): return self.pld_params()[PLD_THETA] def pld_gamma(self): return self.pld_params()[PLD_GAMMA] def tensorboard_enabled(self): return self._config.tensorboard_enabled def tensorboard_output_path(self): return self._config.tensorboard_output_path def tensorboard_job_name(self): return self._config.tensorboard_job_name def get_summary_writer(self, name="DeepSpeedJobName", base=os.path.join(os.environ["HOME"], "tensorboard")): if self.tensorboard_output_path(): base_dir = self.tensorboard_output_path() job_name = self.tensorboard_job_name() log_dir = os.path.join(base_dir, job_name) else: if self.tensorboard_job_name(): name = self.tensorboard_job_name() # Infrastructure-specific job-id if 'DLWS_JOB_ID' in os.environ: infra_job_id = os.environ['DLWS_JOB_ID'] elif 'DLTS_JOB_ID' in os.environ: infra_job_id = os.environ['DLTS_JOB_ID'] else: infra_job_id = 'unknown-job-id' summary_writer_dir_name = os.path.join(infra_job_id, "logs") log_dir = os.path.join(base, summary_writer_dir_name, name) os.makedirs(log_dir, exist_ok=True) return SummaryWriter(log_dir=log_dir) def wall_clock_breakdown(self): return self._config.wall_clock_breakdown def memory_breakdown(self): return self._config.memory_breakdown def sparse_gradients_enabled(self): return self._config.sparse_gradients_enabled def train_batch_size(self): return self._config.train_batch_size def train_micro_batch_size_per_gpu(self): return self._config.train_micro_batch_size_per_gpu def optimizer_name(self): return self.client_optimizer.__class__.__name__ if self.client_optimizer else self._config.optimizer_name def optimizer_params(self): return self._config.optimizer_params def optimizer_legacy_fusion(self): return self._config.optimizer_legacy_fusion def scheduler_name(self): return self._config.scheduler_name def scheduler_params(self): return self._config.scheduler_params def zero_optimization(self): return self._config.zero_enabled def zero_allow_untested_optimizer(self): return self._config.zero_allow_untested_optimizer def zero_reduce_scatter(self): return self._config.zero_config.reduce_scatter def zero_overlap_comm(self): return self._config.zero_config.overlap_comm def zero_cpu_offload(self): return self._config.zero_config.cpu_offload def zero_optimization_stage(self): return self._config.zero_optimization_stage def zero_reduce_bucket_size(self): return self._config.zero_config.reduce_bucket_size def zero_allgather_bucket_size(self): return self._config.zero_config.allgather_bucket_size def zero_optimization_partition_gradients(self): return self.zero_optimization_stage() >= ZERO_OPTIMIZATION_GRADIENTS def zero_contiguous_gradients(self): return self._config.zero_config.contiguous_gradients def zero_load_from_fp32_weights(self): return self._config.zero_config.load_from_fp32_weights def zero_elastic_checkpoint(self): return self._config.zero_config.elastic_checkpoint def fp16_enabled(self): return self._config.fp16_enabled def amp_enabled(self): return self._config.amp_enabled def amp_params(self): return self._config.amp_params def loss_scale(self): return self._config.loss_scale def gradient_accumulation_steps(self): return self._config.gradient_accumulation_steps def allreduce_always_fp32(self): return self._config.allreduce_always_fp32 def postscale_gradients(self): return not self._config.prescale_gradients def gradient_predivide_factor(self): return self._config.gradient_predivide_factor def steps_per_print(self): return self._config.steps_per_print def zero_allgather_partitions(self): return self._config.zero_config.allgather_partitions def dump_state(self): return self._config.dump_state def gradient_clipping(self): return self._config.gradient_clipping def dynamic_loss_scale(self): return self._config.loss_scale == 0 def initial_dynamic_scale(self): return self._config.initial_dynamic_scale def dynamic_loss_scale_args(self): return self._config.dynamic_loss_scale_args def _configure_lr_scheduler(self, client_lr_scheduler): # First check for scheduler in json configuration lr_scheduler = self._scheduler_from_config(self.optimizer) if lr_scheduler: if self.global_rank == 0: logger.info( f'DeepSpeed using configured LR scheduler = {self.scheduler_name()}') self.lr_scheduler = lr_scheduler else: if self.global_rank == 0: logger.info('DeepSpeed using client LR scheduler') self.lr_scheduler = client_lr_scheduler log_dist(f'DeepSpeed LR Scheduler = {self.lr_scheduler}', ranks=[0]) def _configure_checkpointing(self, dist_init_required): dp_rank = self.global_rank if self.mpu: dp_rank = self.mpu.get_data_parallel_rank() # only the first data parallel process needs to store the model checkpoint self.save_non_zero_checkpoint = (dp_rank == 0) if self.zero_optimization() and self.optimizer is not None: param_rank = torch.distributed.get_rank( group=self.optimizer.dp_process_group) # Only the first parameter parallel process needs to store the # optimizer state checkpoints for zero self.save_zero_checkpoint = (param_rank == dp_rank) def _scheduler_from_config(self, optimizer): scheduler_name = self.scheduler_name() if scheduler_name is not None: if hasattr(lr_schedules, scheduler_name): scheduler = getattr(lr_schedules, scheduler_name) else: assert hasattr(torch.optim.lr_scheduler, scheduler_name), \ f"DeepSpeed does not recognize LR scheduler {scheduler_name}" scheduler = getattr(torch.optim.lr_scheduler, scheduler_name) scheduler_params = self.scheduler_params() instantiated_scheduler = scheduler(optimizer, **scheduler_params) return instantiated_scheduler else: return None def _set_distributed_vars(self): if self.local_rank >= 0: torch.cuda.set_device(self.local_rank) self.device = torch.device("cuda", self.local_rank) self.world_size = dist.get_world_size() self.global_rank = dist.get_rank() else: self.world_size = 1 self.global_rank = 0 self.device = torch.device("cuda") # Configure based on command line arguments def _configure_with_arguments(self, args, mpu): self.local_rank = args.local_rank if hasattr(args, 'local_rank') else 0 config_file = args.deepspeed_config if hasattr(args, 'deepspeed_config') else None self._config = DeepSpeedConfig(config_file, mpu, param_dict=self.config_params) # Validate command line arguments def _do_args_sanity_check(self, args): if hasattr(args, 'deepscale_config') and args.deepscale_config is not None: logger.warning( "************ --deepscale_config is deprecated, please use --deepspeed_config ************" ) if hasattr(args, 'deepspeed_config'): assert args.deepspeed_config is None, "Not sure how to proceed, we were given both a deepscale_config and deepspeed_config" args.deepspeed_config = args.deepscale_config assert hasattr(args, 'local_rank') and type(args.local_rank) == int, \ 'DeepSpeed requires integer command line parameter --local_rank' if self.config_params is None: assert hasattr(args, 'deepspeed_config') and args.deepspeed_config is not None, \ 'DeepSpeed requires --deepspeed_config to specify configuration file' assert os.path.isfile(args.deepspeed_config), \ 'DeepSpeed configuration file: {} is not an existing file'.format(args.deepspeed_config) def _is_supported_optimizer(self, optimizer_name): return optimizer_name in DEEPSPEED_OPTIMIZERS or \ getattr(torch.optim, optimizer_name, None) is not None # Validate configuration based on command line arguments def _do_sanity_check(self): if not self.client_optimizer: if self.optimizer_name() is not None: assert self._is_supported_optimizer(self.optimizer_name()), \ '{} is not a supported DeepSpeed Optimizer'.format(self.optimizer_name()) if self.optimizer_name() == LAMB_OPTIMIZER: assert self.dynamic_loss_scale(), \ 'DeepSpeed {} optimizer requires dynamic loss scaling'.format(self.optimizer_name()) def _broadcast_model(self): for p in self.module.parameters(): if torch.is_tensor(p): dist.broadcast(p, self.broadcast_src_rank, group=self.data_parallel_group) def _configure_distributed_model(self, model): self.module = model if self.fp16_enabled(): self.module.half() self.module.to(self.device) if self.mpu is None: self.data_parallel_group = _initialize_parameter_parallel_groups() self.dp_world_size = dist.get_world_size() self.mp_world_size = 1 self.broadcast_src_rank = 0 else: self.data_parallel_group = self.mpu.get_data_parallel_group() self.dp_world_size = self.mpu.get_data_parallel_world_size() self.mp_world_size = self.mpu.get_model_parallel_world_size() self.broadcast_src_rank = _get_global_rank( self.mpu.get_data_parallel_group(), 0) if not self.amp_enabled(): self._broadcast_model() # Configure optimizer def _configure_optimizer(self, client_optimizer, model_parameters): if client_optimizer is not None: basic_optimizer = client_optimizer if self.global_rank == 0: logger.info('Using client Optimizer as basic optimizer') else: basic_optimizer = self._configure_basic_optimizer(model_parameters) if self.global_rank == 0: logger.info( 'Using DeepSpeed Optimizer param name {} as basic optimizer'.format( self.optimizer_name())) if self.global_rank == 0: logger.info('DeepSpeed Basic Optimizer = {}'.format(basic_optimizer)) if self.zero_optimization(): assert not self.amp_enabled(), "Amp and ZeRO are not currently compatible, please use (legacy) fp16 mode which performs similar to amp opt_mode=O2" if not is_zero_supported_optimizer(basic_optimizer): assert self.zero_allow_untested_optimizer(), \ 'You are using an untested ZeRO Optimizer. Please add <"zero_allow_untested_optimizer": true> in the configuration file to use it.' if self.global_rank == 0: logger.warning( "**** You are using ZeRO with an untested optimizer, proceed with caution *****" ) self.optimizer = self._configure_zero_optimizer(basic_optimizer) elif self.amp_enabled(): assert not self.fp16_enabled(), "Cannot enable both amp with (legacy) fp16 mode" amp_params = self.amp_params() if self.global_rank == 0: logger.info(f"Initializing AMP with these params: {amp_params}") try: logger.info("Initializing Apex amp from: {}".format(amp.__path__)) except NameError: # If apex/amp is available it will be imported above raise RuntimeError( "Unable to import apex/amp, please make sure it is installed") self.module, self.optimizer = amp.initialize(self.module, basic_optimizer, **amp_params) self._broadcast_model() elif self.fp16_enabled(): self.optimizer = self._configure_fp16_optimizer(basic_optimizer) else: self.optimizer = basic_optimizer logger.info('DeepSpeed Final Optimizer = {}'.format(self.optimizer)) logger.info('DeepSpeed Final Optimizer = {}'.format(self.optimizer.state_dict())) def _configure_basic_optimizer(self, model_parameters): optimizer_parameters = self.optimizer_params() # print(optimizer_parameters.keys()) if 'max_grad_norm' in optimizer_parameters.keys(): raise ValueError( "'max_grad_norm' is not supported as an optimizer parameter, please switch to using the deepspeed parameter 'gradient_clipping' see: https://www.deepspeed.ai/docs/config-json/#gradient-clipping for more details" ) if self.optimizer_name() == ADAM_OPTIMIZER: torch_adam = optimizer_parameters.pop(TORCH_ADAM_PARAM, False) adam_w_mode = optimizer_parameters.pop(ADAM_W_MODE_PARAM, True) # zero-offload torch-adam adam_w_mode optimizer # T|F T T torch.optim.AdamW # T|F T F torch.optim.Adam # T F T|F DeepSpeedCPUAdam(adam_w_mode) # F F T|F FusedAdam(adam_w_mode) if torch_adam: if adam_w_mode: optimizer = torch.optim.AdamW(model_parameters, **optimizer_parameters) else: optimizer = torch.optim.Adam(model_parameters, **optimizer_parameters) elif self.zero_cpu_offload(): optimizer = DeepSpeedCPUAdam(model_parameters, **optimizer_parameters, adamw_mode=adam_w_mode) else: optimizer_parameters[ADAM_W_MODE_PARAM] = adam_w_mode optimizer = FusedAdam(model_parameters, **optimizer_parameters) elif self.optimizer_name() == LAMB_OPTIMIZER: from deepspeed.ops.lamb import FusedLamb optimizer = FusedLamb(model_parameters, **optimizer_parameters) elif self.optimizer_name() == ONEBIT_ADAM_OPTIMIZER: from deepspeed.runtime.fp16.onebit_adam import OnebitAdam optimizer = OnebitAdam(model_parameters, self, **optimizer_parameters) else: torch_optimizer = getattr(torch.optim, self.optimizer_name()) optimizer = torch_optimizer(model_parameters, **optimizer_parameters) return optimizer def _configure_fp16_optimizer(self, optimizer): initial_dynamic_scale = self.initial_dynamic_scale() dynamic_loss_args = self.dynamic_loss_scale_args() clip_grad = self.gradient_clipping() if isinstance(optimizer, FusedAdam) or self.optimizer_name() == ONEBIT_ADAM_OPTIMIZER: if self.dynamic_loss_scale(): logger.info('Creating fp16 optimizer with dynamic loss scale') timers = self.timers if self.wall_clock_breakdown() else None optimizer = FP16_Optimizer( optimizer, dynamic_loss_scale=True, initial_dynamic_scale=initial_dynamic_scale, dynamic_loss_args=dynamic_loss_args, mpu=self.mpu, clip_grad=clip_grad, fused_adam_legacy=self.optimizer_legacy_fusion(), timers=timers) else: logger.info('Creating fp16 optimizer with static loss scale: {}'.format( self.loss_scale())) optimizer = FP16_Optimizer( optimizer, static_loss_scale=self.loss_scale(), mpu=self.mpu, clip_grad=clip_grad, fused_adam_legacy=self.optimizer_legacy_fusion()) else: logger.info('Creating fp16 unfused optimizer with dynamic loss scale') optimizer = FP16_UnfusedOptimizer( optimizer, static_loss_scale=self.loss_scale(), dynamic_loss_scale=self.dynamic_loss_scale(), dynamic_loss_args=dynamic_loss_args, mpu=self.mpu, clip_grad=clip_grad, fused_lamb_legacy=self.optimizer_name() == LAMB_OPTIMIZER) return optimizer def _configure_zero_optimizer(self, optimizer): zero_stage = self.zero_optimization_stage() logger.info('Creating fp16 ZeRO stage {} optimizer'.format(zero_stage)) assert not self.allreduce_always_fp32(), "ZeRO does not support 'fp32_allreduce': true" if zero_stage == ZERO_OPTIMIZATION_OPTIMIZER_STATES: assert self.zero_reduce_scatter(), 'Stage 1 only supports reduce scatter mode' optimizer = FP16_DeepSpeedZeroOptimizer_Stage1( optimizer, static_loss_scale=self.loss_scale(), dynamic_loss_scale=self.dynamic_loss_scale(), dynamic_loss_args=self.dynamic_loss_scale_args(), clip_grad=self.gradient_clipping(), all_gather_partitions=self.zero_allgather_partitions(), allgather_size=self.zero_allgather_bucket_size(), max_elements_per_comm=self.zero_reduce_bucket_size(), dp_process_group=self.data_parallel_group, elastic_checkpoint=self.zero_elastic_checkpoint(), mpu=self.mpu) elif zero_stage == ZERO_OPTIMIZATION_GRADIENTS: optimizer = FP16_DeepSpeedZeroOptimizer( optimizer, timers=self.timers, static_loss_scale=self.loss_scale(), dynamic_loss_scale=self.dynamic_loss_scale(), dynamic_loss_args=self.dynamic_loss_scale_args(), clip_grad=self.gradient_clipping(), contiguous_gradients=self.zero_contiguous_gradients(), reduce_bucket_size=self.zero_reduce_bucket_size(), allgather_bucket_size=self.zero_allgather_bucket_size(), dp_process_group=self.data_parallel_group, reduce_scatter=self.zero_reduce_scatter(), overlap_comm=self.zero_overlap_comm(), cpu_offload=self.zero_cpu_offload(), mpu=self.mpu, postscale_gradients=self.postscale_gradients(), gradient_predivide_factor=self.gradient_predivide_factor(), gradient_accumulation_steps=self.gradient_accumulation_steps()) else: raise NotImplementedError("ZeRO stage {} not implemented".format(zero_stage)) return optimizer def _configure_progressive_layer_drop(self): pld = ProgressiveLayerDrop(theta=self.pld_theta(), gamma=self.pld_gamma()) return pld def deepspeed_io(self, dataset, batch_size=None, route=ROUTE_TRAIN, pin_memory=True, data_sampler=None, collate_fn=None, num_local_io_workers=None): if not isinstance(dataset, torch.utils.data.Dataset): raise ValueError("Training data must be a torch Dataset") if data_sampler is None and (route == ROUTE_PREDICT or route == ROUTE_EVAL): data_sampler = torch.utils.data.SequentialSampler(dataset) if batch_size is None: batch_size = self.train_micro_batch_size_per_gpu() if collate_fn is None: collate_fn = self.collate_fn # Currently we only use timer in train route deepspeed_io_timer = None if route == ROUTE_TRAIN: deepspeed_io_timer = self.tput_timer # If mpu is provied, forward world size and parallel rank to sampler. data_parallel_world_size = None data_parallel_rank = None if self.mpu is not None: data_parallel_world_size = self.mpu.get_data_parallel_world_size() data_parallel_rank = self.mpu.get_data_parallel_rank() return DeepSpeedDataLoader(dataset=dataset, batch_size=batch_size, pin_memory=pin_memory, collate_fn=collate_fn, local_rank=self.local_rank, tput_timer=deepspeed_io_timer, num_local_io_workers=num_local_io_workers, data_sampler=data_sampler, data_parallel_world_size=data_parallel_world_size, data_parallel_rank=data_parallel_rank) def train(self, mode=True): r""" """ self.warn_unscaled_loss = True self.module.train(mode) def eval(self): r""" """ self.warn_unscaled_loss = True self.module.train(False) def _scale_loss(self, prescaled_loss): if isinstance(prescaled_loss, torch.Tensor): scaled_loss = prescaled_loss / self.gradient_accumulation_steps() elif isinstance(prescaled_loss, tuple) or isinstance(prescaled_loss, list): scaled_loss = [] for l in prescaled_loss: if isinstance(l, torch.Tensor): scaled_loss.append(l / self.gradient_accumulation_steps()) else: scaled_loss.append(l) else: scaled_loss = prescaled_loss if self.warn_unscaled_loss: logger.warning( f'DeepSpeed unable to scale loss because of type: {type(prescaled_loss)}' ) self.warn_unscaled_loss = False return scaled_loss def forward(self, *inputs, **kwargs): r"""Execute forward propagation Arguments: *inputs: Variable length input list **kwargs: variable length keyword arguments """ if self.module.training and self.progressive_layer_drop: kwargs.update(self.progressive_layer_drop.get_state()) if self.wall_clock_breakdown(): self.timers('forward_microstep').start() self.timers('forward').start() if self.training_dataloader is None: self.tput_timer.start() loss = self.module(*inputs, **kwargs) if self.wall_clock_breakdown(): self.timers('forward').stop() self.timers('forward_microstep').stop() return loss def allreduce_gradients(self, bucket_size=MEMORY_OPT_ALLREDUCE_SIZE): #Zero stage 2 communicates during non gradient accumulation boundaries as well if self.zero_optimization_partition_gradients(): self.optimizer.overlapping_partition_gradients_reduce_epilogue() #Communicate only at gradient accumulation boundaries elif self.is_gradient_accumulation_boundary(): if self.zero_optimization_stage() == ZERO_OPTIMIZATION_OPTIMIZER_STATES: assert self.zero_reduce_scatter() self.optimizer.reduce_scatter_gradients( postscale_gradients=self.postscale_gradients(), gradient_predivide_factor=self.gradient_predivide_factor(), gradient_average=self.gradient_average) else: self.buffered_allreduce_fallback(elements_per_buffer=bucket_size) def backward(self, loss, allreduce_gradients=True, release_loss=False): r"""Execute backward pass on the loss Arguments: loss: Torch tensor on which to execute backward propagation allreduce_gradients: If this is False, then gradient averaging will be skipped. Default is True. """ if not allreduce_gradients: logger.warning( f'Argument `allreduce_gradients` is deprecated, ignored, and will soon be removed' ) # scale loss w.r.t. gradient accumulation if needed if self.gradient_accumulation_steps() > 1: loss = self._scale_loss(loss.float()) # Log training Loss if self.tensorboard_enabled(): if self.is_gradient_accumulation_boundary(): if self.global_rank == 0: self.summary_events = [ (f'Train/Samples/train_loss', loss.mean().item() * self.gradient_accumulation_steps(), self.global_samples) ] for event in self.summary_events: # write_summary_events self.summary_writer.add_scalar(event[0], event[1], event[2]) self.summary_writer.flush() if self.wall_clock_breakdown(): self.timers('backward_microstep').start() self.timers('backward').start() assert self.optimizer is not None, "must provide optimizer during " \ "init in order to use backward" if self.wall_clock_breakdown(): self.timers('backward_inner_microstep').start() self.timers('backward_inner').start() if self.zero_optimization(): self.optimizer.is_gradient_accumulation_boundary = self.is_gradient_accumulation_boundary( ) self.optimizer.backward(loss) elif self.amp_enabled(): # AMP requires delaying unscale when inside gradient accumulation boundaries # https://nvidia.github.io/apex/advanced.html#gradient-accumulation-across-iterations delay_unscale = not self.is_gradient_accumulation_boundary() with amp.scale_loss(loss, self.optimizer, delay_unscale=delay_unscale) as scaled_loss: scaled_loss.backward() elif self.fp16_enabled(): self.optimizer.backward(loss) else: loss.backward() if self.wall_clock_breakdown(): self.timers('backward_inner').stop() self.timers('backward_inner_microstep').stop() if self.wall_clock_breakdown(): self.timers('backward_allreduce_microstep').start() self.timers('backward_allreduce').start() if self.enable_backward_allreduce: self.allreduce_gradients() if self.wall_clock_breakdown(): self.timers('backward_allreduce').stop() self.timers('backward_allreduce_microstep').stop() self.timers('backward').stop() self.timers('backward_microstep').stop() if release_loss: # loss.data = None pass return loss def is_gradient_accumulation_boundary(self): """Query whether the current micro-batch is at the boundary of gradient accumulation, and thus will trigger gradient reductions and an optimizer step. Returns: bool: if the current step is a gradient accumulation boundary. """ return (self.micro_steps + 1) % \ self.gradient_accumulation_steps() == 0 def zero_grad(self): """ Zero parameter grads. """ for param_name, param in self.module.named_parameters(): param.grad = None def clip_fp32_gradients(self): torch.nn.utils.clip_grad_norm_(parameters=self.module.parameters(), max_norm=self.gradient_clipping()) def _take_model_step(self, lr_kwargs): if self.gradient_clipping() > 0.0: if not self.fp16_enabled() and not self.amp_enabled(): self.clip_fp32_gradients() elif self.amp_enabled(): # AMP's recommended way of doing clipping # https://nvidia.github.io/apex/advanced.html#gradient-clipping master_params = amp.master_params(self.optimizer) torch.nn.utils.clip_grad_norm_(parameters=master_params, max_norm=self.gradient_clipping()) self.optimizer.step() #zero grad in basic optimizer could be unreliable and may not exhibit #the behaviour that we want if not self.zero_optimization() and not self.fp16_enabled( ) and not self.amp_enabled(): self.zero_grad() else: self.optimizer.zero_grad() report_progress = self.global_rank == 0 if self.global_rank else True # Check overlow here since in DS fp16 optimizer, the overflow is updated in above step() function. overflow = False if hasattr(self.optimizer, 'overflow'): overflow = self.optimizer.overflow if overflow: self.skipped_steps += 1 else: if self.lr_scheduler is not None: self.lr_scheduler.step(**(lr_kwargs or {})) if report_progress and (self.global_steps + 1) % self.steps_per_print() == 0: self._report_progress(self.global_steps + 1) self.global_steps += 1 self.global_samples += self.train_batch_size() def step(self, lr_kwargs=None): r"""Execute the weight update step after forward and backward propagation on effective_train_batch. """ if self.wall_clock_breakdown(): self.timers('step_microstep').start() self.timers('step').start() assert self.optimizer is not None, "must provide optimizer during " \ "init in order to use step" report_progress = self.global_rank == 0 if self.global_rank else True # Update the model when we reach gradient accumulation boundaries if self.is_gradient_accumulation_boundary(): if self.progressive_layer_drop: self.progressive_layer_drop.update_state(self.global_steps) self._take_model_step(lr_kwargs) self.tput_timer.stop(report_progress) # Log learning rate if self.tensorboard_enabled(): if self.is_gradient_accumulation_boundary(): if self.global_rank == 0: self.summary_events = [(f'Train/Samples/lr', self.get_lr()[0], self.global_samples)] for event in self.summary_events: # write_summary_events self.summary_writer.add_scalar(event[0], event[1], event[2]) if self.fp16_enabled() and hasattr(self.optimizer, 'cur_scale'): self.summary_events.append((f'Train/Samples/loss_scale', self.optimizer.cur_scale, self.global_samples)) for event in self.summary_events: # write_summary_events self.summary_writer.add_scalar(event[0], event[1], event[2]) self.summary_writer.flush() if self.wall_clock_breakdown(): self.timers('step').stop() self.timers('step_microstep').stop() timer_names = [ 'forward_microstep', 'backward_microstep', 'backward_inner_microstep', 'backward_allreduce_microstep', 'step_microstep' ] self.timers.log(names=timer_names, memory_breakdown=self.memory_breakdown()) # Log timing if self.is_gradient_accumulation_boundary(): if self.tensorboard_enabled(): if self.global_rank == 0: self.summary_events = [ (f'Train/Samples/elapsed_time_ms_forward', self.timers('forward').elapsed(reset=False) * 1000.0, self.global_samples), (f'Train/Samples/elapsed_time_ms_backward', self.timers('backward').elapsed(reset=False) * 1000.0, self.global_samples), (f'Train/Samples/elapsed_time_ms_backward_inner', self.timers('backward_inner').elapsed(reset=False) * 1000.0, self.global_samples), (f'Train/Samples/elapsed_time_ms_backward_allreduce', self.timers('backward_allreduce').elapsed(reset=False) * 1000.0, self.global_samples), (f'Train/Samples/elapsed_time_ms_step', self.timers('step').elapsed(reset=False) * 1000.0, self.global_samples) ] for event in self.summary_events: # write_summary_events self.summary_writer.add_scalar(event[0], event[1], event[2]) self.summary_writer.flush() if self.wall_clock_breakdown(): self.timers.log([ 'forward', 'backward', 'backward_inner', 'backward_allreduce', 'step' ]) self.micro_steps += 1 def _get_optimizer_param(self, param_name): result = [] if not self.optimizer: return result for group in self.optimizer.param_groups: if param_name in group: result.append(group[param_name]) else: result.append(0.0) return result def get_lr(self): return self._get_optimizer_param('lr') def get_type(self): return self._get_optimizer_param('type') def get_mom(self): if self.optimizer_name() in ['SGD', 'RMSprop']: return self._get_optimizer_param('momentum') else: return self._get_optimizer_param('betas') def get_pld_theta(self): if self.progressive_layer_drop: return self.progressive_layer_drop.get_theta() else: return None def _report_progress(self, step): lr = self.get_lr() mom = self.get_mom() log_dist(f'step={step}, skipped={self.skipped_steps}, lr={lr}, mom={mom}', ranks=[0]) def allreduce_bucket(self, bucket): tensor = self.flatten(bucket) tensor_to_allreduce = tensor if self.allreduce_always_fp32(): tensor_to_allreduce = tensor.float() if self.postscale_gradients(): if self.gradient_predivide_factor() != 1.0: tensor_to_allreduce.mul_(1. / self.gradient_predivide_factor()) dist.all_reduce(tensor_to_allreduce, group=self.data_parallel_group) if self.gradient_average: if self.gradient_predivide_factor() != self.dp_world_size: tensor_to_allreduce.mul_(self.gradient_predivide_factor() / self.dp_world_size) else: tensor_to_allreduce.div_(self.dp_world_size) dist.all_reduce(tensor_to_allreduce, group=self.data_parallel_group) if self.allreduce_always_fp32() and tensor is not tensor_to_allreduce: tensor.copy_(tensor_to_allreduce) return tensor def allreduce_and_copy(self, small_bucket): allreduced = self.allreduce_bucket(small_bucket) for buf, synced in zip(small_bucket, self.unflatten(allreduced, small_bucket)): buf.copy_(synced) def allreduce_no_retain(self, bucket, numel_per_bucket=500000000): small_bucket = [] numel = 0 for tensor in bucket: small_bucket.append(tensor) numel = numel + tensor.numel() if numel > numel_per_bucket: self.allreduce_and_copy(small_bucket) small_bucket = [] numel = 0 if len(small_bucket) > 0: self.allreduce_and_copy(small_bucket) def buffered_allreduce_fallback(self, grads=None, elements_per_buffer=500000000): grads = [] for param_name, param in self.module.named_parameters(): if param.grad is None: # In cases where there is an imbalance of empty grads across # ranks we must create empty grads, this will ensure that every # rank is reducing the same size. In some cases it may make # sense in the future to support the ability to average not # w.r.t. world size but with a different value. param.grad = torch.zeros(param.size(), dtype=param.dtype, device=param.device) grads.append(param.grad.data) else: grad_data = param.grad.data if self.sparse_gradients_enabled( ) and param_name in self.csr_tensor_module_names: grads.append(CSRTensor(grad_data)) else: grads.append(grad_data) split_buckets = split_half_float_double_csr(grads) for i, bucket_tuple in enumerate(split_buckets): bucket_type, bucket = bucket_tuple if bucket_type == CSRTensor.type(): self.csr_allreduce_no_retain(bucket) else: self.allreduce_no_retain(bucket, numel_per_bucket=elements_per_buffer) def csr_allreduce_no_retain(self, bucket): allreduced_csrs = self.csr_allreduce_bucket(bucket) # Densify csr tensor and copy back to original location for csr in allreduced_csrs: dense_tensor = csr.to_dense() csr.orig_dense_tensor.copy_(dense_tensor) def csr_allreduce_bucket(self, bucket): csr_list = [] for csr in bucket: csr_list.append(self.csr_allreduce(csr)) return csr_list def csr_allreduce(self, csr): # Pre-divide for fp16 stability csr.values.div_(self.dp_world_size) indices_device_list = self.csr_all_gather(csr.indices) values_device_list = self.csr_all_gather(csr.values) csr.indices = torch.cat(indices_device_list) csr.values = torch.cat(values_device_list) return csr def csr_all_gather(self, value): my_size = torch.LongTensor([value.size()[0]]).to(self.device) all_sizes = self.all_gather_scalar(my_size) max_size = torch.cat(all_sizes).max() fill_size = (max_size - my_size) assert value.dim() in [1, 2] if value.dim() == 1: if fill_size > 0: value = torch.cat([value, value.new_zeros(fill_size)]) tensor_list = [value.new_zeros(max_size) for _ in range(self.dp_world_size)] else: if fill_size > 0: value = torch.cat([value, value.new_zeros(fill_size, value.size()[1])]) tensor_list = [ value.new_zeros(max_size, value.size()[1]) for _ in range(self.dp_world_size) ] dist.all_gather(tensor_list, value, group=self.data_parallel_group) tensors = [] for dev_idx, t in enumerate(tensor_list): size = all_sizes[dev_idx][0] tensors.append( t.index_select(0, torch.LongTensor(range(size)).to(self.device))) return tensors def all_gather_scalar(self, value): tensor_list = [value.new_zeros(value.size()) for _ in range(self.dp_world_size)] dist.all_gather(tensor_list, value, group=self.data_parallel_group) return tensor_list def module_state_dict(self, destination=None, prefix='', keep_vars=False): sd = self.module.state_dict(destination, prefix, keep_vars) return sd def load_module_state_dict(self, state_dict, strict=True): self.module.load_state_dict(state_dict, strict=strict) def _get_rank_zero_ckpt_name(self, checkpoints_path, tag, mp_rank, dp_rank): filename = 'zero_pp_rank_{}'.format(dp_rank) zero_ckpt_name = os.path.join( checkpoints_path, str(tag), filename + '_mp_rank_{:02d}'.format(mp_rank) + 'optim_states.pt') return zero_ckpt_name def _get_zero_ckpt_name(self, checkpoints_path, tag): mp_rank = 0 if self.mpu is None else self.mpu.get_model_parallel_rank() pp_rank = torch.distributed.get_rank(group=self.optimizer.dp_process_group) return self._get_rank_zero_ckpt_name(checkpoints_path, tag, mp_rank, pp_rank) def _get_ckpt_name(self, checkpoints_path, tag): mp_rank = 0 if self.mpu is None else self.mpu.get_model_parallel_rank() ckpt_name = os.path.join(checkpoints_path, str(tag), 'mp_rank_{:02d}'.format(mp_rank) + '_model_states.pt') return ckpt_name def load_checkpoint(self, load_dir, tag=None, load_module_strict=True, load_optimizer_states=True, load_lr_scheduler_states=True): """Load training checkpoint Arguments: load_dir: Required. Directory to load the checkpoint from tag: Checkpoint tag used as a unique identifier for checkpoint, if not provided will attempt to load tag in 'latest' file load_module_strict: Optional. Boolean to strictly enforce that the keys in state_dict of module and checkpoint match. load_optimizer_states: Optional. Boolean to load the training optimizer states from Checkpoint. Ex. ADAM's momentum and variance load_lr_scheduler_states: Optional. Boolean to add the learning rate scheduler states from Checkpoint. Returns: A tuple of ``load_path`` and ``client_state``. *``load_path``: Path of the loaded checkpoint. ``None`` if loading the checkpoint failed. *``client_state``: State dictionary used for loading required training states in the client code. """ if tag is None: latest_path = os.path.join(load_dir, 'latest') if os.path.isfile(latest_path): with open(latest_path, 'r') as fd: tag = fd.read().strip() else: logger.warning(f"Unable to find latest file at {latest_path}, if trying to load latest " \ "checkpoint please ensure this file exists or pass an explicit checkpoint tag when loading a checkpoint.") return None, None load_path, client_states = self._load_checkpoint(load_dir, tag, load_module_strict=load_module_strict, load_optimizer_states=load_optimizer_states, load_lr_scheduler_states=load_lr_scheduler_states) if self.zero_optimization() and load_path is not None: self._load_zero_checkpoint(load_dir, tag, load_optimizer_states=load_optimizer_states) return load_path, client_states def _load_checkpoint(self, load_dir, tag, load_module_strict=True, load_optimizer_states=True, load_lr_scheduler_states=True): load_path = self._get_ckpt_name(load_dir, tag) if not os.path.exists(load_path): logger.warn( 'Client provided checkpoint load path: {} does not exist ... skip checkpoint load' .format(load_path)) return None, None logger.info(f'rank: {self.global_rank} loading checkpoint: {load_path}') checkpoint = torch.load(load_path, map_location=lambda storage, loc: storage) if isinstance(self.module, PipelineModule): # Pipeline parallelism uses this to load its own checkpoint files. self._curr_ckpt_path = os.path.join(load_dir, tag) self.load_module_state_dict(state_dict=checkpoint['module'], strict=load_module_strict) if self.optimizer is not None and not self.zero_optimization(): if self.fp16_enabled(): self.optimizer.load_state_dict( checkpoint['optimizer'], load_optimizer_states=load_optimizer_states) elif load_optimizer_states: self.optimizer.load_state_dict(checkpoint['optimizer']) if load_lr_scheduler_states and self.lr_scheduler is not None: self.lr_scheduler.load_state_dict(checkpoint['lr_scheduler']) self.csr_tensor_module_names = checkpoint['csr_tensor_module_names'] self.global_steps = checkpoint['global_steps'] self.global_samples = checkpoint.get('global_samples', self.global_steps * self.train_batch_size()) self.skipped_steps = checkpoint['skipped_steps'] self.loaded_checkpoint_mp_world_size = checkpoint['mp_world_size'] self.loaded_checkpoint_dp_world_size = checkpoint['dp_world_size'] deepspeed_states = [ 'module', 'optimizer', 'lr_scheduler', 'csr_tensor_module_names', 'skipped_steps', 'global_steps', 'dp_world_size', 'mp_world_size' ] client_state = { key: value for key, value in checkpoint.items() if not key in deepspeed_states } return load_path, client_state def _load_zero_checkpoint(self, load_dir, tag, load_optimizer_states=True): zero_sd_list = self._get_all_zero_checkpoints(load_dir, tag) if zero_sd_list is None: return self.optimizer.load_state_dict( state_dict_list=zero_sd_list, load_optimizer_states=load_optimizer_states, load_from_fp32_weights=self.zero_load_from_fp32_weights()) print( f'loading {len(zero_sd_list)} zero partition checkpoints for rank {self.global_rank}' ) def _get_mp_rank_zero_checkpoint_names(self, load_dir, tag, mp_rank, dp_world_size): zero_ckpt_names = [] for dp_rank in range(dp_world_size): ckpt_name = self._get_rank_zero_ckpt_name(checkpoints_path=load_dir, tag=tag, mp_rank=mp_rank, dp_rank=dp_rank) zero_ckpt_names.append(ckpt_name) return zero_ckpt_names def _get_all_zero_checkpoint_names(self, load_dir, tag, mp_world_size, dp_world_size): zero_ckpt_names = [] for mp_rank in range(mp_world_size): mp_rank_ckpt_names = self._get_mp_rank_zero_checkpoint_names( load_dir=load_dir, tag=tag, mp_rank=mp_rank, dp_world_size=dp_world_size) zero_ckpt_names += mp_rank_ckpt_names return zero_ckpt_names def _get_all_zero_checkpoints(self, load_dir, tag): mp_rank = 0 if self.mpu is None else self.mpu.get_model_parallel_rank() zero_ckpt_names = self._get_mp_rank_zero_checkpoint_names( load_dir=load_dir, tag=tag, mp_rank=mp_rank, dp_world_size=self.loaded_checkpoint_dp_world_size) invalid_zero_ckpt_paths = [] for ckpt_name in zero_ckpt_names: if not os.path.exists(ckpt_name): invalid_zero_ckpt_paths.append(ckpt_name) if len(invalid_zero_ckpt_paths) > 0: logger.warn( f"Client provided zero checkpoint load paths: {invalid_zero_ckpt_paths} does not exist" ) return None zero_sd_list = [] for ckpt_name in zero_ckpt_names: zero_sd_list.append(torch.load(ckpt_name, map_location='cpu')) zero_optimizer_sd = [sd['optimizer_state_dict'] for sd in zero_sd_list] print( f"successfully loaded {len(zero_optimizer_sd)} ZeRO state_dicts for rank {self.global_rank}" ) return zero_optimizer_sd def save_checkpoint(self, save_dir, tag=None, client_state={}, save_latest=True, save_zero=True): r"""Save training checkpoint Arguments: save_dir: Required. Directory for saving the checkpoint tag: Optional. Checkpoint tag used as a unique identifier for the checkpoint, global step is used if not provided. client_state: Optional. State dictionary used for saving required training states in the client code. save_latest: Optional. Save a file 'latest' pointing to the latest saved checkpoint. """ # This is to make sure the checkpoint names are created without collision # There seems to be issue creating them in parallel # Ensure save_dir directory exists os.makedirs(save_dir, exist_ok=True) if tag is None: tag = f"global_step{self.global_steps}" if self.save_non_zero_checkpoint: self._create_checkpoint_file(save_dir, tag, False) self._save_checkpoint(save_dir, tag, client_state=client_state) if self.save_zero_checkpoint and save_zero: self._create_zero_checkpoint_files(save_dir, tag) self._save_zero_checkpoint(save_dir, tag) # Save latest checkpoint tag if save_latest: with open(os.path.join(save_dir, 'latest'), 'w') as fd: fd.write(tag) return True def _create_checkpoint_file(self, save_dir, tag, zero_checkpoint): name_function = self._get_zero_ckpt_name if zero_checkpoint else self._get_ckpt_name try: checkpoint_name = name_function(save_dir, tag) ensure_directory_exists(checkpoint_name) except: logger.error(f'Failed saving model checkpoint to {save_dir} with tag {tag}') return False return True def _create_zero_checkpoint_files(self, save_dir, tag): # zero checkpoint files are created sequentially try: checkpoint_name = self._get_zero_ckpt_name(save_dir, tag) if self.local_rank == 0: ensure_directory_exists(checkpoint_name) else: while not os.path.exists(os.path.dirname(checkpoint_name)): time.sleep(1) except: logger.error(f'Failed saving model checkpoint to {save_dir} with tag {tag}') return False return True # dist.barrier() def _save_checkpoint(self, save_dir, tag, client_state={}): save_path = self._get_ckpt_name(save_dir, tag) # A hack to save the checkpointing directory. Pipeline parallelism overrides # module_state_dict() and uses this path to save the model. module_state_dict() # then instead just returns None. self._curr_ckpt_path = os.path.join(save_dir, tag) state = { 'module': self.module_state_dict(), 'optimizer': self.optimizer.state_dict() if self.optimizer and not self.zero_optimization() else None, 'lr_scheduler': self.lr_scheduler.state_dict() if self.lr_scheduler is not None else None, 'csr_tensor_module_names': self.csr_tensor_module_names, 'skipped_steps': self.skipped_steps, 'global_steps': self.global_steps, 'global_samples': self.global_samples, 'dp_world_size': self.dp_world_size, 'mp_world_size': self.mp_world_size } state.update(client_state) log_dist(message=f'Saving model checkpoint: {save_path}', ranks=[0]) #logger.info('Saving model checkpoint: {}'.format(save_path)) torch.save(state, save_path) self._curr_save_path = None def _save_zero_checkpoint(self, save_path, tag): zero_checkpoint_name = self._get_zero_ckpt_name(save_path, tag) zero_sd = {'optimizer_state_dict': self.optimizer.state_dict()} torch.save(zero_sd, zero_checkpoint_name) logger.info('zero checkpoint saved {}'.format(zero_checkpoint_name))

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/tools/ds_fix/stage1.py

import math import torch import torch.distributed as dist from torch._utils import _flatten_dense_tensors, _unflatten_dense_tensors from collections import defaultdict from deepspeed.runtime.zero.utils import _initialize_parameter_parallel_groups from deepspeed.runtime.fp16.loss_scaler import LossScaler, DynamicLossScaler from deepspeed.runtime.utils import get_grad_norm, CheckOverflow from deepspeed.runtime.zero.config import ZERO_OPTIMIZATION_OPTIMIZER_STATES from deepspeed.utils import logger, log_dist def get_alignment_padding(flattened_lean_size, sub_partition_id, sub_partition_size): sub_partition_high_limit = (sub_partition_id + 1) * sub_partition_size if sub_partition_high_limit <= flattened_lean_size: return 0 else: return min(sub_partition_size, sub_partition_high_limit - flattened_lean_size) def get_group_alignment_padding(tensor_list, sub_partition_size, sub_partition_count): group_paddings = [] flattened_size = sum([tensor.numel() for tensor in tensor_list]) for i in range(sub_partition_count): padding = get_alignment_padding(flattened_size, i, sub_partition_size) group_paddings.append(padding) return group_paddings def flatten_dense_tensors_sub_partition_aligned(tensor_list, dp, max_elements_per_comm, pg): assert max_elements_per_comm >= dp, f"max_elements_per_comm {max_elements_per_comm} < dp {dp}" num_elements = sum(t.numel() for t in tensor_list) log_dist("Total number of elements in model: {}, max elements per com: {}".format( num_elements, max_elements_per_comm), ranks=[0]) # Compute aligned partition size based on parameter count aligned_param_partition_size = math.ceil(num_elements / dp) # Compute aligned partition size based on communication size aligned_comm_partition_size = int(max_elements_per_comm // dp) if aligned_param_partition_size <= aligned_comm_partition_size: sub_partition_count = 1 sub_partition_size = aligned_param_partition_size else: sub_partition_count = math.ceil(aligned_param_partition_size / aligned_comm_partition_size) sub_partition_size = aligned_comm_partition_size # Compute required padding for alignment to dp and max_elements_per_comm padding = (sub_partition_count * sub_partition_size * dp) - num_elements log_dist( f"sub_partition_count: {sub_partition_count}, sub_partition_size: {sub_partition_size}, padding: {padding}", ranks=[0]) log_dist( f"number of elements with padding: {num_elements} + {padding} = {num_elements + padding}", ranks=[0]) if padding == 0: aligned_tensor_list = tensor_list else: pad_tensor = torch.zeros(padding, device=tensor_list[0].device, dtype=tensor_list[0].dtype) aligned_tensor_list = tensor_list + [pad_tensor] flat_tensors = _flatten_dense_tensors(aligned_tensor_list) return flat_tensors def _single_range_check(current_index, start_index, end_index, tensor_size): offset = 0 if (current_index >= start_index) and (current_index < end_index): # Fully inside bounds return True, offset elif (start_index > current_index) and (start_index < (current_index + tensor_size)): # Partially contained, compute offset offset = start_index - current_index return True, offset else: return False, offset def _range_check(current_index, element_intervals, tensor_size): results = [] for comm_idx, interval in enumerate(element_intervals): start_index, end_index = interval contained, offset = _single_range_check(current_index, start_index, end_index, tensor_size) if contained: results.append((contained, offset, comm_idx)) if len(results) == 0: return [(False, 0, -1)] return results class FP16_DeepSpeedZeroOptimizer_Stage1(object): """ FP16_DeepSpeedZeroOptimizer_Stage1 designed to reduce the memory footprint required for training large deep learning models. For more details please see ZeRO: Memory Optimization Towards Training A Trillion Parameter Models https://arxiv.org/abs/1910.02054 This version aligns with stage-1 in the paper above. """ def __init__(self, init_optimizer, static_loss_scale=1.0, dynamic_loss_scale=False, dynamic_loss_args=None, verbose=True, dp_process_group=None, partition_size=None, mpu=None, all_gather_partitions=True, allgather_size=500000000, clip_grad=0.0, max_elements_per_comm=5e8, elastic_checkpoint=True): if dp_process_group is not None and partition_size is not None: raise ValueError("Cannot specify both dp_process_group " "and partition size") if dp_process_group is None: dp_process_group = _initialize_parameter_parallel_groups(partition_size) if not torch.cuda.is_available: raise SystemError("Cannot use fp16 without CUDA.") self.optimizer = init_optimizer self.verbose = verbose self.dp_process_group = dp_process_group # TODO: automatically turn off if #params > some_limit self.all_gather_partitions = all_gather_partitions self.allgather_size = allgather_size # self.max_elements_per_comm = max_elements_per_comm # logger.info("max_elements_per_comm={}".format(max_elements_per_comm)) self.elastic_checkpoint = elastic_checkpoint logger.info(f'ZeRO Elastic Checkpoint = {elastic_checkpoint}') # param flattened by groups self.fp16_groups = [] self.fp16_groups_flat = [] # Setup bookkeeping data structures depending on partitioning type # parallel_sub_partitioned_fp16_groups[group-idx] -> [comm-ids] -> [rank-ids] self.parallel_sub_partitioned_fp16_groups = [] # same underlying data as above but viewed as: [groups] -> [rank-ids] -> [comm-ids] self.parallel_comm_sub_partitioned_fp16_groups = [] # 32-bit sub-partitions of the parallel partitioned parameters # that this process will update self.local_sub_partitions_of_fp32_groups = [] # param partition info # parameters in each group that will not be updated by this process directly self.params_not_local = [] # parameters that will be updated by this process directly self.params_in_rank_sub_partitions = [] # parameter offsets for parameters in sub-partitions. Parameter # boundaries may not align with sub-partition boundaries # so we need to keep track of the offsets self.params_in_rank_sub_partitions_offsets = [] # number of elements per sub-partition in each group self.sub_partition_sizes = [] # number of communication intervals for each group self.num_comm_intervals_per_group = [] local_rank = dist.get_rank(group=self.dp_process_group) self.group_paddings = [] self.partition_count = dist.get_world_size(group=self.dp_process_group) self.default_device = self.optimizer.param_groups[0]['params'][0].device # max elems per param group self.max_elems_per_comm = [] # loop to deal with groups for i, param_group in enumerate(self.optimizer.param_groups): # push this group to list before modify self.fp16_groups.append(param_group['params']) # calculate best max elements per comm based to minimize padding self.max_elems_per_comm.append( self.best_max_elems_per_comm( num_elements=sum(t.numel() for t in self.fp16_groups[i]), max_elements_per_comm=max_elements_per_comm, dp=dist.get_world_size(group=self.dp_process_group))) # flattens all tensors into single 1d tensor aligned with sub-partition size for later dividing # RS: create aligned sub-partitions flat_aligned_params = flatten_dense_tensors_sub_partition_aligned( tensor_list=self.fp16_groups[i], dp=dist.get_world_size(group=self.dp_process_group), max_elements_per_comm=self.max_elems_per_comm[i], pg=self.dp_process_group) self.fp16_groups_flat.append(flat_aligned_params) # TODO: I don't think this does anything? # set model fp16 weight to slices of flattened buffer updated_params = _unflatten_dense_tensors(self.fp16_groups_flat[i], self.fp16_groups[i]) for p, q in zip(self.fp16_groups[i], updated_params): p.data = q.data # divide the flat weights into near equal partition equal to the data parallel degree # each process will compute on a different part of the partition # RS: split into two layer list -> [comm-id] -> [sub-partitions per rank] comm_partitions, dp_sub_partitions, element_intervals, sub_partition_size, num_comm_intervals = \ self.get_data_parallel_sub_partitions( tensor=self.fp16_groups_flat[i], max_elements_per_comm=self.max_elems_per_comm[i], world_size=dist.get_world_size( group=self.dp_process_group), dp_process_group=self.dp_process_group ) self.parallel_comm_sub_partitioned_fp16_groups.append( comm_partitions) # comm -> rank self.parallel_sub_partitioned_fp16_groups.append( dp_sub_partitions) # rank -> comm self.sub_partition_sizes.append(sub_partition_size) self.num_comm_intervals_per_group.append(num_comm_intervals) # data_parallel_partitions = self.get_data_parallel_partitions(self.fp16_groups_flat[i]) # self.parallel_partitioned_fp16_groups.append(data_parallel_partitions) # a partition of the fp32 master weights that will be updated by this process # RS: store/detach/cast our local sub-partitions local_sub_partitions = [] for sub_partition in self.parallel_sub_partitioned_fp16_groups[i][ local_rank]: fp32_sub_partition = sub_partition.clone().float().detach() fp32_sub_partition.requires_grad = True local_sub_partitions.append(fp32_sub_partition) self.local_sub_partitions_of_fp32_groups.append(local_sub_partitions) # Compute sub_partition paddings sub_partition_paddings = get_group_alignment_padding( tensor_list=self.fp16_groups[i], sub_partition_size=sub_partition_size, sub_partition_count=num_comm_intervals * self.partition_count) self.group_paddings.append(sub_partition_paddings) # modify optimizer of have flat master weight # self.single_partition_of_fp32_groups[i].requires_grad = True # keep this in case internal optimizer uses it param_group['params'] = self.local_sub_partitions_of_fp32_groups[i] # RS: divide up the sub-partitions and keep track of offsets for each param # partition_size = len(self.fp16_groups_flat[i]) / dist.get_world_size(group=self.dp_process_group) params_in_rank_sub_partition, params_in_rank_sub_partitions_offsets, params_not_local = self.get_all_sub_partition_info( tensor_list=self.fp16_groups[i], all_element_intervals=element_intervals, local_rank=local_rank, world_size=dist.get_world_size(group=self.dp_process_group) ) self.params_in_rank_sub_partitions.append(params_in_rank_sub_partition) self.params_not_local.append(params_not_local) self.params_in_rank_sub_partitions_offsets.append( params_in_rank_sub_partitions_offsets) # we may have a way of fusing dynamic scale. Do not support for now if dynamic_loss_scale: if dynamic_loss_args is None: self.loss_scaler = DynamicLossScaler() else: self.loss_scaler = DynamicLossScaler(**dynamic_loss_args) self.dynamic_loss_scale = True else: self.dynamic_loss_scale = False self.loss_scaler = LossScaler(scale=static_loss_scale) self.cur_iter = 0 self.mpu = mpu self.clip_grad = clip_grad self.overflow = False self.overflow_checker = CheckOverflow(self.fp16_groups, mpu=self.mpu, zero_reduce_scatter=True) self._initialize_optimizer_states() self.hack_first_step = True def _initialize_optimizer_states(self): for group_idx, group in enumerate(self.local_sub_partitions_of_fp32_groups): for idx, sub_partition_param in enumerate(group): sub_partition_grad = torch.zeros(int( self.sub_partition_sizes[group_idx]), dtype=sub_partition_param.dtype).cuda() sub_partition_param.grad = sub_partition_grad self.optimizer.step() for group in self.local_sub_partitions_of_fp32_groups: for idx, sub_partition_param in enumerate(group): sub_partition_param.grad = None @staticmethod def best_max_elems_per_comm(num_elements, max_elements_per_comm, dp): # if we use max-elems-per-comm as is, how many comm intervals will there be max_comm_intervals = math.ceil(num_elements / max_elements_per_comm) padding_for_max_comm = (max_elements_per_comm * max_comm_intervals) - num_elements # if we use 1 less comm interval how much extra comm padding would be required min_comm_intervals = num_elements // max_elements_per_comm if min_comm_intervals == 0: log_dist(f'Using default max_elements_per_comm {max_elements_per_comm}', ranks=[0]) return max_elements_per_comm padding_for_min_comm = math.ceil(num_elements / (dp * min_comm_intervals)) # choose padding that uses least amount of overhead if padding_for_max_comm > padding_for_min_comm: new_max_elements_per_comm = padding_for_min_comm + max_elements_per_comm log_dist( f'Updating max_elements_per_comm from {max_elements_per_comm} -> {new_max_elements_per_comm}', ranks=[0]) return new_max_elements_per_comm else: log_dist(f'Using default max_elements_per_comm {max_elements_per_comm}', ranks=[0]) return max_elements_per_comm @staticmethod def get_data_parallel_sub_partitions(tensor, max_elements_per_comm, world_size, dp_process_group=None): total_num_elements = tensor.numel() # if total elements is less than our max, revert to splitting into dp partitions max_elements_per_comm = min(total_num_elements, max_elements_per_comm) sub_partition_size = int(max_elements_per_comm // world_size) # Ensure partition alignment was done correctly num_sub_partitions = int(total_num_elements // sub_partition_size) assert total_num_elements % sub_partition_size == 0, "{} % {} != 0".format(total_num_elements, sub_partition_size) # Ensure comm interval alignment was done correctly. num_comm_intervals = int(num_sub_partitions // world_size) assert num_sub_partitions % world_size == 0, "{} % {} != 0".format(num_sub_partitions, world_size) if not dist.is_initialized() or dist.get_rank(group=dp_process_group) == 0: logger.info("**** partition info:") logger.info("\t total_num_elements=%s", total_num_elements) logger.info("\t world_size=%s", world_size) logger.info("\t max_elements_per_comm=%s", max_elements_per_comm) logger.info("\t sub_partition_size=%s", sub_partition_size) logger.info("\t num_sub_partitions=%s", num_sub_partitions) logger.info("\t num_comm_intervals=%s", num_comm_intervals) logger.info("****") # [comm_id] -> [rank] comm_partitions = [] for _ in range(num_comm_intervals): comm_partitions.append([]) start = 0 comm_id = 0 element_intervals = defaultdict( list) # [rank] -> [(start,end), (start,end), ...] for idx in range(num_sub_partitions): rank_id = idx % world_size sub_partition = tensor.narrow(0, start, sub_partition_size).detach() element_intervals[rank_id].append((start, start + sub_partition_size)) comm_partitions[comm_id].append(sub_partition) start = start + sub_partition_size if rank_id == (world_size - 1): comm_id += 1 # [rank] -> [comm_id] sub_partitions = [] for _ in range(world_size): sub_partitions.append([]) for comm_id, partitions in enumerate(comm_partitions): for rank_id, partition in enumerate(partitions): sub_partitions[rank_id].append(partition) return comm_partitions, sub_partitions, element_intervals, sub_partition_size, num_comm_intervals @staticmethod def get_all_sub_partition_info(tensor_list, all_element_intervals, local_rank, world_size): params_not_local = [] # [rank] -> [comm-id] -> [param/offset] params_in_rank_sub_partition = [] params_in_rank_sub_partitions_offsets = [] for rank in range(world_size): params_in_local_sub_partition = [] local_sub_partition_offsets = [] comm_tensor_list = [] comm_offset_list = [] current_index = 0 prev_comm_idx = 0 for iii, tensor in enumerate(tensor_list): tensor_size = tensor.numel() #if local_rank == 0: # # logger.info("rank={}, current_index={}, tensor_size={}, tensor-idx={}".format(rank, # current_index, tensor_size, iii)) results_list = _range_check(current_index, all_element_intervals[rank], tensor_size) for contained, offset, comm_idx in results_list: #if local_rank == 0: # logger.info("rank={}, contained={}, offset={}, comm_idx={}".format(rank, contained, # offset, comm_idx)) if contained: if prev_comm_idx != comm_idx: params_in_local_sub_partition.append(comm_tensor_list) comm_tensor_list = [] local_sub_partition_offsets.append(comm_offset_list) comm_offset_list = [] comm_tensor_list.append(tensor) comm_offset_list.append(offset) prev_comm_idx = comm_idx elif rank == local_rank: params_not_local.append(tensor) current_index = current_index + tensor_size #assert len(comm_tensor_list) > 0 #assert len(comm_offset_list) > 0 params_in_local_sub_partition.append(comm_tensor_list) local_sub_partition_offsets.append(comm_offset_list) params_in_rank_sub_partition.append(params_in_local_sub_partition) params_in_rank_sub_partitions_offsets.append(local_sub_partition_offsets) return params_in_rank_sub_partition, params_in_rank_sub_partitions_offsets, params_not_local @staticmethod def get_flat_sub_partitions(comm_tensor_list, comm_param_offsets, sub_partition_size, dtype, default_device, num_comm_intervals=None, return_partition_params=False): partition_params = [] final_param_offsets = [] flat_sub_partitions = [] for tensor_list, param_offsets in zip(comm_tensor_list, comm_param_offsets): flat_tensor_list = [] current_size = 0 my_offsets = [] my_params = [] for i, tensor in enumerate(tensor_list): if tensor.grad is None: tensor.grad = torch.zeros(tensor.size(), dtype=tensor.dtype, device=tensor.device) param = tensor tensor = tensor.grad num_elements = tensor.numel() tensor_offset = 0 #we need to offset to get to the right element if i == 0 and param_offsets[i] > 0: tensor_offset = param_offsets[i] num_elements = num_elements - tensor_offset # We don't need all elements of the tensor if this tensor is # larger than we have space for in our curr sub-partition if num_elements > (sub_partition_size - current_size): num_elements = sub_partition_size - current_size #we need a narrow view of the tensor based on the tensor offset and number of elements that #we need from this tensor if tensor_offset > 0 or num_elements < tensor.numel(): flat_tensor_list.append(tensor.contiguous().view(-1).narrow( 0, int(tensor_offset), int(num_elements)).to(dtype)) else: flat_tensor_list.append(tensor.to(dtype)) my_params.append(param) #remember offset into partition and #elems for this tensor my_offsets.append((current_size, num_elements)) current_size = current_size + num_elements #this means its the last partition and does not align with the dp boundary. We need to pad before flattening if current_size < sub_partition_size: my_offsets.append((None, None)) my_params.append(None) if len(tensor_list) == 0: assert default_device != None flat_tensor_list.append( torch.zeros(int(sub_partition_size - current_size), dtype=dtype, device=default_device)) else: flat_tensor_list.append( torch.zeros(int(sub_partition_size - current_size), dtype=dtype, device=tensor_list[0].device)) partition_params.append(my_params) #flat_tensor_list) final_param_offsets.append(my_offsets) assert len(flat_tensor_list) == len(my_offsets), "{} {}".format(len(flat_tensor_list), len(my_offsets)) flat_sub_partitions.append(_flatten_dense_tensors(flat_tensor_list)) if num_comm_intervals is not None and len( flat_sub_partitions) < num_comm_intervals: # logger.info("padding w. sub partitions to ensure uniform communication") device = flat_sub_partitions[0].device for _ in range(num_comm_intervals - len(flat_sub_partitions)): flat_sub_partitions.append( torch.zeros(int(sub_partition_size), dtype=dtype, device=device)) partition_params.append([None]) final_param_offsets.append([(None, None)]) if return_partition_params: assert len(flat_sub_partitions) == len(partition_params) assert len(partition_params) == len(final_param_offsets), "{} {}".format(len(partition_params), len(final_param_offsets)) return flat_sub_partitions, partition_params, final_param_offsets return flat_sub_partitions def zero_grad(self, set_grads_to_None=True): """ Zero FP16 parameter grads. """ # FP32 grad should never exist. # For speed, set model fp16 grad to None by default for group in self.fp16_groups: for p in group: if set_grads_to_None: p.grad = None else: if p.grad is not None: p.grad.detach_() p.grad.zero_() def free_grad_in_param_list(self, param_list): for p in param_list: if isinstance(p, list): for _p in p: _p.grad = None else: p.grad = None def reduce_scatter_gradients(self, postscale_gradients, gradient_predivide_factor, gradient_average): world_size = dist.get_world_size(group=self.dp_process_group) local_rank = dist.get_rank(group=self.dp_process_group) for i, group in enumerate(self.fp16_groups): num_comm_intervals = self.num_comm_intervals_per_group[i] all_sub_partitions = [] for rank in range(world_size): # gsp is list of partitions indexed by comm_idx grad_sub_partitions = self.get_flat_sub_partitions( comm_tensor_list=self.params_in_rank_sub_partitions[i][rank], comm_param_offsets=self.params_in_rank_sub_partitions_offsets[i] [rank], dtype=torch.half, default_device=self.default_device, sub_partition_size=self.sub_partition_sizes[i], num_comm_intervals=self.num_comm_intervals_per_group[i]) all_sub_partitions.append(grad_sub_partitions) assert len(grad_sub_partitions) == num_comm_intervals local_comm_partitions = [] for comm_idx in range(num_comm_intervals): single_comm_all_partitions = [] for rank in range(world_size): single_comm_all_partitions.append(all_sub_partitions[rank][comm_idx]) if postscale_gradients: if gradient_predivide_factor != 1.0: for partition in single_comm_all_partitions: partition.mul_(1. / gradient_predivide_factor) dist.reduce_scatter(output=single_comm_all_partitions[local_rank], input_list=single_comm_all_partitions, group=self.dp_process_group) if gradient_average: # Only need to average our local grads in post scaling if gradient_predivide_factor != world_size: single_comm_all_partitions[local_rank].mul_( gradient_predivide_factor / world_size) else: for partition in single_comm_all_partitions: partition.div_(world_size) dist.reduce_scatter(output=single_comm_all_partitions[local_rank], input_list=single_comm_all_partitions, group=self.dp_process_group) def step(self, closure=None): # First compute norm for all group so we know if there is overflow self.overflow = self.overflow_checker.check() prev_scale = self.loss_scale self._update_scale(self.overflow) if self.overflow: self.zero_grad() if self.verbose: logger.info("[deepspeed] OVERFLOW! Skipping step. Attempted loss " "scale: {}, reducing to {}".format( prev_scale, self.loss_scale)) return self.overflow norm_groups = [] local_sub_partitions_grad_groups = [] partition_id = dist.get_rank(group=self.dp_process_group) for i, group in enumerate(self.fp16_groups): #TODO RS: update get grad norm to support sub partitions norm_groups.append(get_grad_norm(group, mpu=self.mpu)) #RS: update free grads w.r.t. sub partitions #free gradients for all the parameters that are not updated by this process self.free_grad_in_param_list(self.params_not_local[i]) # create flat gradient partitions for parameters updated by this process local_grad_sub_partitions = self.get_flat_sub_partitions( comm_tensor_list=self.params_in_rank_sub_partitions[i][partition_id], comm_param_offsets=self.params_in_rank_sub_partitions_offsets[i] [partition_id], sub_partition_size=self.sub_partition_sizes[i], dtype=self.local_sub_partitions_of_fp32_groups[i][0].dtype, num_comm_intervals=self.num_comm_intervals_per_group[i], default_device=self.default_device) #RS: update all our local params with sub-partition grads for idx, sub_partition_param in enumerate(self.local_sub_partitions_of_fp32_groups[i]): sub_partition_param.grad = local_grad_sub_partitions[idx] #RS: update free grads for sub-partitions #release all the gradient since we have already created a necessary copy in dp_grad_partition self.free_grad_in_param_list( self.params_in_rank_sub_partitions[i][partition_id]) local_sub_partitions_grad_groups.append(local_grad_sub_partitions) #RS: update unscale/clip with sub partitions self.unscale_and_clip_grads(local_sub_partitions_grad_groups, norm_groups) self.optimizer.step() #RS: clear our sub partition grads #get rid of the fp32 gradients. Not needed anymore for group in self.local_sub_partitions_of_fp32_groups: for idx, sub_partition_param in enumerate(group): sub_partition_param.grad = None #group.grad = None #NOTE RS: removed norm_groups outer loop from original code, i don't think it's needed #RS: copy all sub-partition fp32 data to fp16 sub partitions # copy fp32 param data to fp16 partitions w.r.t. our local rank for fp16_all_sub_partitions, fp32_local_sub_partitions in zip(self.parallel_sub_partitioned_fp16_groups, self.local_sub_partitions_of_fp32_groups): for local_sub_partition_param_fp16, local_sub_partition_param_fp32 in zip(fp16_all_sub_partitions[partition_id], fp32_local_sub_partitions): if self.hack_first_step == True: local_sub_partition_param_fp32.data.copy_(local_sub_partition_param_fp16.data) else: local_sub_partition_param_fp16.data.copy_(local_sub_partition_param_fp32.data) self.hack_first_step = False #RS: all_gather/broadcast sub-partitions in separate comm calls #gather the updated weights from everyone for fp16_all_sub_partitions in self.parallel_comm_sub_partitioned_fp16_groups: for comm_id, sub_partitions in enumerate(fp16_all_sub_partitions): dist.all_gather(sub_partitions, sub_partitions[partition_id], group=self.dp_process_group) # TODO: we probably don't need this? just to be safe for i in range(len(norm_groups)): updated_params = _unflatten_dense_tensors(self.fp16_groups_flat[i], self.fp16_groups[i]) for p, q in zip(self.fp16_groups[i], updated_params): p.data = q.data return self.overflow def unscale_and_clip_grads(self, grad_groups_flat, norm_groups): total_norm = 0.0 for norm in norm_groups: total_norm += norm**2.0 total_norm = math.sqrt(total_norm) # compute combined scale factor for this group combined_scale = self.loss_scale if self.clip_grad > 0.: # norm is in fact norm*scale clip = ((total_norm / self.loss_scale) + 1e-6) / self.clip_grad if clip > 1: combined_scale = clip * self.loss_scale for grad in grad_groups_flat: if isinstance(grad, list): sub_partitions = grad for g in sub_partitions: g.data.mul_(1. / combined_scale) else: grad.data.mul_(1. / combined_scale) def backward(self, loss, retain_graph=False): self.loss_scaler.backward(loss.float(), retain_graph=retain_graph) def _update_scale(self, has_overflow=False): self.loss_scaler.update_scale(has_overflow) # Promote state so it can be retrieved or set via "fp16_optimizer_instance.state" def _get_state(self): return self.optimizer.state def _set_state(self, value): self.optimizer.state = value state = property(_get_state, _set_state) # Promote param_groups so it can be retrieved or set via "fp16_optimizer_instance.param_groups" # (for example, to adjust the learning rate) def _get_param_groups(self): return self.optimizer.param_groups def _set_param_groups(self, value): self.optimizer.param_groups = value param_groups = property(_get_param_groups, _set_param_groups) # Promote loss scale so it can be retrieved or set via "fp16_optimizer_instance.loss_scale" def _get_loss_scale(self): return self.loss_scaler.loss_scale def _set_loss_scale(self, value): self.loss_scaler.cur_scale = value loss_scale = property(_get_loss_scale, _set_loss_scale) cur_scale = property(_get_loss_scale, _set_loss_scale) # Return communication interval paddings for local rank and group def _get_local_group_paddings(self, group_index): local_rank = dist.get_rank(group=self.dp_process_group) sub_partition_indices = [ local_rank + (comm_idx * self.partition_count) for comm_idx in range(self.num_comm_intervals_per_group[group_index]) ] group_paddings = [ self.group_paddings[group_index][sub_idx] for sub_idx in sub_partition_indices ] return group_paddings # Return group tensor after removing paddings that are added for alignment to DP world size. # This method works on the assumption that each group contains sub partitions. def _get_groups_without_padding(self, groups_with_padding): groups_without_padding = [] for group_index, group in enumerate(groups_with_padding): group_paddings = self._get_local_group_paddings(group_index) lean_sub_partitions = [] for sub_partition, padding in zip(group, group_paddings): lean_length = sub_partition.numel() - padding lean_sub_partitions.append(sub_partition[:lean_length]) groups_without_padding.append(lean_sub_partitions) return groups_without_padding # Return optimizer state after removing paddings that are added for alignment. def _get_state_without_padding(self, state_with_padding, padding): lean_state = {} for key, value in state_with_padding.items(): if torch.is_tensor(value): lean_length = value.numel() - padding lean_state[key] = value[:lean_length] else: lean_state[key] = value return lean_state # Return base optimizer states. # This method assumes that each param group contains a single flattened tensor. def _get_base_optimizer_state(self): optimizer_groups_state = [] for group_index, group in enumerate(self.optimizer.param_groups): param_paddings = self._get_local_group_paddings(group_index) group_lean_state = [] for param_idx, param in enumerate(group['params']): lean_state = self._get_state_without_padding(self.optimizer.state[param], param_paddings[param_idx]) group_lean_state.append(lean_state) optimizer_groups_state.append(group_lean_state) return optimizer_groups_state def _rigid_state_dict(self): """ Returns a dict that can be loaded for continued training with same DP degree """ """ Returns a dict containing the current state of this :class:`FP16_Optimizer` instance. This dict contains attributes of :class:`FP16_Optimizer`, as well as the state_dict of the contained Pytorch optimizer. Example:: checkpoint = {} checkpoint['model'] = model.state_dict() checkpoint['optimizer'] = optimizer.state_dict() torch.save(checkpoint, "saved.pth") """ state_dict = {} state_dict['loss_scaler'] = self.loss_scaler state_dict['dynamic_loss_scale'] = self.dynamic_loss_scale state_dict['overflow'] = self.overflow state_dict['base_optimizer_state'] = self.optimizer.state_dict() state_dict[ 'local_sub_partitions_of_fp32_groups'] = self.local_sub_partitions_of_fp32_groups return state_dict def _elastic_state_dict(self): """ Returns a dict that can be loaded for elastic training with different DP degree """ state_dict = {} state_dict['loss_scaler'] = self.loss_scaler state_dict['dynamic_loss_scale'] = self.dynamic_loss_scale state_dict['overflow'] = self.overflow state_dict['base_optimizer_state'] = self._get_base_optimizer_state() state_dict['zero_stage'] = ZERO_OPTIMIZATION_OPTIMIZER_STATES state_dict['partition_count'] = self.partition_count state_dict['num_comm_intervals_per_group'] = self.num_comm_intervals_per_group # Remove paddings for DP alignment to enable loading for other alignment values fp32_groups_without_padding = self._get_groups_without_padding( self.local_sub_partitions_of_fp32_groups) state_dict['local_sub_partitions_of_fp32_groups'] = fp32_groups_without_padding return state_dict def state_dict(self): """ Returns a dict containing the current state of this :class:`FP16_Optimizer` instance. This dict contains attributes of :class:`FP16_Optimizer`, as well as the state_dict of the contained Pytorch optimizer. Example:: checkpoint = {} checkpoint['model'] = model.state_dict() checkpoint['optimizer'] = optimizer.state_dict() torch.save(checkpoint, "saved.pth") """ if self.elastic_checkpoint: return self._elastic_state_dict() return self._rigid_state_dict() # Extract the fp32 weights of the current rank from checkpoint by merging the # sub partitions of communication intervals across ranks. # Let sub_i_j = sub partition of rank i and comm interval j # For 2 ranks and 2 comm intervals, checkpoints (minus padding) are as follows: # rank 0 = [sub_0_0, sub_0_1] # rank 1 = [sub_1_0, sub_1_1] # Merge to get [sub_0_0, sub_1_0, sub_0_1, sub_1_1] => original un-padded flattened tensor. def _retrieve_group_sub_partition_weights(self, all_partition_fp32_weights, max_elems_per_comm): num_partitions = len(all_partition_fp32_weights) num_comm_intervals = len(all_partition_fp32_weights[0]) num_sub_partitions = num_partitions * num_comm_intervals all_sub_partition_weights = [None] * num_sub_partitions for rank, partition_weights in enumerate(all_partition_fp32_weights): for comm_idx, sub_partition_weights in enumerate(partition_weights): #all_sub_partition_weights.append(sub_partition_weights) sub_partition_idx = (comm_idx * num_partitions) + rank all_sub_partition_weights[sub_partition_idx] = sub_partition_weights flat_merged_weights = flatten_dense_tensors_sub_partition_aligned( tensor_list=all_sub_partition_weights, dp=dist.get_world_size(group=self.dp_process_group), max_elements_per_comm=max_elems_per_comm, pg=self.dp_process_group) comm_partitions, dp_sub_partitions, element_intervals, sub_partition_size, num_comm_intervals = \ self.get_data_parallel_sub_partitions( tensor=flat_merged_weights, max_elements_per_comm=max_elems_per_comm, world_size=dist.get_world_size(group=self.dp_process_group), dp_process_group=self.dp_process_group ) partition_id = dist.get_rank(group=self.dp_process_group) return [sub_partition for sub_partition in dp_sub_partitions[partition_id]] # Restore base optimizer fp32 weights from checkpoint by: # 1) Merging fp32 weights from checkpoints of all partitions # 2) Extracting fp32 weights for current partition from merged weights # 3) Using extracted weights to update base optimizer weights directly. def _restore_from_fp32_weights(self, all_state_dict): sub_partition_of_fp32_groups = [] for group_idx in range(len(self.local_sub_partitions_of_fp32_groups)): all_partition_fp32_weights = [ sd['local_sub_partitions_of_fp32_groups'][group_idx] for sd in all_state_dict ] max_elems_per_comm = self.max_elems_per_comm[group_idx] sub_partition_weights = self._retrieve_group_sub_partition_weights( all_partition_fp32_weights, max_elems_per_comm) sub_partition_of_fp32_groups.append(sub_partition_weights) for current_group, saved_group in zip(self.local_sub_partitions_of_fp32_groups, sub_partition_of_fp32_groups): for current_sub_part, saved_sub_part in zip(current_group, saved_group): current_sub_part.data.copy_(saved_sub_part.data) # Extract optimizer state for current partition from merged states of all partitions def _partition_base_optimizer_state(self, state_key, all_partition_states, max_elems_per_comm): if not torch.is_tensor(all_partition_states[0]): return all_partition_states[0] alignment = dist.get_world_size(group=self.dp_process_group) flat_merged_partitions = flatten_dense_tensors_sub_partition_aligned( tensor_list=all_partition_states, dp=dist.get_world_size(group=self.dp_process_group), max_elements_per_comm=max_elems_per_comm, pg=self.dp_process_group) comm_partitions, dp_sub_partitions, element_intervals, sub_partition_size, num_comm_intervals = \ self.get_data_parallel_sub_partitions( tensor=flat_merged_partitions, max_elements_per_comm=max_elems_per_comm, world_size=dist.get_world_size(group=self.dp_process_group), dp_process_group=self.dp_process_group ) partition_id = dist.get_rank(group=self.dp_process_group) return [sub_partition for sub_partition in dp_sub_partitions[partition_id]] # Compute the optimizer state partitions for the group by # 1) Merging state values across the previous partitioning. # 2) Repartition state values for the new partitioning # 3) Return state corresponding to local partition def _retrieve_group_optimizer_states(self, all_partition_states, max_elems_per_comm): merged_optimizer_states = {} num_partitions = len(all_partition_states) num_comm_intervals = len(all_partition_states[0]) num_sub_partitions = num_partitions * num_comm_intervals for rank, partition_state in enumerate(all_partition_states): for comm_idx, sub_partition_state in enumerate(partition_state): for key, value in sub_partition_state.items(): if not key in merged_optimizer_states.keys(): merged_optimizer_states[key] = [None] * num_sub_partitions sub_partition_idx = (comm_idx * num_partitions) + rank merged_optimizer_states[key][sub_partition_idx] = value group_optimizer_states = {} for key, value in merged_optimizer_states.items(): group_optimizer_states[key] = self._partition_base_optimizer_state( key, value, max_elems_per_comm) return group_optimizer_states # Restore base optimizer state from checkpoint by # 1) Merging optimizer state from checkpoints of all partitions # 2) Extracting optimizer state for current partition from the merged state # 3) Using the extracted value to directly update the base optimizer. def _restore_base_optimizer_state(self, state_dict_list): base_optimizer_group_states = [] for group_idx in range(len(self.optimizer.param_groups)): all_partition_group_states = [ sd['base_optimizer_state'][group_idx] for sd in state_dict_list ] max_elems_per_comm = self.max_elems_per_comm[group_idx] group_optimizer_states = self._retrieve_group_optimizer_states( all_partition_group_states, max_elems_per_comm) base_optimizer_group_states.append(group_optimizer_states) for group_idx, group in enumerate(self.optimizer.param_groups): for param_idx, param in enumerate(group['params']): for key, saved in base_optimizer_group_states[group_idx].items(): if torch.is_tensor(self.optimizer.state[param][key]): current = self.optimizer.state[param][key] current.data.copy_(saved[param_idx].data) else: self.optimizer.state[param][key] = saved # Restore base optimizer fp32 weights from ZeRO fp16 weights def _restore_from_fp16_weights(self): partition_id = dist.get_rank(group=self.dp_process_group) for fp16_partitions, fp32_partitions in zip(self.parallel_sub_partitioned_fp16_groups, self.local_sub_partitions_of_fp32_groups): for fp16_sub_partition, fp32_sub_partition in zip(fp16_partitions[partition_id], fp32_partitions): fp32_sub_partition.data.copy_(fp16_sub_partition.data) # Refresh the fp32 master params from the fp16 copies. def refresh_fp32_params(self): self._restore_from_fp16_weights() def _rigid_load_state_dict(self, state_dict, load_optimizer_states=True): # I think it should actually be ok to reload the optimizer before the model. self.loss_scaler = state_dict['loss_scaler'] self.dynamic_loss_scale = state_dict['dynamic_loss_scale'] self.overflow = state_dict['overflow'] if load_optimizer_states: self.optimizer.load_state_dict(state_dict['base_optimizer_state']) for curr_group, saved_group in zip(self.local_sub_partitions_of_fp32_groups, state_dict['local_sub_partitions_of_fp32_groups']): for curr_param, saved_param in zip(curr_group, saved_group): curr_param.data.copy_(saved_param.data) def _elastic_load_state_dict(self, state_dict_list, load_optimizer_states=True, load_from_fp32_weights=False): """ Loads a state_dict created by an earlier call to state_dict(). If ``fp16_optimizer_instance`` was constructed from some ``init_optimizer``, whose parameters in turn came from ``model``, it is expected that the user will call ``model.load_state_dict()`` before ``fp16_optimizer_instance.load_state_dict()`` is called. Example:: model = torch.nn.Linear(D_in, D_out).cuda().half() optimizer = torch.optim.SGD(model.parameters(), lr=1e-3) optimizer = FP16_Optimizer(optimizer, static_loss_scale = 128.0) ... checkpoint = torch.load("saved.pth") model.load_state_dict(checkpoint['model']) optimizer.load_state_dict(checkpoint['optimizer']) """ # I think it should actually be ok to reload the optimizer before the model. self.loss_scaler = state_dict_list[0]['loss_scaler'] self.dynamic_loss_scale = state_dict_list[0]['dynamic_loss_scale'] self.overflow = state_dict_list[0]['overflow'] if load_optimizer_states: self._restore_base_optimizer_state(state_dict_list) if load_from_fp32_weights: self._restore_from_fp32_weights(state_dict_list) else: self._restore_from_fp16_weights() def load_state_dict(self, state_dict_list, load_optimizer_states=True, load_from_fp32_weights=False): """ Loads a state_dict created by an earlier call to state_dict(). If ``fp16_optimizer_instance`` was constructed from some ``init_optimizer``, whose parameters in turn came from ``model``, it is expected that the user will call ``model.load_state_dict()`` before ``fp16_optimizer_instance.load_state_dict()`` is called. Example:: model = torch.nn.Linear(D_in, D_out).cuda().half() optimizer = torch.optim.SGD(model.parameters(), lr=1e-3) optimizer = FP16_Optimizer(optimizer, static_loss_scale = 128.0) ... checkpoint = torch.load("saved.pth") model.load_state_dict(checkpoint['model']) optimizer.load_state_dict(checkpoint['optimizer']) """ if self.elastic_checkpoint: self._elastic_load_state_dict(state_dict_list, load_optimizer_states, load_from_fp32_weights) else: self._rigid_load_state_dict( state_dict_list[dist.get_rank(group=self.dp_process_group)], load_optimizer_states) def _dump_optimizer_state(self, message): logger.info(f'{message}') for i, group in enumerate(self.optimizer.param_groups): for j, param in enumerate(group['params']): for key, value in self.optimizer.state[param].items(): t_stats = [ value.min(), value.max(), (value.max() - value.min()), value.mean() ] stats = [float(t) for t in t_stats] logger.info( f'group/param/key/min/max/delta/mean = {i}, {j}, {key}: {stats}')

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/LM/InstructGPT/run_PopQA.py

#!/usr/bin/python # -*- coding: UTF-8 -*- from tqdm import tqdm import argparse import os import time import json import torch import random import numpy as np import pandas as pd import openai openai.api_key = "YOUR_API_KEY" seed = 633 torch.backends.cudnn.deterministic = True random.seed(seed) np.random.seed(seed) torch.manual_seed(seed) torch.cuda.manual_seed_all(seed) print("Cuda:", torch.cuda.is_available()) print("pwd", os.getcwd()) from transformers import AutoTokenizer, AutoModelForCausalLM from util_clm import convert_model_to_int8_on_gpu import jsonlines def load_jsonlines(file): with jsonlines.open(file, "r") as jsonl_f: lst = [obj for obj in jsonl_f] return lst q_templates = { 22: "What is {}'s occupation?", 218: "In what city was {} born?", 91: "What genre is {}?", 257: "Who is the father of {}?", 182: "In what country is {}?", 164: "Who was the producer of {}?", 526: "Who was the director of {}?", 97: "What is {} the capital of?", 533: "Who was the screenwriter for {}?", 639: "Who was the composer of {}?", 472: "What color is {}?", 106: "What is the religion of {}?", 560: "What sport does {} play?", 484: "Who is the author of {}?", 292: "Who is the mother of {}?", 422: "What is the capital of {}?", } completion_template = ( "Q: {} A:" # "{}" # "Query: {}\nResult:" # "Q: {} A:" # "{} The answer is" ) genread_template = "Generate a background document from Wikipedia to answer the given question. {}" # This prompt comes from the GenRead paper def call_request(prompt, model, tokenizer, max_new_tokens=15): max_inpt_tokens = tokenizer.model_max_length if ( len(prompt) > tokenizer.model_max_length ): # conservative lower bound, since each token is at least 1 character inpts = tokenizer(prompt, return_tensors="pt") new_prompt = tokenizer.decode( inpts.input_ids[0, -(max_inpt_tokens - max_new_tokens) :] ) else: new_prompt = prompt # try to get a response from the model multiple times if theres a timeout while True: try: # if i > 0: # print("Retrying request") response = openai.Completion.create( model=model, prompt=new_prompt, temperature=0.0, max_tokens=max_new_tokens, logprobs=5, top_p=1, frequency_penalty=0.0, presence_penalty=0.0, ) break except Exception as e: # print(e) print("Timeout, trying again") time.sleep(1) pred = response["choices"][0]["text"] if pred.startswith("\n\n"): pred = pred[2:] pred = pred.split("\n")[0] return pred, response.to_dict_recursive() def call_model( prompt, model, tokenizer, device, max_new_tokens=15, model_max_length=None ): max_inpt_tokens = ( tokenizer.model_max_length if model_max_length is None else model_max_length ) inpts = tokenizer(prompt, return_tensors="pt").to(device) gen = model.generate( input_ids=inpts.input_ids[:, -(max_inpt_tokens - max_new_tokens) :], attention_mask=inpts.attention_mask[:, -(max_inpt_tokens - max_new_tokens) :], pad_token_id=tokenizer.eos_token_id, max_new_tokens=max_new_tokens, num_beams=1, do_sample=False, ) text = tokenizer.decode(gen[0]) actual_prompt = tokenizer.decode( inpts.input_ids[0, -(max_inpt_tokens - max_new_tokens) :] ) pred = text[len(actual_prompt) :] if pred.startswith("\n\n"): pred = pred[2:] pred = pred.split("\n")[0] return pred, text def clip_paragraph(text, eval_method): if eval_method in ["BM25", "genread"]: return text split = text.split(". ") return ". ".join(split[:-1]) + "." def get_few_shot_text_with_retrieval(row, retrieval_dict, eval_method): if eval_method == "vanilla": return completion_template.format(row.question) + " " + row.obj # retrieval_dict[row.id]["ctxs"][0] if row.question.replace("?", "").lower() not in retrieval_dict: print("missing retrieval") return completion_template.format(row.question) + " " + row.obj else: retrieval = retrieval_dict[row.question.replace("?", "").lower()]["ctxs"][0] retrieved_text = clip_paragraph(retrieval["text"], eval_method) return ( retrieved_text + "\n\n" + completion_template.format(row.question) + " " + row.obj ) def get_few_shot_text(row, eval_method): return completion_template.format(row.question) + " " + row.obj def get_genread_passage( question, genread_template, generate_function, max_new_tokens=150 ): prompt = genread_template.format(question) return generate_function(prompt, max_new_tokens=max_new_tokens)[0] def get_few_shot_examples_genread( knowledge, generate_function, n_examples, genread_template, is_templatedQA, max_new_tokens=150, ): if is_templatedQA: few_shot_examples = dict() all_pids = list(q_templates.keys()) examples_per_template = n_examples // (len(q_templates) - 1) for pid in all_pids: for row2 in ( knowledge[knowledge.prop_id == pid].sample(n=examples_per_template).iloc ): if pid not in few_shot_examples: few_shot_examples[pid] = [] generation = get_genread_passage( row2.question, genread_template, generate_function, max_new_tokens=max_new_tokens, ) few_shot_examples[pid].append( get_few_shot_text_with_retrieval( row2, {row2.question: {"ctxs": [{"id": -1, "text": generation}]}}, "genread", ) ) else: few_shot_examples = [] for row2 in knowledge.sample(n=n_examples + 1).iloc: generation = get_genread_passage( row2.question, genread_template, generate_function, max_new_tokens=max_new_tokens, ) few_shot_examples.append( get_few_shot_text_with_retrieval( row2, {row2.question: {"ctxs": [{"id": -1, "text": generation}]}}, "genread", ) ) return few_shot_examples def main(): parser = argparse.ArgumentParser() parser.add_argument("--model_name", type=str, default="text-davinci-002") parser.add_argument("--input_file", type=str) parser.add_argument("--alias", type=str) parser.add_argument("--n_examples", type=int, default=15) parser.add_argument( "--eval_method", type=str, default="contriever", choices=["vanilla", "BM25", "contriever", "genread"], ) parser.add_argument( "--ret_path", type=str, default=None, required=False, help="path to retrieved documents jsonl", ) parser.add_argument("--device", type=str, default="cuda") parser.add_argument("--max_new_tokens", type=int, default=15) parser.add_argument("--sample", type=int, default=0, help="if 0, use all examples") parser.add_argument( "--continue_from", type=str, help="path to previous results file" ) parser.add_argument("--int8bit", action="store_true") parser.add_argument( "--parallel", type=str, help="string of format 'i.n_workers' where i is the index of the worker", ) args = parser.parse_args() use_gpt3 = args.model_name in { "text-davinci-003", "text-davinci-002", "text-curie-001", "text-babbage-001", "text-ada-001", } if use_gpt3: tokenizer = AutoTokenizer.from_pretrained("gpt2") generate = lambda prompt, max_new_tokens: call_request( prompt, args.model_name, tokenizer, max_new_tokens=max_new_tokens ) else: gpt = args.model_name device = args.device tokenizer = AutoTokenizer.from_pretrained(gpt) tokenizer.pad_token = tokenizer.eos_token tokenizer.pad_token_id = tokenizer.eos_token_id if args.int8bit: model = convert_model_to_int8_on_gpu( AutoModelForCausalLM.from_pretrained(gpt), device ) else: model = AutoModelForCausalLM.from_pretrained(gpt).eval().to(device) if "opt" in args.model_name or args.model_name == "EleutherAI/gpt-neox-20b": generate = lambda prompt, max_new_tokens: call_model( prompt, model=model, tokenizer=tokenizer, device=device, max_new_tokens=max_new_tokens, model_max_length=2048, ) else: generate = lambda prompt, max_new_tokens: call_model( prompt, model=model, tokenizer=tokenizer, device=device, max_new_tokens=max_new_tokens, ) input_path = args.input_file knowledge = pd.read_csv(input_path, sep="\t") if args.continue_from is not None: results = pd.read_csv(args.continue_from, sep="\t") knowledge = knowledge[~knowledge.id.isin(results.id)] n = len(knowledge) if args.sample == 0 else args.sample sample = knowledge.sample(n=n, replace=False) if args.parallel is not None: worker_num, n_workers = map(int, args.parallel.split(".")) sample = sample.iloc[worker_num::n_workers] n_examples = args.n_examples is_templatedQA = True examples_per_template = n_examples // (len(q_templates) - 1) preds = [] prompts = [] accuracy = [] responses = [] if args.eval_method in ["BM25", "contriever"]: has_answer = [] retrieval_ids = [] with open(args.ret_path) as f: retrieval_dict = { json.loads(s)["question"]: json.loads(s) for s in f.readlines() } # print(retrieval_dict) if args.eval_method == "genread": genread_few_shot_examples = get_few_shot_examples_genread( knowledge, generate, n_examples, genread_template, is_templatedQA, max_new_tokens=150, ) has_answer = [] gen_passages = [] # main loop row_num = 0 for row in tqdm(sample.iloc, total=n): if row_num < 10000: row_num += 1 continue # get few shot examples text if n_examples == 0: few_shot_examples_text = "" else: few_shot_examples = [] if args.eval_method == "genread": if is_templatedQA: other_pids = list(q_templates.keys()) other_pids.remove(row.prop_id) few_shot_examples = [] for pid in other_pids: few_shot_examples.extend( random.sample( genread_few_shot_examples[pid], examples_per_template ) ) else: few_shot_examples = random.sample( [ ex for ex in genread_few_shot_examples if row.question not in ex ], n_examples, ) else: if is_templatedQA: other_pids = list(q_templates.keys()) other_pids.remove(row.prop_id) for pid in other_pids: for row2 in ( knowledge[knowledge.prop_id == pid] .sample(n=examples_per_template) .iloc ): few_shot_examples.append( get_few_shot_text_with_retrieval( row2, retrieval_dict, args.eval_method ) if args.eval_method in ["BM25", "contriever"] else get_few_shot_text(row2, args.eval_method) ) else: for row2 in ( knowledge[knowledge.question != row.question] .sample(n=n_examples) .iloc ): few_shot_examples.append( get_few_shot_text_with_retrieval( row2, retrieval_dict, args.eval_method ) if args.eval_method in ["BM25", "contriever"] else get_few_shot_text(row2, args.eval_method) ) np.random.shuffle(few_shot_examples) few_shot_examples_text = "\n\n".join(few_shot_examples) + "\n\n" # get prompt if args.eval_method == "vanilla": prompt = few_shot_examples_text + completion_template.format(row.question) elif args.eval_method in ["BM25", "contriever"]: query = row.question try: retrieval = retrieval_dict[query]["ctxs"][ 0 ] # retrieval_dict[row.id]["ctxs"][0] except: print( "No retrieval for", query, " Example query:", list(retrieval_dict.keys())[0], ) retrieval = {"text": "", "id": np.nan, "hasanswer": False} # retrieved_text = clip_paragraph( # retrieval["text"], eval_method=args.eval_method # ) retrieved_text = ( retrieval_dict[query]["ctxs"][0]["text"] + "\n\n" + retrieval_dict[query]["ctxs"][1]["text"] + "\n\n" + retrieval_dict[query]["ctxs"][2]["text"] ) retrieval_id = retrieval["id"] prompt = ( few_shot_examples_text + retrieved_text + "\n\n" + completion_template.format(row.question) ) has_answer.append(retrieval["hasanswer"]) retrieval_ids.append(retrieval_id) elif args.eval_method == "genread": generation = get_genread_passage( row.question, genread_template, generate, max_new_tokens=150 ) prompt = ( few_shot_examples_text + generation + "\n\n" + completion_template.format(row.question) ) gen_passages.append(generation) # generate response pred, response = generate(prompt, max_new_tokens=args.max_new_tokens) prompts.append(prompt) preds.append(pred) responses.append(response) # compute accuracy possible_answers = json.loads(row.possible_answers) is_correct = False genread_has_answer = False for pa in possible_answers: if pa in pred or pa.lower() in pred or pa.capitalize() in pred: is_correct = True if ( args.eval_method == "genread" and pa in response or pa.lower() in response or pa.capitalize() in response ): genread_has_answer = True accuracy.append(is_correct) if args.eval_method == "genread": has_answer.append(genread_has_answer) if len(preds) % 500 == 0: g = open("gpt3.txt", "a") temp_sample = sample.iloc[: len(preds)].copy() temp_sample["is_correct"] = accuracy # print(temp_sample.is_correct.mean()) g.write(str(temp_sample.is_correct.mean()) + "\n") g.flush() # save results intermittently if len(preds) % 100000 == 0: temp_sample = sample.iloc[: len(preds)].copy() temp_sample["pred"] = preds temp_sample["prompt"] = prompts temp_sample["generation"] = responses temp_sample["is_correct"] = accuracy if args.eval_method in ["BM25", "contriever"]: temp_sample["has_answer"] = has_answer temp_sample["retrieval_id"] = retrieval_ids if args.eval_method == "genread": temp_sample["has_answer"] = has_answer temp_sample["gen_passage"] = gen_passages model_name_alias = args.model_name.replace("/", "_") if not os.path.exists(f"results/temp/"): os.makedirs(f"results/temp/") worker_str = "" if args.parallel is None else f"-worker={args.parallel}" output_path = f"results/temp/model={model_name_alias}-input={args.alias}-method={args.eval_method}-shots={n_examples}-n={len(temp_sample)}{'_int8bit' if args.int8bit is True else ''}{worker_str}.csv" temp_sample.to_csv(output_path, index=False) sample["is_correct"] = accuracy sample["prompt"] = prompts sample["pred"] = preds sample["generation"] = responses if args.eval_method in ["BM25", "contriever"]: sample["has_answer"] = has_answer sample["retrieval_id"] = retrieval_ids if args.eval_method == "genread": sample["has_answer"] = has_answer sample["gen_passage"] = gen_passages print(sample.is_correct.mean()) model_name_alias = args.model_name.replace("/", "_") worker_str = "" if args.parallel is None else f"-worker={args.parallel}" sample.to_csv( f"results/model={model_name_alias}-input={args.alias}-method={args.eval_method}-shots={n_examples}-n={len(sample)}{'_int8bit' if args.int8bit is True else ''}{worker_str}.csv" ) if __name__ == "__main__": main()

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/LM/Flan-T5/arguments.py

# coding=utf-8 """argparser configuration""" import argparse import os import torch import deepspeed def add_model_config_args(parser: argparse.ArgumentParser): """Model arguments""" group = parser.add_argument_group("model", "model configuration") group.add_argument( "--model-config", type=str, default=None, help="the configuration of the base model", ) group.add_argument( "--cpu-optimizer", action="store_true", help="Run optimizer on CPU" ) group.add_argument( "--cpu_torch_adam", action="store_true", help="Use Torch Adam as optimizer on CPU.", ) return parser def add_fp16_config_args(parser: argparse.ArgumentParser): """Mixed precision arguments.""" group = parser.add_argument_group("fp16", "fp16 configurations") group.add_argument("--fp16", action="store_true", help="Run model in fp16 mode") group.add_argument( "--fp32-embedding", action="store_true", help="embedding in fp32" ) group.add_argument( "--fp32-layernorm", action="store_true", help="layer norm in fp32" ) group.add_argument( "--fp32-tokentypes", action="store_true", help="embedding token types in fp32" ) group.add_argument( "--fp32-allreduce", action="store_true", help="all-reduce in fp32" ) group.add_argument( "--hysteresis", type=int, default=2, help="hysteresis for dynamic loss scaling" ) group.add_argument( "--loss-scale", type=float, default=None, help="Static loss scaling, positive power of 2 " "values can improve fp16 convergence. If None, dynamic loss scaling is used.", ) group.add_argument( "--loss-scale-window", type=float, default=1000, help="Window over which to raise/lower dynamic scale", ) group.add_argument( "--min-scale", type=float, default=1, help="Minimum loss scale for dynamic loss scale", ) return parser def add_training_args(parser: argparse.ArgumentParser): """Training arguments.""" group = parser.add_argument_group("train", "training configurations") group.add_argument("--do-train", action="store_true", help="whether do training") group.add_argument("--do-valid", action="store_true", help="whether do validation") group.add_argument("--do-valid-and-eval", action="store_true") group.add_argument("--do-eval", action="store_true", help="whether do testing") group.add_argument( "--do-infer", action="store_true", help="whether do inference (testing without labels)", ) group.add_argument( "--train-ratio", type=float, default=1.0, help="the ratio of the training set used for training", ) group.add_argument( "--train-num", type=int, default=-1, help="the number of training samples, -1 for all sample", ) group.add_argument( "--dev-ratio", type=float, default=1.0, help="the ratio of the training set used for validation", ) group.add_argument( "--dev-num", type=int, default=-1, help="the number of validation samples, -1 for all sample", ) group.add_argument( "--test-ratio", type=float, default=1.0, help="the ratio of the training set used for testing", ) group.add_argument( "--test-num", type=int, default=-1, help="the number of testing samples, -1 for all sample", ) group.add_argument("--epochs", type=int, default=1, help="the epochs for training") group.add_argument( "--batch-size", type=int, default=4, help="Data Loader batch size" ) group.add_argument( "--dev-batch-size", type=int, default=2, help="Data Loader batch size" ) group.add_argument( "--gradient-accumulation-steps", type=int, default=1, help="gradient accumulation steps", ) group.add_argument( "--weight-decay", type=float, default=0.01, help="weight decay coefficient for L2 regularization", ) group.add_argument( "--checkpoint-activations", action="store_true", help="checkpoint activation to allow for training " "with larger models and sequences", ) group.add_argument( "--checkpoint-num-layers", type=int, default=1, help="chunk size (number of layers) for checkpointing", ) group.add_argument( "--num-checkpoints", type=int, default=24, help="For activation checkpointing" ) group.add_argument( "--deepspeed-activation-checkpointing", action="store_true", help="uses activation checkpointing from deepspeed", ) group.add_argument("--clip-grad", type=float, default=1.0, help="gradient clipping") group.add_argument( "--train-iters", type=int, default=1000000, help="total number of iterations to train over all training runs", ) group.add_argument("--log-interval", type=int, default=100, help="report interval") group.add_argument( "--max-save", type=int, default=-1, help="max checkpoints to save" ) group.add_argument("--seed", type=int, default=1234, help="random seed") group.add_argument("--few-data-num", type=int, default=None) group.add_argument("--few-data-names", type=str, default=None) group.add_argument("--data-aug", type=int, default=None) group.add_argument("--rand-real-label", action="store_true") group.add_argument("--rand-pseudo-label", action="store_true") # Learning rate. group.add_argument( "--lr-decay-iters", type=int, default=None, help="number of iterations to decay LR over," " If None defaults to `--train-iters`*`--epochs`", ) group.add_argument( "--lr-decay-style", type=str, default="linear", choices=["constant", "linear", "cosine", "exponential", "noam"], help="learning rate decay function", ) group.add_argument("--lr", type=float, default=1.0e-4, help="initial learning rate") group.add_argument( "--warmup", type=float, default=0.0, help="percentage of data to warmup on (.01 = 1% of all " "training iters). Default 0.01", ) group.add_argument("--warmup-iter", type=int, default=0) # save group.add_argument( "--save", type=str, default=None, help="Output directory to save checkpoints to.", ) group.add_argument( "--save-interval", type=int, default=5000, help="number of iterations between saves", ) group.add_argument( "--no-save-optim", action="store_true", help="Do not save current optimizer." ) # load group.add_argument( "--load", type=str, default=None, help="Path to a directory containing a model checkpoint.", ) group.add_argument( "--load-oprimizer-states", action="store_true", help="whether to load optimizer states", ) group.add_argument( "--load-lr-scheduler-states", action="store_true", help="whether to load learning rate scheduler states", ) group.add_argument( "--no-load-optim", action="store_true", help="Do not load optimizer when loading checkpoint.", ) group.add_argument( "--log-file", type=str, default=None, help="the path to save log.txt file" ) # distributed training args group.add_argument( "--distributed-backend", default="nccl", help="which backend to use for distributed training. One of [gloo, nccl]", ) group.add_argument( "--local_rank", type=int, default=None, help="local rank passed from distributed launcher", ) return parser def add_prompt_args(parser: argparse.ArgumentParser): group = parser.add_argument_group("prompt", "prompt configurations") group.add_argument( "--load_prompt", type=str, default=None, help="the path to load prompt from" ) group.add_argument( "--prompt-tune", action="store_true", help="whether to do prompt tuning" ) group.add_argument( "--prompt-config", type=str, default=None, help="the path of the prompt configuration", ) group.add_argument( "--save-prompt-only", action="store_true", help="whether to save the prompt only. If true, only prompts will be saved otherwise, " "the whole model together with the prompt will be saved.", ) return parser def add_evaluation_args(parser: argparse.ArgumentParser): """Evaluation arguments.""" group = parser.add_argument_group("validation", "validation configurations") group.add_argument( "--eval-batch-size", type=int, default=None, help="Data Loader batch size for evaluation datasets. Defaults to `--batch-size`", ) group.add_argument( "--eval-iters", type=int, default=100, help="number of iterations to run for evaluation validation/test for", ) group.add_argument( "--eval-interval", type=int, default=1000, help="interval between running evaluation on validation set", ) group.add_argument("--eval-per-prompt", action="store_true") group.add_argument("--no-norm-cand-loss", action="store_true") return parser def add_text_generate_args(parser: argparse.ArgumentParser): """Text generate arguments.""" group = parser.add_argument_group("Text generation", "configurations") group.add_argument("--sampling", action="store_true") group.add_argument( "--temperature", type=float, default=1.2, help="The temperature of sampling." ) group.add_argument("--top_p", type=float, default=0.9, help="Top-p sampling.") group.add_argument("--top_k", type=int, default=0, help="Top-k sampling.") group.add_argument( "--max-generation-length", type=int, default=64, help="The maximum sequence length to generate.", ) group.add_argument( "--min-generation-length", type=int, default=0, help="The minimum sequence length to generate.", ) group.add_argument( "--num-beams", type=int, default=1, help="The beam number of beam search." ) group.add_argument( "--no-repeat-ngram-size", type=int, default=0, help="The n-gram whose length is less than this option will appear at most once in the whole dialog.", ) group.add_argument( "--repetition-penalty", type=float, default=1, help="Repetition penalty, to prevent repeated words.", ) group.add_argument( "--early-stopping", action="store_true", help="Early-stopping while generating." ) group.add_argument( "--length-penalty", type=float, default=0, help="Length penalty, to prevent short generation.", ) group.add_argument( "--rule-path", type=str, default=None, help="The directory that contains hand-written rules.", ) return parser def add_data_args(parser: argparse.ArgumentParser): """Train/valid/test data arguments.""" group = parser.add_argument_group("data", "data configurations") group.add_argument( "--model-parallel-size", type=int, default=1, help="size of the model parallel." ) group.add_argument( "--data-path", nargs="+", type=str, default=None, help="Path to combined dataset to split.", ) group.add_argument( "--data-ext", type=str, default=".json", help="the extension of the data file" ) group.add_argument( "--data-name", type=str, default=None, help="the name of the dataset" ) group.add_argument( "--data-names", type=str, default=None, help="the name of the dataset" ) group.add_argument( "--data-prefix", type=str, default=None, help="the prefix to add before each data sample", ) group.add_argument( "--num-workers", type=int, default=2, help="Number of workers to use for dataloading", ) group.add_argument( "--tokenizer-path", type=str, default="tokenizer.model", help="path used to save/load sentencepiece tokenization models", ) group.add_argument( "--enc-seq-length", type=int, default=512, help="Maximum sequence length to process", ) group.add_argument( "--dec-seq-length", type=int, default=256, help="Maximum sequence length to process", ) group.add_argument("--pad-token", type=str, default="<pad>") group.add_argument("--FiD", action="store_true") group.add_argument( "--passage_num", type=int, default=10, help="Number of passages to use for FiD" ) return parser def add_flan_args(parser: argparse.ArgumentParser): group = parser.add_argument_group("flan", "data configurations") group.add_argument("--flan-sample", action="store_true") group.add_argument("--flan-sample-max", type=float, default=1000000) return parser def add_debug_args(parser: argparse.ArgumentParser): group = parser.add_argument_group("debug", "data configurations") group.add_argument("--debug-option", type=int, default=-1) group.add_argument("--shuff-cand-idx", action="store_true") return parser def get_args(): """Parse all the args.""" parser = argparse.ArgumentParser(description="PyTorch BERT Model") parser = add_model_config_args(parser) parser = add_fp16_config_args(parser) parser = add_training_args(parser) parser = add_evaluation_args(parser) parser = add_data_args(parser) parser = add_prompt_args(parser) parser = add_text_generate_args(parser) parser = add_flan_args(parser) parser = add_debug_args(parser) # Include DeepSpeed configuration arguments parser = deepspeed.add_config_arguments(parser) args = parser.parse_args() if not args.data_path: print("WARNING: No training data specified") args.cuda = torch.cuda.is_available() args.rank = int(os.getenv("RANK", "0")) args.world_size = int(os.getenv("WORLD_SIZE", "1")) args.local_rank = int(os.getenv("LOCAL_RANK", "0")) args.model_parallel_size = min(args.model_parallel_size, args.world_size) if args.rank == 0: print( "using world size: {} and model-parallel size: {} ".format( args.world_size, args.model_parallel_size ) ) args.dynamic_loss_scale = False if args.loss_scale is None: args.dynamic_loss_scale = True if args.rank == 0: print(" > using dynamic loss scaling") if args.data_path is not None and len(args.data_path) == 1: args.data_path = args.data_path[0] return args

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/LM/Flan-T5/learning_rates.py

# coding=utf-8 """PyTorch DataLoader for TFRecords""" import torch from torch.optim.lr_scheduler import _LRScheduler import math class AnnealingLR(_LRScheduler): """Anneals the learning rate from start to zero along a cosine curve.""" DECAY_STYLES = ['linear', 'cosine', 'exponential', 'constant', 'None', 'noam'] def __init__(self, optimizer, start_lr, warmup_iter, num_iters, decay_style=None, last_iter=-1, gradient_accumulation_steps=1): self.optimizer = optimizer self.start_lr = start_lr self.warmup_iter = (warmup_iter // gradient_accumulation_steps) + 1 self.num_iters = last_iter + 1 self.end_iter = num_iters self.gradient_accumulation_steps = gradient_accumulation_steps self.decay_style = decay_style.lower() if isinstance(decay_style, str) else None self.step(self.num_iters) if torch.distributed.get_rank() == 0: print('learning rate decaying', decay_style) def get_lr(self): # https://openreview.net/pdf?id=BJYwwY9ll pg. 4 if self.warmup_iter > 0 and self.num_iters <= self.warmup_iter: if self.decay_style != self.DECAY_STYLES[5]: return float(self.start_lr) * self.num_iters / self.warmup_iter else: return float(self.start_lr) / math.sqrt(self.warmup_iter) * self.num_iters / self.warmup_iter #* self.num_iters / self.warmup_iter / math.sqrt(self.warmup_iter) else: if self.decay_style == self.DECAY_STYLES[0]: return self.start_lr*((self.end_iter-(self.num_iters-self.warmup_iter))/self.end_iter) elif self.decay_style == self.DECAY_STYLES[1]: return self.start_lr / 2.0 * (math.cos(math.pi * (self.num_iters - self.warmup_iter) / self.end_iter) + 1) elif self.decay_style == self.DECAY_STYLES[2]: #TODO: implement exponential decay return self.start_lr elif self.decay_style == self.DECAY_STYLES[5]: return self.start_lr / math.sqrt(self.num_iters + 1) else: return self.start_lr def step(self, step_num=None): if step_num is None: step_num = self.num_iters + 1 self.num_iters = step_num new_lr = self.get_lr() for group in self.optimizer.param_groups: group['lr'] = new_lr def state_dict(self): sd = { 'start_lr': self.start_lr, 'warmup_iter': self.warmup_iter, 'num_iters': self.num_iters, 'decay_style': self.decay_style, 'end_iter': self.end_iter } return sd def load_state_dict(self, sd): self.start_lr = sd['start_lr'] self.warmup_iter = sd['warmup_iter'] self.num_iters = sd['num_iters'] self.end_iter = sd['end_iter'] self.decay_style = sd['decay_style'] self.step(self.num_iters)

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/LM/Flan-T5/utils.py

# coding=utf-8 """Utilities for logging and serialization""" import os import random import numpy as np import torch from fp16 import FP16_Optimizer import mpu import deepspeed from apex.optimizers import FusedAdam as Adam from fp16 import FP16_Module from fp16 import FP16_Optimizer from learning_rates import AnnealingLR from model import EncDecModel, EncDecConfig from model import enc_dec_get_params_for_weight_decay_optimization, enc_dec_get_params_for_prompt_optimization from model import DistributedDataParallel as DDP def print_rank_0(message): if torch.distributed.is_initialized(): if torch.distributed.get_rank() == 0: print(message, flush=True) else: print(message, flush=True) def print_args(args): """Print arguments.""" print('arguments:', flush=True) for arg in vars(args): dots = '.' * (29 - len(arg)) print(' {} {} {}'.format(arg, dots, getattr(args, arg)), flush=True) def save_rank_0(args, message): if torch.distributed.is_initialized(): if torch.distributed.get_rank() == 0: with open(args.log_file, "a") as f: f.write(message + "\n") f.flush() else: with open(args.log_file, "a") as f: f.write(message + "\n") f.flush() def save_preds_t0(args, name, prompt_names, step, all_res_prompt, all_preds_prompt, all_labels_prompt): s = np.mean([np.mean([vv for vv in v.values()]) for v in all_res_prompt]) if torch.distributed.is_initialized(): if torch.distributed.get_rank() == 0: os.makedirs(os.path.join(args.save, "preds", name), exist_ok=True) with open(os.path.join(args.save, "preds", name, "{:.2f}_{}.txt".format(s, step)), "w") as f: f.write(str(all_res_prompt) + "\n") for pid in range(len(prompt_names)): f.write("\n" + str(prompt_names[pid]) + "\n") for p, l in zip(all_preds_prompt[pid], all_labels_prompt[pid]): f.write(str(p) + "\t\t" + str(l) + "\n") def save_preds_prompts(args, name, dataset, step, res, all_preds_prompts, all_labels_prompts): s = np.mean([v for v in res[0].values()]) if torch.distributed.is_initialized(): if torch.distributed.get_rank() == 0: os.makedirs(os.path.join(args.save, "preds", name), exist_ok=True) with open(os.path.join(args.save, "preds", name, "{:.2f}_{}.txt".format(s, step)), "w") as f: f.write(str(res) + "\n") for pid in dataset.all_data[name]["prompt_ids"]: f.write("\n" + str(dataset.all_data[name]["prompt_templates"][pid]) + "\n") for p, l in zip(all_preds_prompts[pid], all_labels_prompts[pid]): f.write(str(p) + "\t\t" + str(l) + "\n") def save_preds(args, name, step, res, all_preds, all_labels): s = np.mean([v for v in res[0].values()]) if torch.distributed.is_initialized(): if torch.distributed.get_rank() == 0: os.makedirs(os.path.join(args.save, "preds", name), exist_ok=True) with open(os.path.join(args.save, "preds", name, "{:.2f}_{}.txt".format(s, step)), "w") as f: f.write(str(res) + "\n") for p, l in zip(all_preds, all_labels): f.write(str(p) + "\t\t" + str(l) + "\n") def get_model(args, vocab_size, prompt_config=None): """Build the model.""" print_rank_0('building Enc-Dec model ...') config = EncDecConfig.from_json_file(args.model_config) config.vocab_size = vocab_size model = EncDecModel(config, parallel_output=True, checkpoint_activations=args.checkpoint_activations, checkpoint_num_layers=args.checkpoint_num_layers, prompt_config=prompt_config, args=args) if mpu.get_data_parallel_rank() == 0: print(' > number of parameters on model parallel rank {}: {}'.format( mpu.get_model_parallel_rank(), sum([p.nelement() for p in model.parameters()])), flush=True) # To prevent OOM for model sizes that cannot fit in GPU memory in full precision if args.deepspeed and args.fp16: model.half() # GPU allocation. model.cuda(torch.cuda.current_device()) if args.prompt_tune and prompt_config["init_scratch"]: model.init_prompt_embeds() # Fp16 conversion. if args.fp16: model = FP16_Module(model) # Wrap model for distributed training. model = DDP(model) return model def get_optimizer(model, args, prompt_config=None): """Set up the optimizer.""" # Build parameter groups (weight decay and non-decay). while isinstance(model, (DDP, FP16_Module)): model = model.module if args.prompt_tune and prompt_config["fix_model"]: param_groups = enc_dec_get_params_for_prompt_optimization(model) else: param_groups = enc_dec_get_params_for_weight_decay_optimization(model) # Add model parallel attribute if it is not set. for param_group in param_groups: for param in param_group['params']: if not hasattr(param, 'model_parallel'): param.model_parallel = False if args.cpu_optimizer: if args.cpu_torch_adam: cpu_adam_optimizer = torch.optim.Adam else: from deepspeed.ops.adam import DeepSpeedCPUAdam cpu_adam_optimizer = DeepSpeedCPUAdam optimizer = cpu_adam_optimizer(param_groups, lr=args.lr, weight_decay=args.weight_decay) else: # Use FusedAdam. optimizer = Adam(param_groups, lr=args.lr, weight_decay=args.weight_decay) print(f'Optimizer = {optimizer.__class__.__name__}') if args.deepspeed: # fp16 wrapper is not required for DeepSpeed. return optimizer # Wrap into fp16 optimizer. if args.fp16: optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale, dynamic_loss_scale=args.dynamic_loss_scale, dynamic_loss_args={ 'scale_window': args.loss_scale_window, 'min_scale': args.min_scale, 'delayed_shift': args.hysteresis}) if torch.distributed.get_rank() == 0: print(optimizer.param_groups) return optimizer def get_learning_rate_scheduler(optimizer, args): """Build the learning rate scheduler.""" # Add linear learning rate scheduler. if args.lr_decay_iters is not None: num_iters = args.lr_decay_iters else: num_iters = args.train_iters num_iters = max(1, num_iters) init_step = -1 if args.warmup_iter > 0: warmup_iter = args.warmup_iter else: warmup_iter = args.warmup * num_iters lr_scheduler = AnnealingLR(optimizer, start_lr=args.lr, warmup_iter=warmup_iter, num_iters=num_iters, decay_style=args.lr_decay_style, last_iter=init_step, gradient_accumulation_steps=args.gradient_accumulation_steps) return lr_scheduler def setup_model_and_optimizer(args, vocab_size, ds_config, prompt_config=None, set_optim=True): """Setup model and optimizer.""" model = get_model(args, vocab_size, prompt_config) if set_optim: optimizer = get_optimizer(model, args, prompt_config) lr_scheduler = get_learning_rate_scheduler(optimizer, args) else: optimizer, lr_scheduler = None, None if args.deepspeed: print_rank_0("DeepSpeed is enabled.") model, optimizer, _, lr_scheduler = deepspeed.initialize( model=model, optimizer=optimizer, args=args, lr_scheduler=lr_scheduler, mpu=mpu, dist_init_required=False, config_params=ds_config ) print(args.load) if args.load is not None: args.iteration = load_checkpoint(model, optimizer, lr_scheduler, args, prompt_config) else: args.iteration = 0 return model, optimizer, lr_scheduler def set_deepspeed_activation_checkpointing(args): deepspeed.checkpointing.configure(mpu, deepspeed_config=args.deepspeed_config, num_checkpoints=args.num_checkpoints) mpu.checkpoint = deepspeed.checkpointing.checkpoint mpu.get_cuda_rng_tracker = deepspeed.checkpointing.get_cuda_rng_tracker mpu.model_parallel_cuda_manual_seed = deepspeed.checkpointing.model_parallel_cuda_manual_seed def initialize_distributed(args): """Initialize torch.distributed.""" # Manually set the device ids. device = args.rank % torch.cuda.device_count() if args.local_rank is not None: device = args.local_rank torch.cuda.set_device(device) # Call the init process deepspeed.init_distributed() # Set the model-parallel / data-parallel communicators. mpu.initialize_model_parallel(args.model_parallel_size) # Optional DeepSpeed Activation Checkpointing Features if args.deepspeed and args.deepspeed_activation_checkpointing: set_deepspeed_activation_checkpointing(args) def set_random_seed(seed): """Set random seed for reproducability.""" if seed is not None and seed > 0: random.seed(seed) np.random.seed(seed) torch.manual_seed(seed) mpu.model_parallel_cuda_manual_seed(seed) def save_checkpoint(iteration, model, optimizer, lr_scheduler, args, save_dir=None): """Save a model checkpoint.""" save_ds_checkpoint(iteration, model, args, save_dir) # Wait so everyone is done (necessary) torch.distributed.barrier() # And update the latest iteration if torch.distributed.get_rank() == 0: tracker_filename = os.path.join(args.save if save_dir is None else save_dir, 'latest_checkpointed_iteration.txt') with open(tracker_filename, 'w') as f: f.write(str(iteration)) # Wait so everyone is done (not necessary) torch.distributed.barrier() def save_ds_checkpoint(iteration, model, args, save_dir=None): """Save a model checkpoint.""" sd = {} sd['iteration'] = iteration if args.save_prompt_only: prompt = model.module.module.module.get_prompt_embeds() save_prompt(args.save if save_dir is None else save_dir, iteration, prompt["encoder"]) else: model.save_checkpoint(args.save if save_dir is None else save_dir, str(iteration), client_state = sd, save_zero=False) def save_prompt(save_dir, iteration, prompt_embeds): save_path = os.path.join(save_dir, "prompt-{}.pt".format(iteration)) if torch.distributed.get_rank() == 0: torch.save(prompt_embeds, save_path) def get_checkpoint_iteration(args): # Read the tracker file and set the iteration. tracker_filename = os.path.join(args.load, 'latest_checkpointed_iteration.txt') if not os.path.isfile(tracker_filename): print_rank_0('WARNING: could not find the metadata file {} '.format( tracker_filename)) print_rank_0(' will not load any checkpoints and will start from ' 'random') return 0, False, False iteration = 0 release = False with open(tracker_filename, 'r') as f: metastring = f.read().strip() try: iteration = int(metastring) except ValueError: release = metastring == 'release' if not release: print_rank_0('ERROR: Invalid metadata file {}. Exiting'.format( tracker_filename)) exit() assert iteration > 0 or release, 'error parsing metadata file {}'.format( tracker_filename) return iteration, release, True def load_prompt(load_dir): prompt = torch.load(load_dir, map_location=lambda storage, loc: storage) return prompt def load_checkpoint(model, optimizer, lr_scheduler, args, prompt_config=None): """Load a model checkpoint.""" iteration, release, success = get_checkpoint_iteration(args) if not success: return 0 mp_rank = mpu.get_model_parallel_rank() checkpoint_name = os.path.join(args.load, str(iteration), 'mp_rank_{:02d}'.format(mp_rank) + '_model_states.pt') if not os.path.exists(checkpoint_name): print('Client provided checkpoint load path: {} does not exist ... skip checkpoint load'.format(checkpoint_name)) if mpu.get_data_parallel_rank() == 0: print("Unable to load checkpoint.") return iteration print('loading checkpoint: {}'.format(checkpoint_name)) sd = torch.load(checkpoint_name, map_location=lambda storage, loc: storage) if args.prompt_tune: load_prompt_path = prompt_config.get("load_prompt") if load_prompt_path is not None and len(load_prompt_path) > 0: prompt_embeds = load_prompt(load_prompt_path) sd["module"]["encoder.prompt_embeds.weight"] = prompt_embeds model.module.load_state_dict(sd["module"], strict=False) iteration = sd['iteration'] torch.distributed.barrier() if mpu.get_data_parallel_rank() == 0: print(' successfully loaded {}'.format(checkpoint_name)) return iteration

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/LM/Flan-T5/generation_utils.py

# coding=utf-8 import os import torch import torch.nn.functional as F from collections import defaultdict from tokenization_t5 import EncDecTokenizer class BeamHypotheses(object): def __init__( self, num_beams, max_length, length_penalty, early_stopping, tokenizer=None ): """ Initialize n-best list of hypotheses. """ self.max_length = max_length - 1 # ignoring bos_token self.length_penalty = length_penalty self.early_stopping = early_stopping self.num_beams = num_beams self.length_fact = [] self.beams = [] self.worst_score = 1e9 self.raw_worst_score = 1e9 self.tokenizer = tokenizer def __len__(self): """ Number of hypotheses in the list. """ return len(self.beams) def add(self, hyp, sum_logprobs): """ Add a new hypothesis to the list. """ score = sum_logprobs / len(hyp) ** self.length_penalty if len(self) < self.num_beams or score > self.worst_score: self.beams.append((score, hyp)) self.length_fact.append(len(hyp) ** self.length_penalty) if len(self) > self.num_beams: sorted_scores = sorted( [(s, idx, _) for idx, (s, _) in enumerate(self.beams)] ) del self.beams[sorted_scores[0][1]] self.worst_score = sorted_scores[1][0] self.raw_worst_score = self.worst_score * ( len(sorted_scores[1][2]) ** self.length_penalty ) else: self.worst_score = min(score, self.worst_score) self.raw_worst_score = sum_logprobs def is_done(self, best_sum_logprobs, cur_len): """ If there are enough hypotheses and that none of the hypotheses being generated can become better than the worst one in the heap, then we are done with this sentence. """ if len(self) < self.num_beams: return False elif self.early_stopping: return True else: cur_score = best_sum_logprobs / cur_len ** self.length_penalty ret = self.worst_score >= cur_score return ret def construct_antonym_dict(args): if args.rule_path is None: return None with open(os.path.join(args.rule_path, "./antonym/antonym.txt"), "r") as f: data = f.read().split("\n") data = [eval(item) for item in data if item] antonym_dict = defaultdict(list) for first, second in data: antonym_dict[first].append(second) antonym_dict[second].append(first) return antonym_dict def calc_banned_antonym_words_ids(input_tokens, tokenizer, antonym_dict): if antonym_dict is None: return [] antonym_words = [set()] * len(input_tokens) # only consider tokens occurring in current sentence for idx, tokens in enumerate(input_tokens): for word in tokenizer.convert_ids_to_tokens(reversed(tokens.tolist())): if word == "<sep>": break antonym_words[idx].update( tokenizer.convert_tokens_to_ids(antonym_dict[word]) ) return [list(tokens) for tokens in antonym_words] def calc_banned_ngram_tokens( prev_input_ids, num_hypos: int, no_repeat_ngram_size: int, tokenizer: EncDecTokenizer, ) -> None: """Copied from fairseq for no_repeat_ngram in beam_search""" generated_ngrams = [{} for _ in range(num_hypos)] for idx in range(num_hypos): gen_words = prev_input_ids[idx] generated_ngram = generated_ngrams[idx] for ngram in zip(*[gen_words[i:] for i in range(no_repeat_ngram_size)]): ngram = tuple(ngram) generated_ngram[ngram] = generated_ngram.get(ngram, set()) | set([ngram]) def _get_generated_ngrams(hypo_idx): # Before decoding the next token, prevent decoding of ngrams that have already appeared cur_len = len(prev_input_ids[hypo_idx]) generated_ngram_idx = [] for prefix_len in range(no_repeat_ngram_size): ngram_words = tuple(prev_input_ids[hypo_idx][cur_len - prefix_len :]) generated_ngram_words = generated_ngrams[hypo_idx].get(ngram_words, []) generated_ngram_idx += tokenizer.convert_tokens_to_ids( generated_ngram_words ) return generated_ngram_idx banned_tokens = [_get_generated_ngrams(hypo_idx) for hypo_idx in range(num_hypos)] return banned_tokens def top_k_logits(logits, top_k=0, top_p=0.0, filter_value=-10000): # This function has been mostly taken from huggingface conversational ai code at # https://medium.com/huggingface/how-to-build-a-state-of-the-art-conversational-ai-with-transfer-learning-2d818ac26313 if top_k > 0: # Remove all tokens with a probability less than the last token of the top-k indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None] logits[indices_to_remove] = filter_value batch_size = logits.size()[0] if top_p > 0.0: # logits : (batch_size, vocab_size) logits = logits.view(batch_size, -1).contiguous() # logits : (batch_size, vocab_size) for logit in logits: # logit: (vocab_size) sorted_logits, sorted_indices = torch.sort(logit, descending=True) cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1) # Remove tokens with cumulative probability above the threshold sorted_indices_to_remove = cumulative_probs > top_p # Shift the indices to the right to keep also the first token above the threshold sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[ ..., :-1 ].clone() sorted_indices_to_remove[..., 0] = 0 indices_to_remove = sorted_indices[sorted_indices_to_remove] logit[indices_to_remove] = filter_value logits = logits.view(batch_size, -1).contiguous() return logits def calc_banned_bad_words_ids(prev_input_ids, bad_words_ids): banned_tokens = [] def _tokens_match(prev_tokens, tokens): if len(tokens) == 0: # if bad word tokens is just one token always ban it return True if len(tokens) > len(prev_input_ids): # if bad word tokens are longer then prev input_ids they can't be equal return False if prev_tokens[-len(tokens) :] == tokens: # if tokens match return True else: return False for prev_input_ids_slice in prev_input_ids: banned_tokens_slice = [] for banned_token_seq in bad_words_ids: assert ( len(banned_token_seq) > 0 ), "Banned words token sequences {} cannot have an empty list".format( bad_words_ids ) if ( _tokens_match(prev_input_ids_slice.tolist(), banned_token_seq[:-1]) is False ): # if tokens do not match continue continue banned_tokens_slice.append(banned_token_seq[-1]) banned_tokens.append(banned_tokens_slice) return banned_tokens def enforce_repetition_penalty_( tokenizer, lprobs, batch_size, num_beams, prev_output_tokens, repetition_penalty ): """repetition penalty (from CTRL paper https://arxiv.org/abs/1909.05858).""" for i in range(batch_size * num_beams): for previous_token in set(prev_output_tokens[i].tolist()): if previous_token != tokenizer.eos_id: # if score < 0 then repetition penalty has to multiplied to reduce the previous token probability if lprobs[i, previous_token] < 0: lprobs[i, previous_token] *= repetition_penalty else: lprobs[i, previous_token] /= repetition_penalty def postprocess_next_token_scores( tokenizer: EncDecTokenizer, scores, input_ids, no_repeat_ngram_size, bad_words_ids, cur_len, min_length, max_length, eos_token_id, repetition_penalty, batch_size, num_beams, antonym_dict, ): # repetition penalty (from CTRL paper https://arxiv.org/abs/1909.05858) if repetition_penalty != 1.0: enforce_repetition_penalty_( tokenizer, scores, batch_size, num_beams, input_ids, repetition_penalty, ) # set eos token prob to zero if min_length is not reached if eos_token_id is not None and cur_len < min_length: scores[:, eos_token_id] = -10000 if no_repeat_ngram_size > 0: # calculate a list of banned tokens to prevent repetitively generating the same ngrams num_batch_hypotheses = batch_size * num_beams # from fairseq: https://github.com/pytorch/fairseq/blob/a07cb6f40480928c9e0548b737aadd36ee66ac76/fairseq/sequence_generator.py#L345 banned_batch_tokens = calc_banned_ngram_tokens( input_ids, num_batch_hypotheses, no_repeat_ngram_size, tokenizer=tokenizer ) for i, banned_tokens in enumerate(banned_batch_tokens): scores[i, banned_tokens] = -10000 if bad_words_ids is not None: # calculate a list of banned tokens according to bad words banned_tokens = calc_banned_bad_words_ids(input_ids, bad_words_ids) for i, banned_tokens in enumerate(banned_tokens): scores[i, banned_tokens] = -10000 # add antonym banned list banned_tokens = calc_banned_antonym_words_ids(input_ids, tokenizer, antonym_dict) for i, banned_tokens in enumerate(banned_tokens): scores[i, banned_tokens] = -10000 scores[:, 0] = -50000 return scores def generate_no_beam( model_batch, full_context, model, tokenizer: EncDecTokenizer, args, device ): batch_size = args.batch_size target_length = args.max_generation_length dec_init_length = 1 # +1 for s_0 if args.FiD: model.module.module.module.reset_score_storage() batch_size, _, sequence_length = model_batch["passage_input_ids"].size() enc_input_ids = model_batch["passage_input_ids"].view( batch_size * args.passage_num, sequence_length ) enc_attention_mask = model_batch["passage_attention_mask"].view( batch_size * args.passage_num, 1, sequence_length, sequence_length ) enc_outputs = model( enc_input_ids=enc_input_ids, enc_attention_mask=enc_attention_mask, only_encoder=True, ) enc_hidden_states = enc_outputs["encoder_last_hidden_state"].view( batch_size, sequence_length * args.passage_num, -1 ) else: enc_input_ids = model_batch["enc_input_ids"] enc_attention_mask = model_batch["enc_attention_mask"] enc_outputs = model( enc_input_ids=enc_input_ids, enc_attention_mask=enc_attention_mask, only_encoder=True, ) enc_hidden_states = enc_outputs["encoder_last_hidden_state"] # for generating responses # we only use the <go> token, so truncate other tokens dec_input_ids = model_batch["dec_input_ids"][..., :dec_init_length] dec_attention_mask = model_batch["dec_attention_mask"][ ..., :dec_init_length, :dec_init_length ] # we use past_key_values, so only the current token mask is needed cross_attention_mask = model_batch["cross_attention_mask"][..., :dec_init_length, :] unfinished_sents = enc_input_ids.new(enc_hidden_states.size(0)).fill_(1) output_ids = enc_input_ids.new_zeros( [enc_hidden_states.size(0), 0] ) # not include the prompt prob_idx = torch.arange(batch_size) past_key_values = None gen_len = 0 # construct antonym dict antonym_dict = None while gen_len < target_length: if unfinished_sents.max() == 0: tokens_to_add = tokenizer.eos_id * (1 - unfinished_sents) output_ids = torch.cat([output_ids, tokens_to_add.unsqueeze(-1)], dim=-1) else: dec_outputs = model( dec_input_ids=dec_input_ids, dec_attention_mask=dec_attention_mask, cross_attention_mask=cross_attention_mask, enc_hidden_states=enc_hidden_states, past_key_values=past_key_values, ) past_key_values = dec_outputs["past_key_values"] lm_logits = dec_outputs["lm_logits"] logits = lm_logits[:, -1, :] / args.temperature prev_output_tokens = torch.cat([full_context, output_ids], dim=-1) logits = postprocess_next_token_scores( tokenizer=tokenizer, scores=logits, input_ids=prev_output_tokens, no_repeat_ngram_size=args.no_repeat_ngram_size, bad_words_ids=[[0]], cur_len=gen_len, min_length=args.min_generation_length, max_length=args.max_generation_length, eos_token_id=tokenizer.eos_id, repetition_penalty=args.repetition_penalty, batch_size=batch_size, num_beams=1, antonym_dict=antonym_dict, ) if args.sampling: logits = top_k_logits(logits, top_k=args.top_k, top_p=args.top_p) probs = F.softmax(logits.float(), dim=-1) next_token = torch.multinomial(probs, num_samples=1).squeeze(1) else: next_token = torch.argmax(logits, -1) tokens_to_add = next_token * unfinished_sents + tokenizer.pad_id * ( 1 - unfinished_sents ) dec_input_ids = tokens_to_add.unsqueeze(-1) output_ids = torch.cat([output_ids, tokens_to_add.unsqueeze(-1)], dim=-1) # let the current token attend to all previous tokens dec_attention_mask = torch.cat( [dec_attention_mask[:, :, -1:, :], dec_attention_mask[:, :, -1:, -1:]], dim=-1, ) cross_attention_mask = cross_attention_mask[:, :, -1:, :] gen_len += 1 unfinished_sents.mul_(tokens_to_add.ne(tokenizer.eos_id).long()) output_ids = output_ids.cpu().tolist() generation_token_ids_list = [] generation_str_list = [] for e in output_ids: generation_token_ids = ( e[: e.index(tokenizer.eos_id)] if tokenizer.eos_id in e else e ) generation_token_ids_list.append(generation_token_ids) generation_str_list.append(tokenizer.decode(generation_token_ids)) return generation_str_list, generation_token_ids_list def generate_beam( model_batch, full_context, model, tokenizer: EncDecTokenizer, args, device ): """ Since the context in model batch is truncated, we need full_context to store the tokens in the entire context. """ batch_size = args.batch_size num_beams = args.num_beams target_length = args.max_generation_length do_sample = args.sampling and (args.top_p > 0 or args.top_k > 0) vocab_size = tokenizer.vocab_size enc_input_ids = model_batch["enc_input_ids"] enc_attention_mask = model_batch["enc_attention_mask"] enc_input_length = enc_input_ids.size(-1) enc_input_ids = enc_input_ids.unsqueeze(1).expand( batch_size, num_beams, enc_input_length ) enc_attention_mask = enc_attention_mask.unsqueeze(1).expand( batch_size, num_beams, 1, enc_input_length, enc_input_length ) enc_input_ids = enc_input_ids.contiguous().view( batch_size * num_beams, enc_input_length ) enc_attention_mask = enc_attention_mask.contiguous().view( batch_size * num_beams, 1, enc_input_length, enc_input_length ) full_context = full_context.unsqueeze(1).expand( batch_size, num_beams, full_context.size(-1) ) full_context = full_context.contiguous().view( batch_size * num_beams, full_context.size(-1) ) enc_outputs = model( enc_input_ids=enc_input_ids, enc_attention_mask=enc_attention_mask, only_encoder=True, ) enc_hidden_states = enc_outputs["encoder_last_hidden_state"] dec_init_length = 1 # 1 for s_0 # for generating responses dec_input_ids = model_batch["dec_input_ids"][..., :dec_init_length] dec_attention_mask = model_batch["dec_attention_mask"][ ..., :dec_init_length, :dec_init_length ] # we use past_key_values, so only the current token mask is needed cross_attention_mask = model_batch["cross_attention_mask"][..., :dec_init_length, :] dec_input_ids = dec_input_ids.unsqueeze(1).expand( batch_size, num_beams, dec_init_length ) dec_attention_mask = dec_attention_mask.unsqueeze(1).expand( batch_size, num_beams, 1, dec_init_length, dec_init_length ) cross_attention_mask = cross_attention_mask.unsqueeze(1).expand( batch_size, num_beams, 1, dec_init_length, enc_input_length ) dec_input_ids = dec_input_ids.contiguous().view( batch_size * num_beams, dec_init_length ) dec_attention_mask = dec_attention_mask.contiguous().view( batch_size * num_beams, 1, dec_init_length, dec_init_length ) cross_attention_mask = cross_attention_mask.contiguous().view( batch_size * num_beams, 1, dec_init_length, enc_input_length ) done = [False for _ in range(batch_size)] output_ids = enc_input_ids.new_zeros( [enc_input_ids.size(0), 0] ) # not include the prompt past_key_values = None gen_len = 0 # construct antonym dict antonym_dict = None # generated hypotheses generated_hyps = [ BeamHypotheses( num_beams, target_length, args.length_penalty, early_stopping=args.early_stopping, tokenizer=tokenizer, ) for _ in range(batch_size) ] beam_scores = torch.zeros( (batch_size, num_beams), dtype=torch.float, device=dec_input_ids.device ) beam_scores = beam_scores.view(-1) # shape (batch_size * num_beams,) while gen_len < target_length: dec_outputs = model( dec_input_ids=dec_input_ids, dec_attention_mask=dec_attention_mask, cross_attention_mask=cross_attention_mask, enc_hidden_states=enc_hidden_states, past_key_values=past_key_values, ) past_key_values = dec_outputs["past_key_values"] lm_logits = dec_outputs["lm_logits"] logits = lm_logits[:, -1, :] / args.temperature scores = F.log_softmax(logits, dim=-1) prev_output_tokens = torch.cat([full_context, output_ids], dim=-1) scores = postprocess_next_token_scores( tokenizer=tokenizer, scores=scores, input_ids=prev_output_tokens, no_repeat_ngram_size=args.no_repeat_ngram_size, bad_words_ids=None, cur_len=gen_len, min_length=args.min_generation_length, max_length=args.max_generation_length, eos_token_id=tokenizer.eos_id, repetition_penalty=args.repetition_penalty, batch_size=batch_size, num_beams=num_beams, antonym_dict=antonym_dict, ) if do_sample: _scores = scores + beam_scores[:, None].expand_as(scores) if args.temperature != 1.0: _scores = _scores / args.temperature _scores = top_k_logits(_scores, top_k=args.top_k, top_p=args.top_p) _scores = _scores.contiguous().view(batch_size, num_beams * vocab_size) # Sample 2 next tokens for each beam (so we have some spare tokens and match output of greedy beam search) probs = F.softmax(_scores, dim=-1) next_tokens = torch.multinomial( probs, num_samples=2 * num_beams ) # (batch_size, num_beams * 2) # Compute next scores next_scores = torch.gather( _scores, -1, next_tokens ) # (batch_size, num_beams * 2) # sort the sampled vector to make sure that the first num_beams samples are the best next_scores, next_scores_indices = torch.sort( next_scores, descending=True, dim=1 ) next_tokens = torch.gather( next_tokens, -1, next_scores_indices ) # (batch_size, num_beams * 2) else: next_scores = scores + beam_scores[:, None].expand_as( scores ) # (batch_size * num_beams, vocab_size) # re-organize to group the beam together (we are keeping top hypothesis accross beams) next_scores = next_scores.view( batch_size, num_beams * vocab_size ) # (batch_size, num_beams * vocab_size) next_scores, next_tokens = torch.topk( next_scores, 2 * num_beams, dim=1, largest=True, sorted=True ) assert next_scores.size() == next_tokens.size() == (batch_size, 2 * num_beams) # next batch beam content next_batch_beam = [] for batch_idx in range(batch_size): # if we are done with this sentence, add a pad token if done[batch_idx]: assert ( len(generated_hyps[batch_idx]) >= num_beams ), "Batch can only be done if at least {} beams have been generated".format( num_beams ) next_batch_beam.extend( [(0, tokenizer.pad_id, 0)] * num_beams ) # pad the batch continue # next sentence beam content, this will get added to next_batch_beam next_sent_beam = [] # next tokens for this sentence for beam_token_rank, (beam_token_id, beam_token_score) in enumerate( zip(next_tokens[batch_idx], next_scores[batch_idx]) ): # get beam and token IDs beam_id = beam_token_id // vocab_size token_id = beam_token_id % vocab_size effective_beam_id = batch_idx * num_beams + beam_id # add to generated hypotheses if end of sentence if token_id.item() == tokenizer.eos_id: # if beam_token does not belong to top num_beams tokens, it should not be added is_beam_token_worse_than_top_num_beams = ( beam_token_rank >= num_beams ) if is_beam_token_worse_than_top_num_beams: continue generated_hyps[batch_idx].add( output_ids[effective_beam_id].clone(), beam_token_score.item(), ) else: # add next predicted token since it is not eos_token next_sent_beam.append( (beam_token_score, token_id, effective_beam_id) ) # once the beam for next step is full, don't add more tokens to it. if len(next_sent_beam) == num_beams: break # Check if we are done so that we can save a pad step if all(done) done[batch_idx] = done[batch_idx] or generated_hyps[batch_idx].is_done( next_scores[batch_idx].max().item(), gen_len ) # update next beam content assert len(next_sent_beam) == num_beams, "Beam should always be full" next_batch_beam.extend(next_sent_beam) assert len(next_batch_beam) == num_beams * ( batch_idx + 1 ), "We should have added num_beams each step" # stop when we are done with each sentence if all(done): break # sanity check / prepare next batch assert len(next_batch_beam) == batch_size * num_beams beam_scores = torch.tensor( [x[0] for x in next_batch_beam], device=dec_input_ids.device ) beam_tokens = torch.tensor( [x[1] for x in next_batch_beam], device=dec_input_ids.device ) beam_idx = torch.tensor( [x[2] for x in next_batch_beam], device=dec_input_ids.device ) # re-order batch and update current length output_ids = output_ids[beam_idx, :] output_ids = torch.cat([output_ids, beam_tokens.unsqueeze(1)], dim=-1) dec_input_ids = beam_tokens.unsqueeze(1) dec_attention_mask = torch.cat( [dec_attention_mask[:, :, -1:, :], dec_attention_mask[:, :, -1:, -1:]], dim=-1, ) cross_attention_mask = cross_attention_mask[:, :, -1:, :] # past_key_values = num_layer * 2 * (2, beam_size, 32, prefix_len, 64) first 2: self/cross attention, second 2: key/value past_key_values = [ [ torch.index_select(layer_past_type, 1, beam_idx) for layer_past_type in layer_past ] for layer_past in past_key_values ] gen_len += 1 # finalize all open beam hypotheses and add to generated hypotheses for batch_idx in range(batch_size): if done[batch_idx]: continue # need to add best num_beams hypotheses to generated hyps for beam_id in range(num_beams): effective_beam_id = batch_idx * num_beams + beam_id final_score = beam_scores[effective_beam_id].item() final_tokens = output_ids[effective_beam_id] generated_hyps[batch_idx].add(final_tokens, final_score) best = [] best_ids = [] # retrieve best hypotheses for i, hypotheses in enumerate(generated_hyps): sorted_hyps = sorted(hypotheses.beams, key=lambda x: x[0]) best_hyp = sorted_hyps.pop()[1] best.append(tokenizer.decode(best_hyp.cpu().tolist())) best_ids.append(best_hyp.cpu().tolist()) return best, best_ids

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/LM/Flan-T5/train_t0.py

# coding=utf-8 """Training Enc-Dec""" import os import torch import json import numpy as np from arguments import get_args from data_utils.T0Datasets import T0Dataset from data_utils.data_config import ( DATA_GROUP_CONFIG, DATA_NO_EVAL, DATA_NO_VALID, DATA_NO_TRAIN, DATA_EVAL_GEN, DATA_RETRIEVAL_AUGMENTATION, RA_PASSAGE_NUM, ) from data_utils import ANSWER_POST_FN from tokenization_t5 import EncDecTokenizer import mpu from utils import save_checkpoint from utils import print_args from utils import print_rank_0, save_rank_0 from utils import save_preds_t0 from utils import setup_model_and_optimizer, set_random_seed, initialize_distributed from samplers import DistributedBatchSampler, RandomSampler from data_utils import * from metrics import * from torch.utils.data import DataLoader, SequentialSampler from generation_utils import generate_beam, generate_no_beam from promptsource.templates import TemplateCollection from tqdm import tqdm def forward_step( args, model_batch, no_model_batch, model, device, keep_enc_hidden=False, do_infer=False, ): for k in model_batch: model_batch[k] = model_batch[k].to(device) for k in no_model_batch: no_model_batch[k] = no_model_batch[k].to(device) if args.FiD: batch_size, _, sequence_length = model_batch["passage_input_ids"].size() enc_outputs = model( enc_input_ids=model_batch["passage_input_ids"].view( batch_size * args.passage_num, sequence_length ), enc_attention_mask=model_batch["passage_attention_mask"].view( batch_size * args.passage_num, 1, sequence_length, sequence_length ), only_encoder=True, ) enc_hidden_states = enc_outputs["encoder_last_hidden_state"].view( batch_size, sequence_length * args.passage_num, -1 ) new_model_batch = {} for k in model_batch: if k not in ["passage_input_ids", "passage_attention_mask"]: new_model_batch[k] = model_batch[k] output = model(**new_model_batch, enc_hidden_states=enc_hidden_states) else: if keep_enc_hidden: enc_outputs = model(**model_batch, only_encoder=True) enc_hidden_states = enc_outputs["encoder_last_hidden_state"] output = model(**model_batch, enc_hidden_states=enc_hidden_states) else: output = model(**model_batch) logits = output["lm_logits"] forw_out = {"logits": logits} if keep_enc_hidden: forw_out["enc_hidden_states"] = enc_hidden_states if not do_infer: losses = mpu.vocab_parallel_cross_entropy( logits.contiguous().float(), no_model_batch["labels"] ) loss_mask = no_model_batch["loss_mask"] losses = (losses * loss_mask).sum(-1) / loss_mask.sum(-1) loss = losses.mean() forw_out["loss"] = loss forw_out["loss_batch"] = losses return forw_out def backward_step(args, loss, model, optimizer): # backward if args.deepspeed: model.backward(loss) else: optimizer.zero_grad() if args.fp16: optimizer.backward(loss, update_master_grads=False) else: loss.backward() # Update master gradients. if not args.deepspeed: if args.fp16: optimizer.update_master_grads() # Clipping gradients helps prevent the exploding gradient. if args.clip_grad > 0: if not args.fp16: mpu.clip_grad_norm(model.parameters(), args.clip_grad) else: optimizer.clip_master_grads(args.clip_grad) def train( args, data_names, tokenizer: EncDecTokenizer, model, optimizer, lr_scheduler, train_data_utils, dev_data_utils, device, ): """Train the model.""" train_dataloader, train_dataset, random_sampler = train_data_utils # Turn on training mode which enables dropout. model.train() # Tracking loss. total_loss = 0.0 step, global_step = 1, 1 best_scores = [] for e in range(args.epochs): model.train() random_sampler.set_epoch(e) train_dataset.set_epoch(e) for model_batch, no_model_batch, _, _ in train_dataloader: forw_out = forward_step(args, model_batch, no_model_batch, model, device) loss = forw_out["loss"] if torch.distributed.get_rank() == 0: print(loss) backward_step(args, loss, model, optimizer) # Update losses. total_loss += loss.item() if args.deepspeed: model.step() else: optimizer.step() if not (args.fp16 and optimizer.overflow): lr_scheduler.step() # Logging. if ( global_step % args.log_interval == 0 and step % args.gradient_accumulation_steps == 0 ): learning_rate = optimizer.param_groups[0]["lr"] avg_lm_loss = total_loss / ( args.log_interval * args.gradient_accumulation_steps ) log_string = "epoch {:3d}/{:3d} |".format(e, args.epochs) log_string += " global iteration {:8d}/{:8d} |".format( global_step, args.train_iters ) log_string += " learning rate {:.3} |".format(learning_rate) log_string += " lm loss {:.6} |".format(avg_lm_loss) if args.fp16: log_string += " loss scale {:.1f} |".format( optimizer.cur_scale if args.deepspeed else optimizer.loss_scale ) print_rank_0(log_string) save_rank_0(args, log_string) total_loss = 0.0 # Checkpointing if ( args.save and args.save_interval and global_step % args.save_interval == 0 and step % args.gradient_accumulation_steps == 0 ): save_dir_path = os.path.join(args.save, str(global_step)) if torch.distributed.get_rank() == 0: os.makedirs(save_dir_path, exist_ok=True) save_checkpoint( global_step, model, optimizer, lr_scheduler, args, save_dir=save_dir_path, ) # Evaluation if ( args.eval_interval and global_step % args.eval_interval == 0 and step % args.gradient_accumulation_steps == 0 and args.do_valid ): prefix = "iteration {} | ".format(global_step) metric_values = [] for name, dev_data_util_prompt in dev_data_utils.items(): if DATA_CONFIG[name].get("selfsup", False): if DATA_CONFIG[name]["type"] == "gen": dev_dataloader, dev_dataset, _ = dev_data_util_prompt[0] dev_loss = evaluate_lm( args, tokenizer, name, dev_dataset, dev_dataloader, model, device, mode="dev", train_step=global_step, save_res=True, ) log_string = ( prefix + name + " | dev_loss: " + str(np.mean(dev_loss)) ) print_rank_0(log_string) save_rank_0(args, log_string) else: dev_dataloader, dev_dataset, _ = dev_data_util_prompt[0] dev_loss, dev_res, dev_preds, dev_labels = evaluate_rank( args, tokenizer, name, dev_dataset, dev_dataloader, model, device, mode="dev", train_step=global_step, save_res=True, ) log_string = ( prefix + name + " | dev_loss: " + str(np.mean(dev_loss)) + " | dev res: " + str(dev_res) ) print_rank_0(log_string) save_rank_0(args, log_string) else: dev_res_prompt = [] dev_loss_prompt = [] dev_preds_prompt = [] dev_labels_prompt = [] dev_prompt_names = [] for pid, dev_data_util in enumerate(dev_data_util_prompt): dev_dataloader, dev_dataset, _ = dev_data_util dev_prompt_names.append( dev_dataset.all_data[name]["prompt_names"][0] ) if ( dev_dataset.data_prompts[name][0].answer_choices is not None ): eval_func = evaluate_rank else: eval_func = evaluate_gen dev_loss, dev_res, dev_preds, dev_labels = eval_func( args, tokenizer, name, dev_dataset, dev_dataloader, model, device, mode="dev", train_step=global_step, save_res=True, ) dev_loss_prompt.append(dev_loss) dev_res_prompt.append(dev_res) dev_preds_prompt.append(dev_preds) dev_labels_prompt.append(dev_labels) log_string = ( prefix + name + " | dev_loss: " + str(np.mean(dev_loss_prompt)) + " | dev res: " + str(dev_res_prompt) ) print_rank_0(log_string) save_rank_0(args, log_string) save_preds_t0( args, name, dev_prompt_names, global_step, dev_res_prompt, dev_preds_prompt, dev_labels_prompt, ) values = [ v for dev_res in dev_res_prompt for v in dev_res.values() ] metric_values.extend(values) if len(metric_values) != 0: metric_avg = sum(metric_values) / len(metric_values) log_string = prefix + "Average: " + str(metric_avg) print_rank_0(log_string) save_rank_0(args, log_string) model.train() step += 1 if step % args.gradient_accumulation_steps == 0: global_step += 1 return global_step def evaluate_lm( args, tokenizer: EncDecTokenizer, name, eval_dataset: T0Dataset, eval_data_loader, model, device, mode="dev", train_step=0, save_res=False, ): model.eval() total_loss = 0.0 step = 0 with torch.no_grad(): for model_batch, no_model_batch, _, _ in eval_data_loader: for k in model_batch: model_batch[k] = model_batch[k].to(device) for k in no_model_batch: no_model_batch[k] = no_model_batch[k].to(device) forw_out = forward_step( args, model_batch, no_model_batch, model, device, keep_enc_hidden=True ) loss = forw_out["loss"].item() if "loss" in forw_out else 0 total_loss += loss step += 1 if step == 0: if torch.distributed.get_rank() == 0: print(name) print(eval_dataset.data_prompts[name][0].name) print(len(eval_dataset)) total_loss /= step return total_loss def evaluate_gen( args, tokenizer: EncDecTokenizer, name, eval_dataset: T0Dataset, eval_data_loader, model, device, mode="dev", train_step=0, save_res=False, ): # Turn on evaluation mode which disables dropout. model.eval() total_loss = 0.0 step = 0 all_output_ids = [] all_idxs = [] if args.FiD: all_scores = [] with torch.no_grad(): if not args.FiD: for model_batch, no_model_batch, _, _ in eval_data_loader: for k in model_batch: model_batch[k] = model_batch[k].to(device) for k in no_model_batch: no_model_batch[k] = no_model_batch[k].to(device) forw_out = forward_step( args, model_batch, no_model_batch, model, device, keep_enc_hidden=True, ) loss = forw_out["loss"].item() if "loss" in forw_out else 0 total_loss += loss step += 1 if step == 0: if torch.distributed.get_rank() == 0: print(name) print(eval_dataset.data_prompts[name][0].name) print(len(eval_dataset)) total_loss /= step for e, (model_batch, no_model_batch, _, _) in tqdm( enumerate(eval_data_loader), desc="Evaluating" ): for k in model_batch: model_batch[k] = model_batch[k].to(device) for k in no_model_batch: no_model_batch[k] = no_model_batch[k].to(device) if args.num_beams == 1: generation_str_list, generation_id_list = generate_no_beam( model_batch, model_batch["enc_input_ids"], model, tokenizer, args, device, ) if args.FiD: scores = model.module.module.module.get_crossattention_scores( model_batch["passage_attention_mask"][:, :, 0, 0, :].bool() ) all_scores.append(scores) else: generation_str_list, generation_id_list = generate_beam( model_batch, model_batch["enc_input_ids"], model, tokenizer, args, device, ) output_ids = [ x + [tokenizer.pad_id] + (args.max_generation_length - len(x)) * [tokenizer.pad_id] for x in generation_id_list ] output_ids = torch.tensor(output_ids).to(device) tmp_idxs = [ torch.zeros_like(no_model_batch["idxs"]).to(device) for _ in range(mpu.get_data_parallel_world_size()) ] torch.distributed.all_gather( tmp_idxs, no_model_batch["idxs"].data, group=mpu.get_data_parallel_group(), ) tmp_output_ids = [ torch.zeros_like(output_ids).to(device) for _ in range(mpu.get_data_parallel_world_size()) ] torch.distributed.all_gather( tmp_output_ids, output_ids.data, group=mpu.get_data_parallel_group() ) all_idxs.extend(tmp_idxs) all_output_ids.extend(tmp_output_ids) all_output_ids = torch.cat(all_output_ids, dim=0).cpu().tolist() all_idxs = torch.cat(all_idxs, dim=0).tolist() if args.FiD: all_scores = torch.cat(all_scores, dim=0) print(all_scores.size()) torch.save( all_scores, os.path.join(args.save, f"stored_FiD_scores.pt"), ) all_preds_real = [] all_labels_real = [] eval_res = {} for idxs, output_ids in zip(all_idxs, all_output_ids): _, _, sid = idxs output_ids = ( output_ids[: output_ids.index(tokenizer.pad_id)] if tokenizer.pad_id in output_ids else output_ids ) all_preds_real.append(tokenizer.decode(output_ids)) all_labels_real.append(eval_dataset.all_data[name]["data"][sid]["answer"]) metric_names = eval_dataset.data_prompts[name][0].metadata.metrics for metric_name in metric_names: if (name, metric_name) in ANSWER_POST_FN: all_labels_real, all_preds_real = ANSWER_POST_FN[(name, metric_name)]( all_labels_real, all_preds_real ) res = T0_METRICS[metric_name](all_labels_real, all_preds_real) eval_res.update(res) # if save_res: # save_preds_t0(args, name, eval_dataset, train_step, eval_res, all_preds_real, all_labels_real) return total_loss, eval_res, all_preds_real, all_labels_real def evaluate_rank( args, tokenizer: EncDecTokenizer, name, eval_dataset: T0Dataset, eval_data_loader, model, device, mode="dev", train_step=0, save_res=False, ): """Evaluation.""" # Turn on evaluation mode which disables dropout. model.eval() total_loss = 0.0 step = 0 all_idxs = [] all_preds = [] if args.prompt_tune: all_prompt = torch.load( f"data/{args.data_names}/cache/stored_dembeds.pt", map_location=lambda storage, loc: storage, ) if args.FiD: all_scores = [] tmp_pos_index = torch.arange(1, eval_dataset.max_cand_len + 1, device=device) with torch.no_grad(): for ( model_batch, no_model_batch, cand_model_batch, cand_no_model_batch, ) in tqdm(eval_data_loader, desc="Evaluating"): for k in model_batch: model_batch[k] = model_batch[k].to(device) for k in no_model_batch: no_model_batch[k] = no_model_batch[k].to(device) for k in cand_model_batch: cand_model_batch[k] = cand_model_batch[k].to(device) for k in cand_no_model_batch: cand_no_model_batch[k] = cand_no_model_batch[k].to(device) if args.prompt_tune: prompt = all_prompt[step] model.module.module.module.encoder.load_prompt_embeds(prompt) if args.FiD: model.module.module.module.reset_score_storage() batch_size, _, sequence_length = model_batch["passage_input_ids"].size() enc_outputs = model( enc_input_ids=model_batch["passage_input_ids"].view( batch_size * args.passage_num, sequence_length ), enc_attention_mask=model_batch["passage_attention_mask"].view( batch_size * args.passage_num, 1, sequence_length, sequence_length, ), only_encoder=True, ) enc_hidden_states = enc_outputs["encoder_last_hidden_state"].view( batch_size, sequence_length * args.passage_num, -1 ) else: enc_outputs = model(**model_batch, only_encoder=True) enc_hidden_states = enc_outputs["encoder_last_hidden_state"] # enc_hidden_states[0, :10, :] = prompt output = model(**cand_model_batch, enc_hidden_states=enc_hidden_states) if args.FiD: scores = model.module.module.module.get_crossattention_scores( model_batch["passage_attention_mask"][:, :, 0, 0, :].bool() ) all_scores.append(scores) logits = output["lm_logits"] losses = mpu.vocab_parallel_cross_entropy( logits.contiguous().float(), cand_no_model_batch["target_ids"] ) loss_mask = cand_no_model_batch["loss_mask"] losses = losses * loss_mask gold_loss = 0 preds = [] for samp_loss, cand_pos, cand_label in zip( losses, cand_no_model_batch["pos"], cand_no_model_batch["labels"] ): cum_loss = torch.cumsum(samp_loss, dim=0) # print(samp_loss) sum_loss = torch.masked_select(cum_loss, cand_pos) cand_loss = torch.diff( sum_loss, dim=0, prepend=torch.zeros(1, device=device) ) # print("cand loss", cand_loss) # print("samp loss", samp_loss) cand_pos_idx = torch.masked_select(tmp_pos_index, cand_pos) cand_lens = torch.diff( cand_pos_idx, dim=0, prepend=torch.zeros(1, device=device) ) # print("cand_lens", cand_lens) if args.no_norm_cand_loss: normed_cand_loss = cand_loss else: normed_cand_loss = cand_loss / cand_lens # print(normed_cand_loss) # exit(0) max_res = torch.min(normed_cand_loss, dim=0) preds.append(max_res.indices.item()) gold_loss += normed_cand_loss[cand_label.item()].item() gold_loss /= len(losses) total_loss += gold_loss preds = torch.tensor(preds, dtype=torch.long, device=device) gathered_preds = [ torch.zeros_like(preds) for _ in range(mpu.get_data_parallel_world_size()) ] torch.distributed.all_gather( gathered_preds, preds.contiguous(), mpu.get_data_parallel_group() ) all_preds.extend(gathered_preds) gathered_idx = [ torch.zeros_like(no_model_batch["idxs"]) for _ in range(mpu.get_data_parallel_world_size()) ] torch.distributed.all_gather( gathered_idx, no_model_batch["idxs"].contiguous(), mpu.get_data_parallel_group(), ) all_idxs.extend(gathered_idx) step += 1 if step == 0: if torch.distributed.get_rank() == 0: print(name) print(eval_dataset.data_prompts[name][0].name) print(len(eval_dataset)) total_loss /= step all_idxs = torch.cat(all_idxs, dim=0).cpu().tolist() all_preds = torch.cat(all_preds, dim=0).cpu().tolist() if args.FiD: all_scores = torch.cat(all_scores, dim=0) print(all_scores.size()) torch.save( all_scores, os.path.join(args.save, f"stored_FiD_scores.pt"), ) all_preds_real = [] all_labels_real = [] eval_res = {} for idxs, pred in zip(all_idxs, all_preds): _, _, sid = idxs sample = eval_dataset.all_data[name]["data"][sid] all_preds_real.append(sample["options"][pred]) all_labels_real.append(sample["answer"]) if eval_dataset.data_prompts[name] is None: # selfsup metric_names = ["Other"] else: metric_names = eval_dataset.data_prompts[name][0].metadata.metrics for metric_name in metric_names: if (name, metric_name) in ANSWER_POST_FN: all_labels_real, all_preds_real = ANSWER_POST_FN[(name, metric_name)]( all_labels_real, all_preds_real ) res = T0_METRICS[metric_name](all_labels_real, all_preds_real) eval_res.update(res) # if save_res: # save_preds_t0(args, name, eval_dataset, train_step, eval_res, all_preds_real, all_labels_real) return total_loss, eval_res, all_preds_real, all_labels_real def load_data( args, data_prompts, split, tokenizer, ratio=1, num=-1, few_data_names=None, drop_last=True, ): # Data parallel arguments. world_size = mpu.get_data_parallel_world_size() rank = mpu.get_data_parallel_rank() if args.eval_batch_size is None: args.eval_batch_size = args.batch_size if split == "train": global_batch_size = args.batch_size * world_size elif split == "validation": global_batch_size = args.dev_batch_size * world_size else: global_batch_size = args.eval_batch_size * world_size num_workers = args.num_workers dataset = T0Dataset( args, tokenizer, data_prompts, split, ratio=ratio, few_data_names=few_data_names, num=num, ) if split == "train": sampler = RandomSampler(dataset) sampler.set_seed(args.seed) else: sampler = SequentialSampler(dataset) batch_sampler = DistributedBatchSampler( sampler=sampler, batch_size=global_batch_size, drop_last=drop_last, rank=rank, world_size=world_size, ) data_loader = DataLoader( dataset, batch_sampler=batch_sampler, num_workers=num_workers, pin_memory=True, collate_fn=dataset.collate, ) # Torch dataloader. return data_loader, dataset, sampler def main(): """Main training program.""" # Disable CuDNN. torch.backends.cudnn.enabled = False # Arguments. args = get_args() os.makedirs(args.save, exist_ok=True) # Pytorch distributed. initialize_distributed(args) if torch.distributed.get_rank() == 0: print("Training Enc-Dec model") print_args(args) with open(os.path.join(args.save, "args.json"), "w") as f: json.dump(vars(args), f) # Random seeds for reproducability. set_random_seed(args.seed) device = torch.cuda.current_device() # setup tokenizer tokenizer = EncDecTokenizer( os.path.join(args.tokenizer_path, "spiece.model"), pad_token=args.pad_token ) with open(args.deepspeed_config, "r") as f: ds_config = json.load(f) ds_config["gradient_accumulation_steps"] = args.gradient_accumulation_steps ds_config["train_micro_batch_size_per_gpu"] = args.batch_size data_group_names = args.data_names.split("-") data_names = [] for name in data_group_names: if name in DATA_GROUP_CONFIG: data_names.extend(DATA_GROUP_CONFIG[name]) else: data_names.append(name) few_data_names = None if args.few_data_names is not None: few_data_group_names = args.few_data_names.split("-") few_data_names = [] for name in few_data_group_names: if name in DATA_GROUP_CONFIG: few_data_names.extend(DATA_GROUP_CONFIG[name]) else: few_data_names.append(name) data_prompts = {} for name in data_names: for ra_name in DATA_RETRIEVAL_AUGMENTATION: if ra_name in name: DATA_CONFIG[name] = DATA_CONFIG[ra_name] DATA_CONFIG[name]["data_dir"] = f"data/{name}/cache" break if name in RA_PASSAGE_NUM: args.passage_num = RA_PASSAGE_NUM[name] if DATA_CONFIG[name].get("selfsup", False): data_prompts[name] = None else: collection = TemplateCollection() if "name" in DATA_CONFIG[name]: templates = collection.get_dataset( DATA_CONFIG[name]["name"][0], DATA_CONFIG[name]["name"][1] ) else: templates = collection.get_dataset(name, None) data_prompts[name] = [] for template_name in templates.all_template_names: if "mmlu" in name or "ai2_arc" in name: if template_name == "heres_a_problem": data_prompts[name].append(templates[template_name]) continue if ( "popQA" in name or "marco_qa" in name or "kilt" in name ) and template_name != "question_with_instruction": continue if (name, template_name) not in DATA_NO_TRAIN: if "popQA" in name: prompt = templates[template_name] prompt.metadata.metrics = ["popQA"] data_prompts[name].append(prompt) elif "marco_qa" in name: prompt = templates[template_name] prompt.metadata.metrics = ["BLEU", "ROUGE"] data_prompts[name].append(prompt) elif "kilt" in name: prompt = templates[template_name] prompt.metadata.metrics = ["Trivia QA"] data_prompts[name].append(prompt) else: data_prompts[name].append(templates[template_name]) print("All Data group:", data_group_names, "All Data:", data_names) if args.do_train: train_data_utils = load_data( args, data_prompts, "train", tokenizer, ratio=args.train_ratio, few_data_names=few_data_names, num=args.train_num, ) dev_data_utils = {} for name in data_prompts: if DATA_CONFIG[name].get("selfsup", False): dev_data_utils[name] = [ load_data( args, {name: None}, "validation", tokenizer, ratio=args.dev_ratio, few_data_names=few_data_names, num=args.dev_num, ) ] else: if (name, None) not in DATA_NO_VALID: dev_data_utils[name] = [] for template in data_prompts[name]: if (name, template.name) not in DATA_NO_VALID: dev_data_utils[name].append( load_data( args, {name: [template]}, "validation", tokenizer, ratio=args.dev_ratio, few_data_names=few_data_names, num=args.dev_num, ) ) if args.train_iters == -1: args.train_iters = ( len(train_data_utils[1]) * args.epochs // ( mpu.get_data_parallel_world_size() * args.batch_size * args.gradient_accumulation_steps ) ) if args.save_interval == -1: args.save_interval = len(train_data_utils[1]) // ( mpu.get_data_parallel_world_size() * args.batch_size * args.gradient_accumulation_steps ) if args.eval_interval == -1: args.eval_interval = len(train_data_utils[1]) // ( mpu.get_data_parallel_world_size() * args.batch_size * args.gradient_accumulation_steps ) else: args.train_iters = 10 # a magic number log_string = "Total train epochs {} | Total train iters {} | ".format( args.epochs, args.train_iters ) print_rank_0(log_string) save_rank_0(args, log_string) # Model, optimizer, and learning rate. prompt_config = None if args.prompt_tune: with open(args.prompt_config, "r") as f: prompt_config = json.load(f) model, optimizer, lr_scheduler = setup_model_and_optimizer( args, tokenizer.vocab_size, ds_config, set_optim=args.do_train, prompt_config=prompt_config, ) if args.do_train: train( args, data_names, tokenizer, model, optimizer, lr_scheduler, train_data_utils, dev_data_utils, device, ) if args.do_eval: for name in data_names: if (name, None) not in DATA_NO_EVAL: eval_loss_prompt = [] eval_res_prompt = [] eval_preds_prompt = [] eval_labels_prompt = [] eval_prompt_names = [] for template in data_prompts[name]: if (name, template.name) not in DATA_NO_EVAL: eval_data_utils = load_data( args, {name: [template]}, "validation", tokenizer, ratio=args.test_ratio, few_data_names=few_data_names, num=args.test_num, ) eval_dataloader, eval_dataset, _ = eval_data_utils eval_prompt_names.append( eval_dataset.all_data[name]["prompt_names"][0] ) if ( eval_dataset.data_prompts[name][0].answer_choices is not None and (name, template.name) not in DATA_EVAL_GEN ): eval_func = evaluate_rank else: eval_func = evaluate_gen eval_loss, eval_res, eval_preds, eval_labels = eval_func( args, tokenizer, name, eval_dataset, eval_dataloader, model, device, mode="test", save_res=True, ) eval_loss_prompt.append(eval_loss) eval_res_prompt.append(eval_res) eval_preds_prompt.append(eval_preds) eval_labels_prompt.append(eval_labels) avg_eval_res = { k: np.mean([res[k] for res in eval_res_prompt]) for k in eval_res_prompt[0] } log_string = ( "Eval result: loss: {:.6} | avg_res: {} | all_res: {}".format( np.mean(eval_loss_prompt), avg_eval_res, eval_res_prompt ) ) print_rank_0(log_string) save_rank_0(args, log_string) save_preds_t0( args, name, eval_prompt_names, 0, eval_res_prompt, eval_preds_prompt, eval_labels_prompt, ) if __name__ == "__main__": main()

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/LM/Flan-T5/samplers.py

# coding=utf-8 """Batch samplers that work with either random or sequential data samplers.""" import torch from torch.utils import data class RandomSampler(data.sampler.Sampler): """Based off of pytorch RandomSampler and DistributedSampler. Essentially a RandomSampler, but this class lets the user set an epoch like DistributedSampler Samples elements randomly. If without replacement, then sample from a shuffled dataset. If with replacement, then user can specify ``num_samples`` to draw. Arguments: data_source (Dataset): dataset to sample from num_samples (int): number of samples to draw, default=len(dataset) replacement (bool): samples are drawn with replacement if ``True``, default=False """ def __init__(self, data_source, replacement=False, num_samples=None, diff_order=False): self.data_source = data_source self.replacement = replacement self._num_samples = num_samples self.epoch = -1 self.seed = -1 self.diff_order = diff_order if self._num_samples is not None and replacement is False: raise ValueError("With replacement=False, num_samples should not " "be specified, since a random permute will be " "performed.") if not isinstance(self.num_samples, int) or self.num_samples <= 0: raise ValueError("num_samples should be a positive integer " "value, but got num_samples={}".format( self.num_samples)) if not isinstance(self.replacement, bool): raise ValueError("replacement should be a boolean value, but got " "replacement={}".format(self.replacement)) @property def num_samples(self): # dataset size might change at runtime if self._num_samples is None: return len(self.data_source) return self._num_samples def __iter__(self): n = len(self.data_source) g = torch.Generator() if self.diff_order: if self.epoch >= 0 and self.seed >= 0: g.manual_seed(self.epoch + self.seed) elif self.epoch >= 0: g.manual_seed(self.epoch) elif self.seed >= 0: g.manual_seed(self.seed) else: if self.seed >= 0 and self.seed != 1234: # hack g.manual_seed(self.seed) if self.replacement: return iter(torch.randint(high=n, size=(self.num_samples,), dtype=torch.int64, generator=g).tolist()) return iter(torch.randperm(n, generator=g).tolist()) def __len__(self): return self.num_samples def set_epoch(self, epoch): self.epoch = epoch def set_seed(self, seed): self.seed = seed class DistributedBatchSampler(data.sampler.BatchSampler): """Similar to normal implementation of distributed sampler, except implementation is at the batch sampler level, instead of just the sampler level. This allows wrapping of arbitrary data samplers (sequential, random, WeightedRandomSampler, etc.) with this batch sampler. The `interleave` argument specifies how to distribute a batch. A value of True combined with the above random sampler is equivalent to pytorch's torch.utils.data.distributed.DistributedSampler. For the following batch [0,1,2,3,4,5,6,7] and data parallelism of 2 specifying True will result in the following samples for each gpu: GPU0: [0,2,4,6] GPU1: [1,3,5,7] specifying False will result in the following samples: GPU0: [0,1,2,3] GPU1: [4,5,6,7]""" def __init__(self, sampler, batch_size, drop_last, rank=-1, world_size=2, wrap_last=False, interleave=False): super(DistributedBatchSampler, self).__init__(sampler, batch_size, drop_last) if rank == -1: assert False, 'should not be here' rank = torch.distributed.get_rank() self.rank = rank self.world_size = world_size self.sampler.wrap_around = 0 self.wrap_around = 0 self.wrap_last = wrap_last self.start_iter = 0 self.interleave = interleave def __iter__(self): batch = [] i = 0 for idx in self.data_iterator(self.sampler, wrap_around=False): batch.append(idx) if len(batch) == self.batch_size: tbatch = self._batch(batch) if i >= self.start_iter: yield tbatch self.start_iter = 0 i += 1 batch = [] batch_len = len(batch) if batch_len > 0 and not self.drop_last: if self.wrap_last: self.sampler.wrap_around -= (self.batch_size) self.wrap_around += (len(batch)) self.wrap_around %= self.batch_size yield self._batch(batch) if self.wrap_last: self.sampler.wrap_around += self.batch_size def data_iterator(self, _iter, wrap_around=False): """iterates through data and handles wrap around""" for i, idx in enumerate(_iter): if i < self.wrap_around % self.batch_size: continue if wrap_around: self.wrap_around += 1 self.wrap_around %= self.batch_size yield idx def _batch(self, batch): """extracts samples only pertaining to this worker's batch""" if self.interleave: return batch[self.rank:self.batch_size:self.world_size] start = self.rank * self.batch_size // self.world_size end = (self.rank + 1) * self.batch_size // self.world_size return batch[start:end]

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/LM/Flan-T5/fp16/fp16util.py

# coding=utf-8 import torch import torch.nn as nn from torch.autograd import Variable from torch._utils import _flatten_dense_tensors, _unflatten_dense_tensors import mpu class tofp16(nn.Module): """ Utility module that implements:: def forward(self, input): return input.half() """ def __init__(self): super(tofp16, self).__init__() def forward(self, input): return input.half() def BN_convert_float(module): """ Utility function for network_to_half(). Retained for legacy purposes. """ if isinstance(module, torch.nn.modules.batchnorm._BatchNorm) and module.affine is True: module.float() for child in module.children(): BN_convert_float(child) return module def network_to_half(network): """ Convert model to half precision in a batchnorm-safe way. Retained for legacy purposes. It is recommended to use FP16Model. """ return nn.Sequential(tofp16(), BN_convert_float(network.half())) def convert_module(module, dtype): """ Converts a module's immediate parameters and buffers to dtype. """ for param in module.parameters(recurse=False): if param is not None: if param.data.dtype.is_floating_point: param.data = param.data.to(dtype=dtype) if param._grad is not None and param._grad.data.dtype.is_floating_point: param._grad.data = param._grad.data.to(dtype=dtype) for buf in module.buffers(recurse=False): if buf is not None and buf.data.dtype.is_floating_point: buf.data = buf.data.to(dtype=dtype) def convert_network(network, dtype): """ Converts a network's parameters and buffers to dtype. """ for module in network.modules(): if isinstance(module, torch.nn.modules.batchnorm._BatchNorm) and module.affine is True: continue convert_module(module, dtype) return network class FP16Model(nn.Module): """ Convert model to half precision in a batchnorm-safe way. """ def __init__(self, network): super(FP16Model, self).__init__() self.network = convert_network(network, dtype=torch.half) def forward(self, *inputs): inputs = tuple(t.half() for t in inputs) return self.network(*inputs) def backwards_debug_hook(grad): raise RuntimeError("master_params recieved a gradient in the backward pass!") def prep_param_lists(model, flat_master=False): """ Creates a list of FP32 master parameters for a given model, as in `Training Neural Networks with Mixed Precision: Real Examples`_. Args: model (torch.nn.Module): Existing Pytorch model flat_master (bool, optional, default=False): Flatten the master parameters into a single tensor, as a performance optimization. Returns: A tuple (``model_params``, ``master_params``). ``model_params`` is a list of the model's parameters for later use with :func:`model_grads_to_master_grads` and :func:`master_params_to_model_params`. ``master_params`` is a list of FP32 master gradients. If ``flat_master=True``, ``master_params`` will be a list with one element. Example:: model_params, master_params = prep_param_lists(model) .. warning:: Currently, if ``flat_master=True``, all the model's parameters must be the same type. If the model has parameters of different types, use ``flat_master=False``, or use :class:`FP16_Optimizer`. .. _`Training Neural Networks with Mixed Precision: Real Examples`: http://on-demand.gputechconf.com/gtc/2018/video/S81012/ """ model_params = [param for param in model.parameters() if param.requires_grad] if flat_master: # Give the user some more useful error messages try: # flatten_dense_tensors returns a contiguous flat array. # http://pytorch.org/docs/master/_modules/torch/_utils.html master_params = _flatten_dense_tensors([param.data for param in model_params]).float() except: print("Error in prep_param_lists: model may contain a mixture of parameters " "of different types. Use flat_master=False, or use F16_Optimizer.") raise master_params = torch.nn.Parameter(master_params) master_params.requires_grad = True # master_params.register_hook(backwards_debug_hook) if master_params.grad is None: master_params.grad = master_params.new(*master_params.size()) return model_params, [master_params] else: master_params = [param.clone().float().detach() for param in model_params] for param in master_params: param.requires_grad = True return model_params, master_params def model_grads_to_master_grads(model_params, master_params, flat_master=False): """ Copy model gradients to master gradients. Args: model_params: List of model parameters created by :func:`prep_param_lists`. master_params: List of FP32 master parameters created by :func:`prep_param_lists`. If ``master_params`` was created with ``flat_master=True``, ``flat_master=True`` should also be supplied to :func:`model_grads_to_master_grads`. """ if flat_master: # The flattening may incur one more deep copy than is necessary. master_params[0].grad.data.copy_( _flatten_dense_tensors([p.grad.data for p in model_params])) else: for model, master in zip(model_params, master_params): if model.grad is not None: if master.grad is None: master.grad = Variable(master.data.new(*master.data.size())) master.grad.data.copy_(model.grad.data) else: master.grad = None def master_params_to_model_params(model_params, master_params, flat_master=False): """ Copy master parameters to model parameters. Args: model_params: List of model parameters created by :func:`prep_param_lists`. master_params: List of FP32 master parameters created by :func:`prep_param_lists`. If ``master_params`` was created with ``flat_master=True``, ``flat_master=True`` should also be supplied to :func:`master_params_to_model_params`. """ if flat_master: for model, master in zip(model_params, _unflatten_dense_tensors(master_params[0].data, model_params)): model.data.copy_(master) else: for model, master in zip(model_params, master_params): model.data.copy_(master.data) # Backward compatibility fixes def to_python_float(t): if hasattr(t, 'item'): return t.item() else: return t[0] TORCH_MAJOR = int(torch.__version__.split('.')[0]) TORCH_MINOR = int(torch.__version__.split('.')[1]) clip_grad_norm = mpu.clip_grad_norm #elif TORCH_MAJOR == 0 and TORCH_MINOR <= 4: # clip_grad_norm = torch.nn.utils.clip_grad_norm #else: # clip_grad_norm = torch.nn.utils.clip_grad_norm_

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/LM/Flan-T5/fp16/loss_scaler.py

import torch import mpu # item() is a recent addition, so this helps with backward compatibility. def to_python_float(t): if hasattr(t, 'item'): return t.item() else: return t[0] class LossScaler: """ Class that manages a static loss scale. This class is intended to interact with :class:`FP16_Optimizer`, and should not be directly manipulated by the user. Use of :class:`LossScaler` is enabled via the ``static_loss_scale`` argument to :class:`FP16_Optimizer`'s constructor. Args: scale (float, optional, default=1.0): The loss scale. """ def __init__(self, scale=1): self.cur_scale = scale # `params` is a list / generator of torch.Variable def has_overflow(self, params): return False # `x` is a torch.Tensor def _has_inf_or_nan(x): return False def update_scale(self, overflow): pass @property def loss_scale(self): return self.cur_scale def scale_gradient(self, module, grad_in, grad_out): return tuple(self.loss_scale * g for g in grad_in) def backward(self, loss, retain_graph=False): scaled_loss = loss*self.loss_scale scaled_loss.backward(retain_graph=retain_graph) class DynamicLossScaler: """ Class that manages dynamic loss scaling. It is recommended to use :class:`DynamicLossScaler` indirectly, by supplying ``dynamic_loss_scale=True`` to the constructor of :class:`FP16_Optimizer`. However, it's important to understand how :class:`DynamicLossScaler` operates, because the default options can be changed using the the ``dynamic_loss_args`` argument to :class:`FP16_Optimizer`'s constructor. Loss scaling is designed to combat the problem of underflowing gradients encountered at long times when training fp16 networks. Dynamic loss scaling begins by attempting a very high loss scale. Ironically, this may result in OVERflowing gradients. If overflowing gradients are encountered, :class:`DynamicLossScaler` informs :class:`FP16_Optimizer` that an overflow has occurred. :class:`FP16_Optimizer` then skips the update step for this particular iteration/minibatch, and :class:`DynamicLossScaler` adjusts the loss scale to a lower value. If a certain number of iterations occur without overflowing gradients detected, :class:`DynamicLossScaler` increases the loss scale once more. In this way :class:`DynamicLossScaler` attempts to "ride the edge" of always using the highest loss scale possible without incurring overflow. Args: init_scale (float, optional, default=2**32): Initial loss scale attempted by :class:`DynamicLossScaler.` scale_factor (float, optional, default=2.0): Factor used when adjusting the loss scale. If an overflow is encountered, the loss scale is readjusted to loss scale/``scale_factor``. If ``scale_window`` consecutive iterations take place without an overflow, the loss scale is readjusted to loss_scale*``scale_factor``. scale_window (int, optional, default=1000): Number of consecutive iterations without an overflow to wait before increasing the loss scale. """ def __init__(self, init_scale=2**32, scale_factor=2., scale_window=1000, min_scale=1, delayed_shift=1, consecutive_hysteresis=False): self.cur_scale = init_scale self.cur_iter = 0 self.last_overflow_iter = -1 self.scale_factor = scale_factor self.scale_window = scale_window self.min_scale = min_scale self.delayed_shift = delayed_shift self.cur_hysteresis = delayed_shift self.consecutive_hysteresis = consecutive_hysteresis # `params` is a list / generator of torch.Variable def has_overflow_serial(self, params): for p in params: if p.grad is not None and DynamicLossScaler._has_inf_or_nan(p.grad.data): return True return False def has_overflow(self, params): overflow = self.has_overflow_serial(params) # Since each model parallel GPU carries only part of the model, # make sure overflow flag is synced across all the model parallel GPUs overflow_gpu = torch.cuda.ByteTensor([overflow]) torch.distributed.all_reduce(overflow_gpu, op=torch.distributed.ReduceOp.MAX, group=mpu.get_model_parallel_group()) overflow = overflow_gpu[0].item() return bool(overflow) # `x` is a torch.Tensor def _has_inf_or_nan(x): try: # if x is half, the .float() incurs an additional deep copy, but it's necessary if # Pytorch's .sum() creates a one-element tensor of the same type as x # (which is true for some recent version of pytorch). cpu_sum = float(x.float().sum()) # More efficient version that can be used if .sum() returns a Python scalar # cpu_sum = float(x.sum()) except RuntimeError as instance: # We want to check if inst is actually an overflow exception. # RuntimeError could come from a different error. # If so, we still want the exception to propagate. if "value cannot be converted" not in instance.args[0]: raise return True else: if cpu_sum == float('inf') or cpu_sum == -float('inf') or cpu_sum != cpu_sum: return True return False # `overflow` is boolean indicating whether the gradient overflowed def update_scale(self, overflow): if not hasattr(self, 'min_scale'): self.min_scale = 1 if not hasattr(self, 'delayed_shift'): self.delayed_shift = 1 if not hasattr(self, 'cur_hysteresis'): self.cur_hysteresis = 1 if not hasattr(self, 'consecutive_hysteresis'): self.consecutive_hysteresis = True if overflow: # self.cur_scale /= self.scale_factor if self.delayed_shift == 1 or self.cur_hysteresis == 1: self.cur_scale = max(self.cur_scale/self.scale_factor, self.min_scale) else: self.cur_hysteresis -= 1 self.last_overflow_iter = self.cur_iter else: if self.consecutive_hysteresis: self.cur_hysteresis = self.delayed_shift if (self.cur_iter - self.last_overflow_iter) % self.scale_window == 0: if not self.consecutive_hysteresis: self.cur_hysteresis = self.delayed_shift self.cur_scale *= self.scale_factor self.cur_iter += 1 @property def loss_scale(self): return self.cur_scale def scale_gradient(self, module, grad_in, grad_out): return tuple(self.loss_scale * g for g in grad_in) def backward(self, loss, retain_graph=False): scaled_loss = loss*self.loss_scale scaled_loss.backward(retain_graph=retain_graph) ############################################################## # Example usage below here -- assuming it's in a separate file ############################################################## """ TO-DO separate out into an example. if __name__ == "__main__": import torch from torch.autograd import Variable from dynamic_loss_scaler import DynamicLossScaler # N is batch size; D_in is input dimension; # H is hidden dimension; D_out is output dimension. N, D_in, H, D_out = 64, 1000, 100, 10 # Create random Tensors to hold inputs and outputs, and wrap them in Variables. x = Variable(torch.randn(N, D_in), requires_grad=False) y = Variable(torch.randn(N, D_out), requires_grad=False) w1 = Variable(torch.randn(D_in, H), requires_grad=True) w2 = Variable(torch.randn(H, D_out), requires_grad=True) parameters = [w1, w2] learning_rate = 1e-6 optimizer = torch.optim.SGD(parameters, lr=learning_rate) loss_scaler = DynamicLossScaler() for t in range(500): y_pred = x.mm(w1).clamp(min=0).mm(w2) loss = (y_pred - y).pow(2).sum() * loss_scaler.loss_scale print('Iter {} loss scale: {}'.format(t, loss_scaler.loss_scale)) print('Iter {} scaled loss: {}'.format(t, loss.data[0])) print('Iter {} unscaled loss: {}'.format(t, loss.data[0] / loss_scaler.loss_scale)) # Run backprop optimizer.zero_grad() loss.backward() # Check for overflow has_overflow = DynamicLossScaler.has_overflow(parameters) # If no overflow, unscale grad and update as usual if not has_overflow: for param in parameters: param.grad.data.mul_(1. / loss_scaler.loss_scale) optimizer.step() # Otherwise, don't do anything -- ie, skip iteration else: print('OVERFLOW!') # Update loss scale for next iteration loss_scaler.update_scale(has_overflow) """

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/LM/Flan-T5/fp16/fp16.py

# coding=utf-8 """Stable version of apex FP16 Optimizer""" import torch from torch import nn from torch.autograd import Variable from torch.nn.parameter import Parameter from torch._utils import _flatten_dense_tensors, _unflatten_dense_tensors from .loss_scaler import DynamicLossScaler, LossScaler from .fp16util import model_grads_to_master_grads, master_params_to_model_params, clip_grad_norm FLOAT_TYPES = (torch.FloatTensor, torch.cuda.FloatTensor) HALF_TYPES = (torch.HalfTensor, torch.cuda.HalfTensor) def conversion_helper(val, conversion): """Apply conversion to val. Recursively apply conversion if `val` is a nested tuple/list structure.""" if not isinstance(val, (tuple, list)): return conversion(val) rtn = [conversion_helper(v, conversion) for v in val] if isinstance(val, tuple): rtn = tuple(rtn) return rtn def fp32_to_fp16(val): """Convert fp32 `val` to fp16""" def half_conversion(val): val_typecheck = val if isinstance(val_typecheck, (Parameter, Variable)): val_typecheck = val.data if isinstance(val_typecheck, FLOAT_TYPES): val = val.half() return val return conversion_helper(val, half_conversion) def fp16_to_fp32(val): """Convert fp16 `val` to fp32""" def float_conversion(val): val_typecheck = val if isinstance(val_typecheck, (Parameter, Variable)): val_typecheck = val.data if isinstance(val_typecheck, HALF_TYPES): val = val.float() return val return conversion_helper(val, float_conversion) class FP16_Module(nn.Module): def __init__(self, module): super(FP16_Module, self).__init__() self.add_module('module', module.half()) def forward(self, *inputs, **kwargs): return fp16_to_fp32(self.module(*(fp32_to_fp16(inputs)), **kwargs)) def state_dict(self, destination=None, prefix='', keep_vars=False): return self.module.state_dict(destination, prefix, keep_vars) def load_state_dict(self, state_dict, strict=True): self.module.load_state_dict(state_dict, strict=strict) # TODO: Update overflow check + downscale to use Carl's fused kernel. class FP16_Optimizer(object): """ :class:`FP16_Optimizer` is designed to wrap an existing PyTorch optimizer, and manage static or dynamic loss scaling and master weights in a manner transparent to the user. For standard use, only two lines must be changed: creating the :class:`FP16_Optimizer` instance, and changing the call to ``backward``. Example:: model = torch.nn.Linear(D_in, D_out).cuda().half() optimizer = torch.optim.SGD(model.parameters(), lr=1e-3) # Name the FP16_Optimizer instance to replace the existing optimizer # (recommended but not required): optimizer = FP16_Optimizer(optimizer, static_loss_scale = 128.0) ... # loss.backward() becomes: optimizer.backward(loss) ... Example with dynamic loss scaling:: ... optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True) # optional arg to control dynamic loss scaling behavior # dynamic_loss_args={'scale_window' : 500}) # Usually, dynamic_loss_args is not necessary. Args: init_optimizer (torch.optim.optimizer): Existing optimizer created with the parameters to optimize. Internally, :class:`FP16_Optimizer` replaces the passed optimizer's fp16 parameters, if any, with fp32 master parameters copied from the original ones. :class:`FP16_Optimizer` also stores references to the original fp16 parameters, and updates these fp16 parameters from the master fp32 copy at the end of each :attr:`step`. static_loss_scale (float, optional, default=1.0): Loss scale used internally to scale gradients computed by the model. Any fp16 gradients will be copied to fp32, then downscaled before being applied to the fp32 master params, so ``static_loss_scale`` should not affect learning rate. dynamic_loss_scale (bool, optional, default=False): Use dynamic loss scaling. If True, this will override any ``static_loss_scale`` option. dynamic_loss_args (dict, optional, default=None): Dict of kwargs that will be forwarded to the internal :class:`DynamicLossScaler` instance's constructor. Keys of this dict must match kwargs accepted by :class:`DynamicLossScaler`'s constructor. If ``dynamic_loss_args`` is unspecified, :class:`DynamicLossScaler`'s defaults will be used. verbose (bool, optional, default=True): By default, FP16_Optimizer's constructor prints out the parameters and parameter groups it is ingesting, as a sanity check. If this becomes annoying (e.g. for large models), it can be disabled by passing ``verbose=False``. ``verbose=False`` will not disable printing when the loss scale is readjusted during dynamic loss scaling. ``init_optimizer`` is expected to have been constructed in the ordinary way. It is recommended (although not required) that the newly constructed :class:`FP16_Optimizer` instance be named to replace ``init_optimizer``, for two reasons: First, it means that references to the same name later in the file will not have to change. Second, :class:`FP16_Optimizer` reserves the right (as an implementation detail) to modify ``init_optimizer``. If you do choose a unique name for the new :class:`FP16_Optimizer` instance, you should only work with this new instance, because the preexisting optimizer might no longer behave as expected. ``init_optimizer`` may be any Pytorch optimizer. It may contain a mixture of fp16 and fp32 parameters organized into any number of ``param_groups`` with different hyperparameters. The :class:`FP16_Optimizer` constructor will ingest these ``param_groups`` and remember them. Calls to :: loss.backward() must be replaced with :: optimizer.backward(loss) because :class:`FP16_Optimizer` requires ownership of the backward pass to implement loss scaling and copies to master gradients. .. note:: Loss scaling, either static or dynamic, is orthogonal to learning rate, because gradients are downscaled before being applied. This means that adjusting the loss scale, or using dynamic loss scaling, should not require retuning the learning rate or any other hyperparameters. **Advanced options** **Closures**: :class:`FP16_Optimizer` can wrap a Pytorch optimizer that receives a closure. See docstring for :attr:`step`. **Gradient clipping**: Use :attr:`clip_master_grads`. **Multiple losses**: If your model accumulates gradients from multiple losses, this can be made more efficient by supplying ``update_master_grads=False`` to :attr:`backward`. See docstring for :attr:`backward`. **Manually adjusting loss scale**: The current loss scale can be retrieved or set via :: print(optimizer.loss_scale) optimizer.loss_scale = new_loss_scale For static loss scaling, manually adjusting the loss scale over time is a reasonable thing to do. During later epochs, gradients may become smaller, and a higher loss scale may be required, analogous to scheduling the learning rate. Dynamic loss scaling is more subtle (see :class:`DynamicLossScaler`) and in this case, manually adjusting the loss scale is not recommended. **Multi_GPU training**: If the wrapped ``init_optimizer`` was created from a model wrapped in Pytorch DistributedDataParallel or Apex DistributedDataParallel, :class:`FP16_Optimizer` should still work as intended. """ def __init__(self, init_optimizer, static_loss_scale=1.0, dynamic_loss_scale=False, dynamic_loss_args=None, verbose=False): if not torch.cuda.is_available: raise SystemError("Cannot use fp16 without CUDA.") self.verbose = verbose self.optimizer = init_optimizer # init_state_dict sets up an alternative way to cast per-param state tensors. # Stashing here in case https://github.com/pytorch/pytorch/issues/7733 makes it necessary. # init_state_dict = init_optimizer.state_dict() self.fp16_groups = [] self.fp32_from_fp16_groups = [] self.fp32_from_fp32_groups = [] for i, param_group in enumerate(self.optimizer.param_groups): self.maybe_print("FP16_Optimizer processing param group {}:".format(i)) fp16_params_this_group = [] fp32_params_this_group = [] fp32_from_fp16_params_this_group = [] for i, param in enumerate(param_group['params']): if param.requires_grad: if param.type() == 'torch.cuda.HalfTensor': self.maybe_print("FP16_Optimizer received torch.cuda.HalfTensor with {}" .format(param.size())) fp16_params_this_group.append(param) master_param = param.detach().clone().float() master_param.requires_grad = True # Copythe model parallel flag. master_param.model_parallel = param.model_parallel param_group['params'][i] = master_param fp32_from_fp16_params_this_group.append(master_param) # Reset existing state dict key to the new master param. # We still need to recast per-param state tensors, if any, to FP32. if param in self.optimizer.state: self.optimizer.state[master_param] = self.optimizer.state.pop(param) elif param.type() == 'torch.cuda.FloatTensor': self.maybe_print("FP16_Optimizer received torch.cuda.FloatTensor with {}" .format(param.size())) fp32_params_this_group.append(param) param_group['params'][i] = param else: raise TypeError("Wrapped parameters must be either " "torch.cuda.FloatTensor or torch.cuda.HalfTensor. " "Received {}".format(param.type())) self.fp16_groups.append(fp16_params_this_group) self.fp32_from_fp16_groups.append(fp32_from_fp16_params_this_group) self.fp32_from_fp32_groups.append(fp32_params_this_group) # Leverage state_dict() and load_state_dict() to recast preexisting per-param state tensors self.optimizer.load_state_dict(self.optimizer.state_dict()) # alternative way to cast per-param state tensors: # self.optimizer.load_state_dict(init_state_dict) if dynamic_loss_scale: self.dynamic_loss_scale = True if dynamic_loss_args is not None: self.loss_scaler = DynamicLossScaler(**dynamic_loss_args) else: self.loss_scaler = DynamicLossScaler() else: self.dynamic_loss_scale = False self.loss_scaler = LossScaler(static_loss_scale) self.overflow = False self.first_closure_call_this_step = True self.clip_grad_norm = clip_grad_norm def maybe_print(self, msg): if self.verbose: print(msg) def __getstate__(self): raise RuntimeError("FP16_Optimizer should be serialized using state_dict().") def __setstate__(self, state): raise RuntimeError("FP16_Optimizer should be deserialized using load_state_dict().") def zero_grad(self, set_grads_to_None=False): """ Zero fp32 and fp16 parameter grads. """ # In principle, only the .grad attributes of the model params need to be zeroed, # because gradients are copied into the FP32 master params. However, we zero # all gradients owned by the optimizer, just to be safe: for group in self.optimizer.param_groups: for p in group['params']: if set_grads_to_None: p.grad = None else: if p.grad is not None: p.grad.detach_() p.grad.zero_() # Zero fp16 gradients owned by the model: for fp16_group in self.fp16_groups: for param in fp16_group: if set_grads_to_None: param.grad = None else: if param.grad is not None: param.grad.detach_() # as in torch.optim.optimizer.zero_grad() param.grad.zero_() def _check_overflow(self): params = [] for group in self.fp16_groups: for param in group: params.append(param) for group in self.fp32_from_fp32_groups: for param in group: params.append(param) self.overflow = self.loss_scaler.has_overflow(params) def _update_scale(self, has_overflow=False): self.loss_scaler.update_scale(has_overflow) def _master_params_to_model_params(self): for fp16_group, fp32_from_fp16_group in zip(self.fp16_groups, self.fp32_from_fp16_groups): master_params_to_model_params(fp16_group, fp32_from_fp16_group) def _model_params_to_master_params(self): for fp16_group, fp32_from_fp16_group in zip(self.fp16_groups, self.fp32_from_fp16_groups): master_params_to_model_params(fp32_from_fp16_group, fp16_group) # To consider: Integrate distributed with this wrapper by registering a hook on each variable # that does the overflow check, gradient copy + downscale, and fp32 allreduce in a different stream. def _model_grads_to_master_grads(self): for fp16_group, fp32_from_fp16_group in zip(self.fp16_groups, self.fp32_from_fp16_groups): model_grads_to_master_grads(fp16_group, fp32_from_fp16_group) def _downscale_master(self): if self.loss_scale != 1.0: for group in self.optimizer.param_groups: for param in group['params']: if param.grad is not None: param.grad.data.mul_(1./self.loss_scale) def clip_master_grads(self, max_norm, norm_type=2): """ Clips fp32 master gradients via ``torch.nn.utils.clip_grad_norm``. Args: max_norm (float or int): max norm of the gradients norm_type (float or int): type of the used p-norm. Can be ``'inf'`` for infinity norm. Returns: Total norm of the current fp32 gradients (viewed as a single vector). .. warning:: Returns -1 if the most recently computed fp16 gradients overflowed (that is, if ``self.overflow`` is ``True``). """ if not self.overflow: fp32_params = [] for param_group in self.optimizer.param_groups: for param in param_group['params']: fp32_params.append(param) return self.clip_grad_norm(fp32_params, max_norm, norm_type) else: return -1 def state_dict(self): """ Returns a dict containing the current state of this :class:`FP16_Optimizer` instance. This dict contains attributes of :class:`FP16_Optimizer`, as well as the state_dict of the contained Pytorch optimizer. Example:: checkpoint = {} checkpoint['model'] = model.state_dict() checkpoint['optimizer'] = optimizer.state_dict() torch.save(checkpoint, "saved.pth") """ state_dict = {} state_dict['loss_scaler'] = self.loss_scaler state_dict['dynamic_loss_scale'] = self.dynamic_loss_scale state_dict['overflow'] = self.overflow state_dict['first_closure_call_this_step'] = self.first_closure_call_this_step state_dict['optimizer_state_dict'] = self.optimizer.state_dict() state_dict['fp32_from_fp16'] = self.fp32_from_fp16_groups return state_dict def load_state_dict(self, state_dict): """ Loads a state_dict created by an earlier call to state_dict(). If ``fp16_optimizer_instance`` was constructed from some ``init_optimizer``, whose parameters in turn came from ``model``, it is expected that the user will call ``model.load_state_dict()`` before ``fp16_optimizer_instance.load_state_dict()`` is called. Example:: model = torch.nn.Linear(D_in, D_out).cuda().half() optimizer = torch.optim.SGD(model.parameters(), lr=1e-3) optimizer = FP16_Optimizer(optimizer, static_loss_scale = 128.0) ... checkpoint = torch.load("saved.pth") model.load_state_dict(checkpoint['model']) optimizer.load_state_dict(checkpoint['optimizer']) """ # I think it should actually be ok to reload the optimizer before the model. self.loss_scaler = state_dict['loss_scaler'] self.dynamic_loss_scale = state_dict['dynamic_loss_scale'] self.overflow = state_dict['overflow'] self.first_closure_call_this_step = state_dict['first_closure_call_this_step'] self.optimizer.load_state_dict(state_dict['optimizer_state_dict']) # At this point, the optimizer's references to the model's fp32 parameters are up to date. # The optimizer's hyperparameters and internal buffers are also up to date. # However, the fp32 master copies of the model's fp16 params stored by the optimizer are still # out of date. There are two options. # 1: Refresh the master params from the model's fp16 params. # This requires less storage but incurs precision loss. # 2: Save and restore the fp32 master copies separately. # We choose option 2. # # Pytorch Optimizer.load_state_dict casts saved buffers (e.g. momentum) to the type and device # of their associated parameters, because it's possible those buffers might not exist yet in # the current optimizer instance. In our case, as long as the current FP16_Optimizer has been # constructed in the same way as the one whose state_dict we are loading, the same master params # are guaranteed to exist, so we can just copy_() from the saved master params. for current_group, saved_group in zip(self.fp32_from_fp16_groups, state_dict['fp32_from_fp16']): for current, saved in zip(current_group, saved_group): current.data.copy_(saved.data) def step(self, closure=None): # could add clip option. """ If no closure is supplied, :attr:`step` should be called after ``fp16_optimizer_obj.backward(loss)``. :attr:`step` updates the fp32 master copy of parameters using the optimizer supplied to :class:`FP16_Optimizer`'s constructor, then copies the updated fp32 params into the fp16 params originally referenced by :class:`FP16_Optimizer`'s constructor, so the user may immediately run another forward pass using their model. If a closure is supplied, :attr:`step` may be called without a prior call to :attr:`backward(loss)`. This control flow is identical to `ordinary Pytorch optimizer use`_ with closures. However, the user should take care that any ``loss.backward()`` call within the closure has been replaced by ``fp16_optimizer_obj.backward(loss)``. Args: closure (optional): Closure that will be supplied to the underlying optimizer originally passed to :class:`FP16_Optimizer`'s constructor. closure should call :attr:`zero_grad()` on the :class:`FP16_Optimizer` object, compute the loss, call :attr:`backward(loss)`, and return the loss. Example with closure:: # optimizer is assumed to be an FP16_Optimizer object, previously constructed from an # existing pytorch optimizer. for input, target in dataset: def closure(): optimizer.zero_grad() output = model(input) loss = loss_fn(output, target) # loss.backward() becomes: optimizer.backward(loss) return loss optimizer.step(closure) .. warning:: Currently, calling :attr:`step` with a closure is not compatible with dynamic loss scaling. .. _`ordinary Pytorch optimizer use`: http://pytorch.org/docs/master/optim.html#optimizer-step-closure """ scale = self.loss_scaler.loss_scale self._update_scale(self.overflow) if self.overflow: self.maybe_print("OVERFLOW! Skipping step. Attempted loss scale: {}, reducing to {}" .format(scale, self.loss_scale)) return if closure is not None: retval = self._step_with_closure(closure) else: retval = self.optimizer.step() self._master_params_to_model_params() return retval def _step_with_closure(self, closure): def wrapped_closure(): # helpful for debugging # print("Calling wrapped_closure, first_closure_call_this_step = {}" # .format(self.first_closure_call_this_step)) if self.first_closure_call_this_step: # We expect that the fp16 params are initially fresh on entering self.step(), # so _master_params_to_model_params() is unnecessary the first time wrapped_closure() # is called within self.optimizer.step(). self.first_closure_call_this_step = False else: # If self.optimizer.step() internally calls wrapped_closure more than once, # it may update the fp32 params after each call. However, self.optimizer # doesn't know about the fp16 params at all. If the fp32 params get updated, # we can't rely on self.optimizer to refresh the fp16 params. We need # to handle that manually: self._master_params_to_model_params() # Our API expects the user to give us ownership of the backward() call by # replacing all calls to loss.backward() with optimizer.backward(loss). # This requirement holds whether or not the call to backward() is made within a closure. # If the user is properly calling optimizer.backward(loss) within "closure," # calling closure() here will give the fp32 master params fresh gradients # for the optimizer to play with, so all wrapped_closure needs to do is call # closure() and return the loss. temp_loss = closure() while(self.overflow): scale = self.loss_scaler.loss_scale self._update_scale(self.overflow) self.maybe_print("OVERFLOW within closure! Skipping step. Attempted loss scale: {}, " "reducing to {}".format(scale, self.loss_scale)) temp_loss = closure() return temp_loss retval = self.optimizer.step(wrapped_closure) self.first_closure_call_this_step = True return retval def backward(self, loss, update_master_grads=True, retain_graph=False): """ :attr:`backward` performs the following conceptual steps: 1. fp32_loss = loss.float() (see first Note below) 2. scaled_loss = fp32_loss*loss_scale 3. scaled_loss.backward(), which accumulates scaled gradients into the ``.grad`` attributes of the model's leaves (which may be fp16, fp32, or a mixture, depending how your model was defined). 4. fp16 grads are then copied to the master params' ``.grad`` attributes (see second Note), which are guaranteed to be fp32. 5. Finally, master grads are divided by loss_scale. In this way, after :attr:`backward`, the master params have fresh gradients, and :attr:`step` may be called. .. note:: :attr:`backward` internally converts the loss to fp32 before applying the loss scale. This provides some additional safety against overflow if the user has supplied an fp16 loss value. However, for maximum overflow safety, the user should compute the loss criterion (MSE, cross entropy, etc) in fp32 before supplying it to :attr:`backward`. .. warning:: The gradients found in a model's leaves after the call to :attr:`backward` should not be regarded as valid in general, because it's possible they have been scaled (and in the case of dynamic loss scaling, the scale factor may change over time). If the user wants to inspect gradients after a call to :attr:`backward`, only the master gradients should be regarded as valid. These can be retrieved via :attr:`inspect_master_grad_data()`. Args: loss: The loss output by the user's model. loss may be either float or half (but see first Note above). update_master_grads (bool, optional, default=True): Option to copy fp16 grads to fp32 grads on this call. By setting this to False, the user can delay the copy, which is useful to eliminate redundant fp16->fp32 grad copies if :attr:`backward` is being called on multiple losses in one iteration. If set to False, the user becomes responsible for calling :attr:`update_master_grads` before calling :attr:`step`. retain_graph (bool, optional, default=False): Forwards the usual ``retain_graph=True`` option to the internal call to ``loss.backward``. If ``retain_graph`` is being used to accumulate gradient values from multiple backward passes before calling ``optimizer.step``, passing ``update_master_grads=False`` is also recommended (see Example below). Example:: # Ordinary operation: optimizer.backward(loss) # Naive operation with multiple losses (technically valid, but less efficient): # fp32 grads will be correct after the second call, but # the first call incurs an unnecessary fp16->fp32 grad copy. optimizer.backward(loss1) optimizer.backward(loss2) # More efficient way to handle multiple losses: # The fp16->fp32 grad copy is delayed until fp16 grads from all # losses have been accumulated. optimizer.backward(loss1, update_master_grads=False) optimizer.backward(loss2, update_master_grads=False) optimizer.update_master_grads() """ # To consider: try multiple backward passes using retain_grad=True to find # a loss scale that works. After you find a loss scale that works, do a final dummy # backward pass with retain_graph=False to tear down the graph. Doing this would avoid # discarding the iteration, but probably wouldn't improve overall efficiency. self.loss_scaler.backward(loss.float(), retain_graph=retain_graph) if update_master_grads: self.update_master_grads() def update_master_grads(self): """ Copy the ``.grad`` attribute from stored references to fp16 parameters to the ``.grad`` attribute of the fp32 master parameters that are directly updated by the optimizer. :attr:`update_master_grads` only needs to be called if ``fp16_optimizer_obj.backward`` was called with ``update_master_grads=False``. """ if self.dynamic_loss_scale: self._check_overflow() if self.overflow: return self._model_grads_to_master_grads() self._downscale_master() def inspect_master_grad_data(self): """ When running with :class:`FP16_Optimizer`, ``.grad`` attributes of a model's fp16 leaves should not be regarded as truthful, because they might be scaled. After a call to :attr:`fp16_optimizer_obj.backward(loss)`, if no overflow was encountered, the fp32 master params' ``.grad`` attributes will contain valid gradients properly divided by the loss scale. However, because :class:`FP16_Optimizer` flattens some parameters, accessing them may be nonintuitive. :attr:`inspect_master_grad_data` allows those gradients to be viewed with shapes corresponding to their associated model leaves. Returns: List of lists (one list for each parameter group). The list for each parameter group is a list of the ``.grad.data`` attributes of the fp32 master params belonging to that group. """ if self.overflow: print("Warning: calling FP16_Optimizer.inspect_master_grad_data while in an overflow state. " "Gradients are currently invalid (may be inf, nan, or stale). Returning None.") return None else: # The optimizer owns only references to master params. master_grads_data = [] for param_group in self.optimizer.param_groups: master_grads_this_group = [] for param in param_group['params']: if param.grad is not None: master_grads_this_group.append(param.grad.data) else: master_grads_this_group.append(None) master_grads_data.append(master_grads_this_group) return master_grads_data # Promote loss scale so it can be retrieved or set via "fp16_optimizer_instance.loss_scale" def _get_loss_scale(self): return self.loss_scaler.loss_scale def _set_loss_scale(self, value): self.loss_scaler.cur_scale = value loss_scale = property(_get_loss_scale, _set_loss_scale) # Promote state so it can be retrieved or set via "fp16_optimizer_instance.state" def _get_state(self): return self.optimizer.state def _set_state(self, value): self.optimizer.state = value state = property(_get_state, _set_state) # Promote param_groups so it can be retrieved or set via "fp16_optimizer_instance.param_groups" # (for example, to adjust the learning rate) def _get_param_groups(self): return self.optimizer.param_groups def _set_param_groups(self, value): self.optimizer.param_groups = value param_groups = property(_get_param_groups, _set_param_groups)

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/LM/Flan-T5/mpu/mappings.py

# coding=utf-8 import torch from .initialize import get_model_parallel_group from .utils import split_tensor_along_last_dim def _reduce(input_): """All-reduce the the input tensor across model parallel group.""" group = get_model_parallel_group() # Bypass the function if we are using only 1 GPU. if torch.distributed.get_world_size(group=group) == 1: return input_ # All-reduce. torch.distributed.all_reduce(input_, group=group) return input_ def _split(input_): """Split the tensor along its last dimension and keep the corresponding slice.""" group = get_model_parallel_group() # Bypass the function if we are using only 1 GPU. if torch.distributed.get_world_size(group=group) == 1: return input_ # Split along last dimension. world_size = torch.distributed.get_world_size(group=group) input_list = split_tensor_along_last_dim(input_, world_size) # Note: torch.split does not create contiguous tensors by default. rank = torch.distributed.get_rank(group=group) output = input_list[rank].contiguous() return output def _gather(input_): """Gather tensors and concatinate along the last dimension.""" group = get_model_parallel_group() # Bypass the function if we are using only 1 GPU. if torch.distributed.get_world_size(group=group) == 1: return input_ # Size and dimension. last_dim = input_.dim() - 1 rank = torch.distributed.get_rank(group=group) world_size = torch.distributed.get_world_size(group=group) tensor_list = [torch.empty_like(input_) for _ in range(world_size)] tensor_list[rank] = input_ torch.distributed.all_gather(tensor_list, input_, group=group) # Note: torch.cat already creates a contiguous tensor. output = torch.cat(tensor_list, dim=last_dim).contiguous() return output class _CopyToModelParallelRegion(torch.autograd.Function): """Pass the input to the model parallel region.""" @staticmethod def forward(ctx, input_): return input_ @staticmethod def backward(ctx, grad_output): return _reduce(grad_output) class _ReduceFromModelParallelRegion(torch.autograd.Function): """All-redcue the input from the model parallel region.""" @staticmethod def forward(ctx, input_): return _reduce(input_) @staticmethod def backward(ctx, grad_output): return grad_output class _ScatterToModelParallelRegion(torch.autograd.Function): """Split the input and keep only the corresponding chuck to the rank.""" @staticmethod def forward(ctx, input_): return _split(input_) @staticmethod def backward(ctx, grad_output): return _gather(grad_output) class _GatherFromModelParallelRegion(torch.autograd.Function): """Gather the input from model parallel region and concatinate.""" @staticmethod def forward(ctx, input_): return _gather(input_) @staticmethod def backward(ctx, grad_output): return _split(grad_output) # ----------------- # Helper functions. # ----------------- def copy_to_model_parallel_region(input_): return _CopyToModelParallelRegion.apply(input_) def reduce_from_model_parallel_region(input_): return _ReduceFromModelParallelRegion.apply(input_) def scatter_to_model_parallel_region(input_): return _ScatterToModelParallelRegion.apply(input_) def gather_from_model_parallel_region(input_): return _GatherFromModelParallelRegion.apply(input_)

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/LM/Flan-T5/mpu/initialize.py

# coding=utf-8 """Model and data parallel groups.""" import torch from .utils import ensure_divisibility # Model parallel group that the current rank belongs to. _MODEL_PARALLEL_GROUP = None # Data parallel group that the current rank belongs to. _DATA_PARALLEL_GROUP = None def initialize_model_parallel(model_parallel_size_): """ Initialize model data parallel groups. Arguments: model_parallel_size: number of GPUs used to parallelize model. Let's say we have a total of 8 GPUs denoted by g0 ... g7 and we use 2 GPUs to parallelize the model. The present function will create 4 model parallel groups and 2 data parallel grous as: 4 model parallel groups: [g0, g1], [g2, g3], [g4, g5], [g6, g7] 2 data parallel groups: [g0, g2, g4, g6], [g1, g3, g5, g7] Note that for efficiency, the caller should make sure adjacent ranks are on the same DGX box. For example if we are using 2 DGX-1 boxes with a total of 16 GPUs, rank 0 to 7 belong to the first box and ranks 8 to 15 belong to the second box. """ if torch.distributed.get_rank() == 0: print('> initializing model parallel with size {}'.format( model_parallel_size_)) # Get world size and rank. Ensure some consistencies. assert torch.distributed.is_initialized() world_size = torch.distributed.get_world_size() model_parallel_size = min(model_parallel_size_, world_size) ensure_divisibility(world_size, model_parallel_size) rank = torch.distributed.get_rank() # Build the data parallel groups. global _DATA_PARALLEL_GROUP assert _DATA_PARALLEL_GROUP is None, \ 'data parallel group is already initialized' for i in range(model_parallel_size): ranks = range(i, world_size, model_parallel_size) group = torch.distributed.new_group(ranks) if i == (rank % model_parallel_size): _DATA_PARALLEL_GROUP = group # Build the model parallel groups. global _MODEL_PARALLEL_GROUP assert _MODEL_PARALLEL_GROUP is None, \ 'model parallel group is already initialized' for i in range(world_size // model_parallel_size): ranks = range(i * model_parallel_size, (i + 1) * model_parallel_size) group = torch.distributed.new_group(ranks) if i == (rank // model_parallel_size): _MODEL_PARALLEL_GROUP = group def model_parallel_is_initialized(): """Check if model and data parallel groups are initialized.""" if _MODEL_PARALLEL_GROUP is None or _DATA_PARALLEL_GROUP is None: return False return True def get_model_parallel_group(): """Get the model parallel group the caller rank belongs to.""" assert _MODEL_PARALLEL_GROUP is not None, \ 'model parallel group is not initialized' return _MODEL_PARALLEL_GROUP def get_data_parallel_group(): """Get the data parallel group the caller rank belongs to.""" assert _DATA_PARALLEL_GROUP is not None, \ 'data parallel group is not initialized' return _DATA_PARALLEL_GROUP def get_model_parallel_world_size(): """Return world size for the model parallel group.""" return torch.distributed.get_world_size(group=get_model_parallel_group()) def get_model_parallel_rank(): """Return my rank for the model parallel group.""" return torch.distributed.get_rank(group=get_model_parallel_group()) def get_model_parallel_src_rank(): """Calculate the global rank corresponding to a local rank zeor in the model parallel group.""" global_rank = torch.distributed.get_rank() local_world_size = get_model_parallel_world_size() return (global_rank // local_world_size) * local_world_size def get_data_parallel_world_size(): """Return world size for the data parallel group.""" return torch.distributed.get_world_size(group=get_data_parallel_group()) def get_data_parallel_rank(): """Return my rank for the data parallel group.""" return torch.distributed.get_rank(group=get_data_parallel_group()) def destroy_model_parallel(): """Set the groups to none.""" global _MODEL_PARALLEL_GROUP _MODEL_PARALLEL_GROUP = None global _DATA_PARALLEL_GROUP _DATA_PARALLEL_GROUP = None

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/LM/Flan-T5/mpu/cross_entropy.py

# coding=utf-8 import torch from .initialize import get_model_parallel_group from .initialize import get_model_parallel_rank from .initialize import get_model_parallel_world_size from .utils import VocabUtility class _VocabParallelCrossEntropy(torch.autograd.Function): @staticmethod def forward(ctx, vocab_parallel_logits, target): # Copy so the input remains unchanged. logits = vocab_parallel_logits.clone() # Maximum value along vocab dimension across all GPUs. logits_max = torch.max(logits, dim=-1)[0] torch.distributed.all_reduce(logits_max, op=torch.distributed.ReduceOp.MAX, group=get_model_parallel_group()) # Subtract the maximum value. logits.sub_(logits_max.unsqueeze(dim=-1)) # Sum of exponential of logits along vocab dimension across all GPUs. exp_logits = logits.exp() sum_exp_logits = exp_logits.sum(dim=-1) torch.distributed.all_reduce(sum_exp_logits, op=torch.distributed.ReduceOp.SUM, group=get_model_parallel_group()) # Get the partition's vocab indecies get_vocab_range = VocabUtility.vocab_range_from_per_partition_vocab_size partition_vocab_size = vocab_parallel_logits.size()[-1] rank = get_model_parallel_rank() world_size = get_model_parallel_world_size() vocab_start_index, vocab_end_index = get_vocab_range( partition_vocab_size, rank, world_size) # Create a mask of valid vocab ids (1 means it needs to be masked). target_mask = (target < vocab_start_index) | (target >= vocab_end_index) masked_target = target.clone() - vocab_start_index masked_target[target_mask] = 0 # Get predicted-logits = logits[target]. # For Simplicity, we convert logits to a 2-D tensor with size # [*, partition-vocab-size] and target to a 1-D tensor of size [*]. logits_2d = logits.view(-1, partition_vocab_size) masked_target_1d = masked_target.view(-1) arange_1d = torch.arange(start=0, end=logits_2d.size()[0], device=logits_2d.device) predicted_logits_1d = logits_2d[arange_1d, masked_target_1d] predicted_logits = predicted_logits_1d.view_as(target) predicted_logits[target_mask] = 0.0 # All reduce is needed to get the chunks from other GPUs. torch.distributed.all_reduce(predicted_logits, op=torch.distributed.ReduceOp.SUM, group=get_model_parallel_group()) # Loss = log(sum(exp(logits))) - predicted-logit. loss = torch.log(sum_exp_logits) - predicted_logits # Store softmax, target-mask and masked-target for backward pass. exp_logits.div_(sum_exp_logits.unsqueeze(dim=-1)) ctx.save_for_backward(exp_logits, target_mask, masked_target_1d) return loss @staticmethod def backward(ctx, grad_output): # Retreive tensors from the forward path. softmax, target_mask, masked_target_1d = ctx.saved_tensors # All the inputs have softmax as thier gradient. grad_input = softmax # For simplicity, work with the 2D gradient. partition_vocab_size = softmax.size()[-1] grad_2d = grad_input.view(-1, partition_vocab_size) # Add the gradient from matching classes. arange_1d = torch.arange(start=0, end=grad_2d.size()[0], device=grad_2d.device) grad_2d[arange_1d, masked_target_1d] -= ( 1.0 - target_mask.view(-1).float()) # Finally elementwise multiplication with the output gradients. grad_input.mul_(grad_output.unsqueeze(dim=-1)) return grad_input, None def vocab_parallel_cross_entropy(vocab_parallel_logits, target): """Helper function for the cross entropy.""" return _VocabParallelCrossEntropy.apply(vocab_parallel_logits, target) class _ParallelKLLoss(torch.autograd.Function): @staticmethod def forward(cls, logits: torch.Tensor, targets: torch.Tensor): # Maximum value along vocab dimension across all GPUs. logits_max = torch.max(logits, dim=-1)[0] torch.distributed.all_reduce(logits_max, op=torch.distributed.ReduceOp.MAX, group=get_model_parallel_group()) # Subtract the maximum value. logits.sub_(logits_max.unsqueeze(dim=-1)) # Sum of exponential of logits along vocab dimension across all GPUs. exp_logits = logits.exp() sum_exp_logits = exp_logits.sum(dim=-1) torch.distributed.all_reduce(sum_exp_logits, op=torch.distributed.ReduceOp.SUM, group=get_model_parallel_group()) targets_max = torch.max(targets, dim=-1)[0] torch.distributed.all_reduce(targets_max, op=torch.distributed.ReduceOp.MAX, group=get_model_parallel_group()) # Subtract the maximum value. targets.sub_(targets_max.unsqueeze(dim=-1)) # Sum of exponential of logits along vocab dimension across all GPUs. exp_targets = targets.exp() sum_exp_targets = exp_targets.sum(dim=-1) torch.distributed.all_reduce(sum_exp_targets, op=torch.distributed.ReduceOp.SUM, group=get_model_parallel_group()) # targets_softmax: [b, s, v_p] targets_softmax = torch.div(exp_targets, sum_exp_targets.unsqueeze(-1)) # sum_targets_softmax_logits: [b, s] sum_targets_softmax_logits = torch.matmul( targets_softmax.unsqueeze(-2), logits.unsqueeze(-1)).squeeze(-1).squeeze(-1) torch.distributed.all_reduce(sum_targets_softmax_logits, op=torch.distributed.ReduceOp.SUM, group=get_model_parallel_group()) log_targets_softmax = torch.log(targets_softmax) sum_log_targets_softmax = torch.matmul( targets_softmax.unsqueeze(-2), log_targets_softmax.unsqueeze(-1)).squeeze(-1).squeeze(-1) torch.distributed.all_reduce(sum_log_targets_softmax, op=torch.distributed.ReduceOp.SUM, group=get_model_parallel_group()) loss = torch.log(sum_exp_logits) - sum_targets_softmax_logits + sum_log_targets_softmax logits_softmax = torch.div(exp_logits, sum_exp_logits.unsqueeze(-1)) cls.save_for_backward(logits_softmax, targets_softmax) return loss @staticmethod def backward(cls, grad_output: torch.Tensor): logits_softmax, targets_softmax = cls.saved_tensors grad_input = (logits_softmax - targets_softmax) * grad_output.unsqueeze(-1) return grad_input, None def parallel_KL_loss(logits, targets): return _ParallelKLLoss.apply(logits, targets) class _ParallelSoftCrossEntropyLoss(torch.autograd.Function): @staticmethod def forward(cls, logits: torch.Tensor, targets: torch.Tensor): # Maximum value along vocab dimension across all GPUs. logits_max = torch.max(logits, dim=-1)[0] torch.distributed.all_reduce(logits_max, op=torch.distributed.ReduceOp.MAX, group=get_model_parallel_group()) # Subtract the maximum value. logits.sub_(logits_max.unsqueeze(dim=-1)) # Sum of exponential of logits along vocab dimension across all GPUs. exp_logits = logits.exp() sum_exp_logits = exp_logits.sum(dim=-1) torch.distributed.all_reduce(sum_exp_logits, op=torch.distributed.ReduceOp.SUM, group=get_model_parallel_group()) # sum_targets_softmax_logits: [b, s] sum_targets_softmax_logits = torch.matmul( targets.unsqueeze(-2), logits.unsqueeze(-1)).squeeze(-1).squeeze(-1) torch.distributed.all_reduce(sum_targets_softmax_logits, op=torch.distributed.ReduceOp.SUM, group=get_model_parallel_group()) loss = torch.log(sum_exp_logits) - sum_targets_softmax_logits logits_softmax = torch.div(exp_logits, sum_exp_logits.unsqueeze(-1)) cls.save_for_backward(logits_softmax, targets) return loss @staticmethod def backward(cls, grad_output: torch.Tensor): logits_softmax, targets = cls.saved_tensors grad_input = (logits_softmax - targets) * grad_output.unsqueeze(-1) return grad_input, None def parallel_soft_cross_entropy_loss(logits, targets): return _ParallelSoftCrossEntropyLoss.apply(logits, targets)

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/LM/Flan-T5/mpu/utils.py

# coding=utf-8 import torch def ensure_divisibility(numerator, denominator): """Ensure that numerator is divisible by the denominator.""" assert numerator % denominator == 0, '{} is not divisible by {}'.format( numerator, denominator) def divide(numerator, denominator): """Ensure that numerator is divisible by the denominator and return the division value.""" ensure_divisibility(numerator, denominator) return numerator // denominator def split_tensor_along_last_dim(tensor, num_partitions, contiguous_split_chunks=False): """Split a tensor along its last dimension. Arguments: tensor: input tensor. num_partitions: number of partitions to split the tensor contiguous_split_chunks: If True, make each chunk contiguous in memory. """ # Get the size and dimension. last_dim = tensor.dim() - 1 last_dim_size = divide(tensor.size()[last_dim], num_partitions) # Split. tensor_list = torch.split(tensor, last_dim_size, dim=last_dim) # Note: torch.split does not create contiguous tensors by default. if contiguous_split_chunks: return tuple(chunk.contiguous() for chunk in tensor_list) return tensor_list class VocabUtility: """Split the vocabulary into `world_size` chunks amd return the first and last index of the vocabulary belonging to the `rank` partition: Note that indecies in [fist, last)""" @staticmethod def vocab_range_from_per_partition_vocab_size(per_partition_vocab_size, rank, world_size): index_f = rank * per_partition_vocab_size index_l = index_f + per_partition_vocab_size return index_f, index_l @staticmethod def vocab_range_from_global_vocab_size(global_vocab_size, rank, world_size): per_partition_vocab_size = divide(global_vocab_size, world_size) return VocabUtility.vocab_range_from_per_partition_vocab_size( per_partition_vocab_size, rank, world_size)

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/LM/Flan-T5/mpu/data.py

# coding=utf-8 import torch from .initialize import get_model_parallel_group from .initialize import get_model_parallel_rank from .initialize import get_model_parallel_src_rank _MAX_DATA_DIM = 4 def _check_data_types(keys, data, target_dtype): """Check that all the keys have the same target data type.""" for key in keys: assert data[key].dtype == target_dtype, '{} has data type {} which '\ 'is different than {}'.format(key, data[key].dtype, target_dtype) def _build_key_size_numel_dictionaries(keys, data): """Build the size on rank 0 and broadcast.""" max_dim = _MAX_DATA_DIM sizes = [0 for _ in range(max_dim) for _ in keys] # Pack the sizes on rank zero. if get_model_parallel_rank() == 0: offset = 0 for key in keys: assert data[key].dim() < max_dim, 'you should increase MAX_DATA_DIM' size = data[key].size() for i, s in enumerate(size): sizes[i + offset] = s offset += max_dim # Move to GPU and broadcast. sizes_cuda = torch.cuda.LongTensor(sizes) torch.distributed.broadcast(sizes_cuda, get_model_parallel_src_rank(), group=get_model_parallel_group()) # Move back to cpu and unpack. sizes_cpu = sizes_cuda.cpu() key_size = {} key_numel = {} total_numel = 0 offset = 0 for key in keys: i = 0 size = [] numel = 1 while sizes_cpu[offset + i] > 0: this_size = sizes_cpu[offset + i] size.append(this_size) numel *= this_size i += 1 key_size[key] = size key_numel[key] = numel total_numel += numel offset += max_dim return key_size, key_numel, total_numel def broadcast_data(keys, data, datatype): """Broadcast data from rank zero of each model parallel group to the members of the same model parallel group. Arguments: keys: list of keys in the data disctionary to be broadcasted data: data dictionary of string keys and cpu tensor values. datatype: torch data type of all tensors in data associated with keys. """ # Build (key, size) and (key, number of elements) dictionaries along # with the total number of elements on all ranks. key_size, key_numel, total_numel = _build_key_size_numel_dictionaries(keys, data) # Pack on rank zero. if get_model_parallel_rank() == 0: # Check that all keys have the same data type. _check_data_types(keys, data, datatype) # Flatten the data associated with the keys flatten_data = torch.cat( [data[key].contiguous().view(-1) for key in keys], dim=0).cuda() else: flatten_data = torch.empty(total_numel, device=torch.cuda.current_device(), dtype=datatype) # Boradcast torch.distributed.broadcast(flatten_data, get_model_parallel_src_rank(), group=get_model_parallel_group()) # Unpack output = {} offset = 0 for key in keys: size = key_size[key] numel = key_numel[key] output[key] = flatten_data.narrow(0, offset, numel).view(size) offset += numel return output

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/LM/Flan-T5/mpu/grads.py

# coding=utf-8 # Parts of the code here are adapted from PyTorch # repo: https://github.com/pytorch/pytorch import torch from torch._six import inf from .initialize import get_model_parallel_group from .initialize import get_model_parallel_rank def clip_grad_norm(parameters, max_norm, norm_type=2): """Clips gradient norm of an iterable of parameters. This is adapted from torch.nn.utils.clip_grad.clip_grad_norm_ and added functionality to handle model parallel parameters. Note that the gradients are modified in place. Arguments: parameters (Iterable[Tensor] or Tensor): an iterable of Tensors or a single Tensor that will have gradients normalized max_norm (float or int): max norm of the gradients norm_type (float or int): type of the used p-norm. Can be ``'inf'`` for infinity norm. Returns: Total norm of the parameters (viewed as a single vector). """ if isinstance(parameters, torch.Tensor): parameters = [parameters] parameters = list(filter(lambda p: p.grad is not None, parameters)) max_norm = float(max_norm) norm_type = float(norm_type) if norm_type == inf: total_norm = max(p.grad.data.abs().max() for p in parameters) total_norm_cuda = torch.cuda.FloatTensor([float(total_norm)]) # Take max across all GPUs. torch.distributed.all_reduce(total_norm_cuda, op=torch.distributed.ReduceOp.MAX, group=get_model_parallel_group()) total_norm = total_norm_cuda[0].item() else: total_norm = 0 for p in parameters: if p.model_parallel or (get_model_parallel_rank() == 0): param_norm = p.grad.data.norm(norm_type) total_norm += param_norm.item() ** norm_type # Sum across all model parallel GPUs. total_norm_cuda = torch.cuda.FloatTensor([float(total_norm)]) torch.distributed.all_reduce(total_norm_cuda, op=torch.distributed.ReduceOp.SUM, group=get_model_parallel_group()) total_norm = total_norm_cuda[0].item() ** (1. / norm_type) clip_coef = max_norm / (total_norm + 1e-6) if clip_coef < 1: for p in parameters: p.grad.data.mul_(clip_coef) return total_norm

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/LM/Flan-T5/mpu/layers.py

# coding=utf-8 # Parts of the code here are adapted from PyTorch # repo: https://github.com/pytorch/pytorch import math import torch import torch.nn.functional as F import torch.nn.init as init from torch.nn.parameter import Parameter from apex.normalization.fused_layer_norm import FusedLayerNorm as LayerNorm from .initialize import get_model_parallel_rank from .initialize import get_model_parallel_world_size from .mappings import copy_to_model_parallel_region from .mappings import gather_from_model_parallel_region from .mappings import reduce_from_model_parallel_region from .mappings import scatter_to_model_parallel_region from .random import get_cuda_rng_tracker from .utils import divide from .utils import split_tensor_along_last_dim from .utils import VocabUtility def _initialize_affine_weight(weight, output_size, input_size, per_partition_size, partition_dim, init_method, stride=1, return_master_weight=False): """Initialize affine weight for model parallel. Build the master weight on all processes and scatter the relevant chunk.""" # If we only use 1 process for model parallelism, bypass scatter. world_size = get_model_parallel_world_size() if world_size == 1: init_method(weight) if return_master_weight: return weight return None # Initialize master weight master_weight = torch.empty(output_size, input_size, dtype=weight.dtype, requires_grad=False) init_method(master_weight) # Split and copy per_partition_per_stride_size = divide(per_partition_size, stride) weight_list = torch.split(master_weight, per_partition_per_stride_size, dim=partition_dim) rank = get_model_parallel_rank() my_weight_list = weight_list[rank::world_size] with torch.no_grad(): torch.cat(my_weight_list, dim=partition_dim, out=weight) if return_master_weight: return master_weight return None class VocabParallelEmbedding(torch.nn.Module): """Embedding parallelized in the vocabulary dimension. This is mainly adapted from torch.nn.Embedding and all the default values are kept. Arguments: num_embeddings: vocabulary size. embedding_dim: size of hidden state. init_method: method to initialize weights. """ def __init__(self, num_embeddings, embedding_dim, init_method=init.xavier_normal_): super(VocabParallelEmbedding, self).__init__() # Keep the input dimensions. self.num_embeddings = num_embeddings self.embedding_dim = embedding_dim # Set the detauls for compatibility. self.padding_idx = None self.max_norm = None self.norm_type = 2. self.scale_grad_by_freq = False self.sparse = False self._weight = None # Divide the weight matrix along the vocaburaly dimension. self.vocab_start_index, self.vocab_end_index = \ VocabUtility.vocab_range_from_global_vocab_size( self.num_embeddings, get_model_parallel_rank(), get_model_parallel_world_size()) self.num_embeddings_per_partition = self.vocab_end_index - \ self.vocab_start_index # Allocate weights. self.weight = Parameter(torch.Tensor(self.num_embeddings_per_partition, self.embedding_dim)) self.weight.model_parallel = True # And initialize. _initialize_affine_weight( self.weight, self.num_embeddings, self.embedding_dim, self.num_embeddings_per_partition, 0, init_method) def forward(self, input_): # Build the mask. input_mask = (input_ < self.vocab_start_index) | \ (input_ >= self.vocab_end_index) # Mask the input. masked_input = input_.clone() - self.vocab_start_index masked_input[input_mask] = 0 # Get the embeddings. output_parallel = F.embedding(masked_input, self.weight, self.padding_idx, self.max_norm, self.norm_type, self.scale_grad_by_freq, self.sparse) # Mask the output embedding. output_parallel[input_mask, :] = 0.0 # Reduce across all the model parallel GPUs. output = reduce_from_model_parallel_region(output_parallel) return output class ParallelEmbedding(torch.nn.Module): """Embedding parallelized in the embedding dimension. This is mainly adapted from torch.nn.Embedding and all the default values are kept. Arguments: num_embeddings: vocabulary size. embedding_dim: size of hidden state. init_method: method to initialize weights. """ def __init__(self, num_embeddings, embedding_dim, init_method=init.xavier_normal_, keep_master_weight_for_test=False): super(ParallelEmbedding, self).__init__() # Keep the input dimensions. self.num_embeddings = num_embeddings self.embedding_dim = embedding_dim # Set some detauls for compatibility. self.padding_idx = None self.max_norm = None self.norm_type = 2. self.scale_grad_by_freq = False self.sparse = False self._weight = None # Divide the weight matrix along the embedding dimension. world_size = get_model_parallel_world_size() self.embedding_dim_per_partition = divide(self.embedding_dim, world_size) # Allocate weights. self.weight = Parameter(torch.Tensor(self.num_embeddings, self.embedding_dim_per_partition)) self.weight.model_parallel = True # And initialize. split the weights to different model parallel devices _initialize_affine_weight( self.weight, self.num_embeddings, self.embedding_dim, self.embedding_dim_per_partition, 1, init_method, stride=1, return_master_weight=False) def forward(self, input_): input_parallel = copy_to_model_parallel_region(input_) output_parallel = F.embedding(input_parallel, self.weight, self.padding_idx, self.max_norm, self.norm_type, self.scale_grad_by_freq, self.sparse) output = gather_from_model_parallel_region(output_parallel) return output class ColumnParallelLinear(torch.nn.Module): """Linear layer with column parallelism. NOTE: This function will NOT do all-reduce unless gather_output is True The linear layer is defined as Y = XA + b. A is parallelized along its second dimension as A = [A_1, ..., A_p]. Arguments: input_size: first dimension of matrix A. output_size: second dimension of matrix A. bias: If true, add bias gather_output: If true, call all-gether on output and make Y avaiable to all GPUs, otherwise, every GPU will have its output which is Y_i = XA_i init_method: method to initialize weights. Note that bias is always set to zero. stride: For the strided linear layers. keep_master_weight_for_test: This was added for testing and should be set to False. It returns the master weights used for initialization. """ def __init__(self, input_size, output_size, bias=True, gather_output=True, init_method=init.xavier_normal_, stride=1, keep_master_weight_for_test=False): super(ColumnParallelLinear, self).__init__() # Keep input parameters self.input_size = input_size self.output_size = output_size self.gather_output = gather_output # Divide the weight matrix along the last dimension. world_size = get_model_parallel_world_size() self.output_size_per_partition = divide(output_size, world_size) # Parameters. # Note: torch.nn.functional.linear performs XA^T + b and as a result # we allocate the transpose. self.weight = Parameter(torch.Tensor(self.output_size_per_partition, self.input_size)) self.weight.model_parallel = True if bias: self.bias = Parameter(torch.Tensor(self.output_size_per_partition)) self.bias.model_parallel = True # Always initialize bias to zero. with torch.no_grad(): self.bias.zero_() else: self.register_parameter('bias', None) # Initialize weight. self.master_weight = _initialize_affine_weight( self.weight, self.output_size, self.input_size, self.output_size_per_partition, 0, init_method, stride=stride, return_master_weight=keep_master_weight_for_test) def forward(self, input_): # Set up backprop all-reduce. input_parallel = copy_to_model_parallel_region(input_) # Matrix multiply. output_parallel = F.linear(input_parallel, self.weight, self.bias) if self.gather_output: # All-gather across the partitions. output = gather_from_model_parallel_region(output_parallel) else: output = output_parallel return output class RowParallelLinear(torch.nn.Module): """Linear layer with row parallelism. NOTE: This function will do all-reduce The linear layer is defined as Y = XA + b. A is parallelized along its first dimension and X along its second dimension as: - - | A_1 | | . | A = | . | X = [X_1, ..., X_p] | . | | A_p | - - Arguments: input_size: first dimension of matrix A. output_size: second dimension of matrix A. bias: If true, add bias. Note that bias is not parallelized. input_is_parallel: If true, we assume that the input is already split across the GPUs and we do not split again. init_method: method to initialize weights. Note that bias is always set to zero. stride: For the strided linear layers. keep_master_weight_for_test: This was added for testing and should be set to False. It returns the master weights used for initialization. """ def __init__(self, input_size, output_size, bias=True, input_is_parallel=False, init_method=init.xavier_normal_, stride=1, keep_master_weight_for_test=False): super(RowParallelLinear, self).__init__() # Keep input parameters self.input_size = input_size self.output_size = output_size self.input_is_parallel = input_is_parallel # Divide the weight matrix along the last dimension. world_size = get_model_parallel_world_size() self.input_size_per_partition = divide(input_size, world_size) # Parameters. # Note: torch.nn.functional.linear performs XA^T + b and as a result # we allocate the transpose. self.weight = Parameter(torch.Tensor(self.output_size, self.input_size_per_partition)) self.weight.model_parallel = True if bias: self.bias = Parameter(torch.Tensor(self.output_size)) # Always initialize bias to zero. with torch.no_grad(): self.bias.zero_() else: self.register_parameter('bias', None) # Initialize weight. self.master_weight = _initialize_affine_weight( self.weight, self.output_size, self.input_size, self.input_size_per_partition, 1, init_method, stride=stride, return_master_weight=keep_master_weight_for_test) def forward(self, input_): # Set up backprop all-reduce. if self.input_is_parallel: input_parallel = input_ else: input_parallel = scatter_to_model_parallel_region(input_) # Matrix multiply. output_parallel = F.linear(input_parallel, self.weight) # All-reduce across all the partitions. output_ = reduce_from_model_parallel_region(output_parallel) if self.bias is not None: output = output_ + self.bias else: output = output_ return output

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/LM/Flan-T5/mpu/random.py

# coding=utf-8 #Modified by Samyam Rajbhandari #Used to partition the activations stored for backward propagation #Therefore reduces the memory consumption # Parts of the code here are adapted from PyTorch # repo: https://github.com/pytorch/pytorch import contextlib import torch.distributed as dist import torch from torch import _C from torch.cuda import _lazy_call, device as device_ctx_manager #from torch.utils.checkpoint import detach_variable import torch.distributed as dist PARTITION_ACTIVATIONS = False PA_CORRECTNESS_TEST= False def see_memory_usage(message, force=False): if not force: return dist.barrier() if dist.get_rank() == 0: print(message) print("Memory Allocated ", torch.cuda.memory_allocated()/(1024*1024*1024), "GigaBytes") print("Max Memory Allocated ", torch.cuda.max_memory_allocated()/(1024*1024*1024), "GigaBytes") print("Cache Allocated ", torch.cuda.memory_cached()/(1024*1024*1024), "GigaBytes") print("Max cache Allocated ", torch.cuda.max_memory_cached()/(1024*1024*1024), "GigaBytes") print(" ") #input("Press Any Key To Continue ..") from .initialize import get_data_parallel_rank from .initialize import get_model_parallel_rank from .initialize import get_model_parallel_world_size from .initialize import get_model_parallel_group mp_rank = None #get_model_parallel_rank() mp_size = None #get_model_parallel_world_size() mp_group = None #get_model_parallel_group() # Default name for the model parallel rng tracker. _MODEL_PARALLEL_RNG_TRACKER_NAME = 'model-parallel-rng' transport_stream = None cuda_device=None def detach_variable(inputs, device=None): if isinstance(inputs, tuple): out = [] for inp in inputs: if not isinstance(inp, torch.Tensor): out.append(inp) continue requires_grad = inp.requires_grad if device is not None: x = inp.to(device=device) else: x = inp x = x.detach() x.requires_grad = requires_grad out.append(x) return tuple(out) else: raise RuntimeError( "Only tuple of tensors is supported. Got Unsupported input type: ", type(inputs).__name__) def _set_cuda_rng_state(new_state, device=-1): """Sets the random number generator state of the current GPU. Argumentss: new_state (torch.ByteTensor): The desired state This function is adapted from PyTorch repo (torch.cuda.set_rng_state) with a single change: the input state is not cloned. Cloning caused major performance issues for +4 GPU cases. """ if hasattr(_C, '_cuda_setRNGState') and callable(_C._cuda_setRNGState): # older PyTorch def cb(): with device_ctx_manager(device): _C._cuda_setRNGState(new_state) else: # newer PyTorch if device == -1: device = torch.device('cuda') elif isinstance(device, str): device = torch.device(device) elif isinstance(device, int): device = torch.device('cuda', device) def cb(): idx = device.index if idx is None: idx = torch.cuda.current_device() default_generator = torch.cuda.default_generators[idx] default_generator.set_state(new_state) _lazy_call(cb) class CudaRNGStatesTracker: """Tracker for the cuda RNG states. Using the `add` method, a cuda rng state is initialized based on the input `seed` and is assigned to `name`. Later, by forking the rng state, we can perform operations and return to our starting cuda state. """ def __init__(self): # Map from a string name to the cuda rng state. self.states_ = {} # Seeds are just for book keeping and ensure no seed is set twice. self.seeds_ = set() def reset(self): """Set to the initial state (no tracker).""" self.states_ = {} self.seeds_ = set() def get_states(self): """Get rng states. Copy the dictionary so we have direct pointers to the states, not just a pointer to the dictionary.""" states = {} for name in self.states_: states[name] = self.states_[name] return states def set_states(self, states): """Set the rng states. For efficiency purposes, we do not check the size of seed for compatibility.""" self.states_ = states def add(self, name, seed): """Track the rng state.""" # Check seed is not already used. if seed in self.seeds_: raise Exception('seed {} already exists'.format(seed)) self.seeds_.add(seed) # Check that state is not already defined. if name in self.states_: raise Exception('cuda rng state {} already exists'.format(name)) # Get the current rng state. orig_rng_state = torch.cuda.get_rng_state() # Set the new state and store it. torch.cuda.manual_seed(seed) self.states_[name] = torch.cuda.get_rng_state() # Reset rng state to what it was. _set_cuda_rng_state(orig_rng_state) @contextlib.contextmanager def fork(self, name=_MODEL_PARALLEL_RNG_TRACKER_NAME): """Fork the cuda rng state, perform operations, and exit with the original state.""" # Check if we have added the state if name not in self.states_: raise Exception('cuda rng state {} is not added'.format(name)) # Store current rng state. orig_cuda_rng_state = torch.cuda.get_rng_state() # Set rng state to the desired one _set_cuda_rng_state(self.states_[name]) # Do the stuff we wanted to do. try: yield finally: # Update the current rng state for later use. self.states_[name] = torch.cuda.get_rng_state() # And set the state to the original state we started with. _set_cuda_rng_state(orig_cuda_rng_state) # RNG tracker object. _CUDA_RNG_STATE_TRACKER = CudaRNGStatesTracker() def get_cuda_rng_tracker(): """Get cuda rng tracker.""" return _CUDA_RNG_STATE_TRACKER def model_parallel_cuda_manual_seed(seed): """Initialize model parallel cuda seed. This function should be called after the model parallel is initialized. Also, no torch.cuda.manual_seed should be called after this function. Basically, this is replacement for that function. Two set of RNG states are tracked: default state: This is for data parallelism and is the same among a set of model parallel GPUs but different across different model paralle groups. This is used for example for dropout in the non-model-parallel regions. model-parallel state: This state is different among a set of model parallel GPUs, but the same across data parallel groups. This is used for example for dropout in model parallel regions. """ # 2718 is just for fun and any POSITIVE value will work. offset = seed + 2718 model_parallel_seed = offset + get_model_parallel_rank() # Data parallel gets the original sedd. data_parallel_seed = seed if torch.distributed.get_rank() == 0: print('> initializing model parallel cuda seeds on global rank {}, ' 'model parallel rank {}, and data parallel rank {} with ' 'model parallel seed: {} and data parallel seed: {}'.format( torch.distributed.get_rank(), get_model_parallel_rank(), get_data_parallel_rank(), model_parallel_seed, data_parallel_seed), flush=True) _CUDA_RNG_STATE_TRACKER.reset() # Set the default state. torch.cuda.manual_seed(data_parallel_seed) # and model parallel state. _CUDA_RNG_STATE_TRACKER.add(_MODEL_PARALLEL_RNG_TRACKER_NAME, model_parallel_seed) def get_partition_start(item): global mp_rank, mp_size, mp_group partition_size = get_partition_size(item) start = partition_size * mp_rank return int(start) def get_partition_size(item): global mp_rank, mp_size, mp_group size = item.numel() partition_size = size/mp_size return int(partition_size) def get_full_inputs(tensors): inputs=[] for i in range(int(len(tensors)/2)-1): item = tensors[2 * i] size = tensors[2* i + 1] partition_size = item.numel() tensor_size = partition_size * mp_size flat_tensor = torch.zeros([tensor_size], dtype=item.dtype, device=item.device) partitions=[] for i in range(mp_size): part_i = flat_tensor.narrow(0, partition_size * i , partition_size) if i == mp_rank: part_i.copy_(item) partitions.append(part_i) dist.all_gather(partitions,partitions[mp_rank], group=mp_group) input_tensor = flat_tensor.view(list(size.numpy())) item.data=input_tensor.data inputs.append(item) inputs.append(tensors[-2]) return tuple(inputs) class CheckpointFunction(torch.autograd.Function): """This function is adapted from torch.utils.checkpoint with two main changes: 1) torch.cuda.set_rng_state is replaced with `_set_cuda_rng_state` 2) the states in the model parallel tracker are also properly tracked/set/reset. """ @staticmethod def forward(ctx, run_function, *args): ctx.run_function = run_function global mp_rank, mp_size, mp_group if mp_rank is None: mp_rank = get_model_parallel_rank() mp_size = get_model_parallel_world_size() mp_group = get_model_parallel_group() global cuda_device, transport_stream, PARTITION_ACTIVATIONS if cuda_device is None: if dist.get_rank() == 0: print(f"Partition Activations {PARTITION_ACTIVATIONS} and Correctness Check {PA_CORRECTNESS_TEST}") cuda_device = torch.cuda.current_device() #The transport stream is used to overlap the allgather communication for the activations #with the computation in the backward pass transport_stream = torch.cuda.Stream(device=cuda_device) if PARTITION_ACTIVATIONS: inputs = [item.detach().contiguous().view(-1).narrow(0, get_partition_start(item), get_partition_size(item)).clone() for item in args[:-1]] inputs.append(args[-1]) #just in case something funky is happening such as reuse of inputs inputs_cuda = [item.to(cuda_device) if isinstance(item, torch.Tensor) else item for item in args] # Copy the rng states. ctx.fwd_cpu_rng_state = torch.get_rng_state() ctx.fwd_cuda_rng_state = torch.cuda.get_rng_state() ctx.fwd_cuda_rng_state_tracker = get_cuda_rng_tracker().get_states() #ctx.save_for_backward(*args) with torch.no_grad(): outputs = run_function(*inputs_cuda) del inputs_cuda if PARTITION_ACTIVATIONS: new_args = [] for arg, inp in zip(args,inputs): size= torch.tensor(arg.size()) arg.data = inp.data new_args.append(arg) new_args.append(size) ctx.save_for_backward(*new_args) else: ctx.save_for_backward(*args) return outputs @staticmethod def backward(ctx, *args): if not torch.autograd._is_checkpoint_valid(): raise RuntimeError("Checkpointing is not compatible with .grad(), " "please use .backward() if possible") global cuda_device, transport_stream, PARTITION_ACTIVATIONS if PARTITION_ACTIVATIONS: with torch.cuda.stream(transport_stream): inputs = get_full_inputs(ctx.saved_tensors) detached_inputs = detach_variable(inputs) else: inputs = ctx.saved_tensors detached_inputs = detach_variable(inputs) # Store the current states. bwd_cpu_rng_state = torch.get_rng_state() bwd_cuda_rng_state = torch.cuda.get_rng_state() bwd_cuda_rng_state_tracker = get_cuda_rng_tracker().get_states() # Set the states to what it used to be before the forward pass. torch.set_rng_state(ctx.fwd_cpu_rng_state) _set_cuda_rng_state(ctx.fwd_cuda_rng_state) get_cuda_rng_tracker().set_states(ctx.fwd_cuda_rng_state_tracker) if PARTITION_ACTIVATIONS: current_stream=torch.cuda.current_stream() current_stream.wait_stream(transport_stream) with torch.enable_grad(): outputs = ctx.run_function(*detached_inputs) # Set the states back to what it was at the start of this function. torch.set_rng_state(bwd_cpu_rng_state) _set_cuda_rng_state(bwd_cuda_rng_state) get_cuda_rng_tracker().set_states(bwd_cuda_rng_state_tracker) if isinstance(outputs, torch.Tensor): outputs = (outputs,) torch.autograd.backward(outputs, args) return (None,) + tuple(inp.grad if isinstance(inp, torch.Tensor) else None for inp in detached_inputs) def checkpoint(function, *args): """Checkpoint a model or part of the model. This has been directly copied from torch.utils.checkpoint.""" return CheckpointFunction.apply(function, *args) def partition_activations_in_checkpoint(partition_activation): global PARTITION_ACTIVATIONS PARTITION_ACTIVATIONS=partition_activation if dist.get_rank() == 0: print(f"**************Partition Activations {PARTITION_ACTIVATIONS}************")

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/LM/Flan-T5/mpu/transformer_enc_dec.py

from audioop import cross import math from numpy.lib.function_base import insert import torch import torch.nn as nn import torch.nn.init as init import torch.nn.functional as F # from apex.normalization.fused_layer_norm import FusedLayerNorm as LayerNorm from .initialize import get_model_parallel_world_size from .layers import ColumnParallelLinear from .layers import RowParallelLinear from .mappings import gather_from_model_parallel_region import deepspeed import pickle from .random import checkpoint from .random import get_cuda_rng_tracker from .utils import divide from .utils import split_tensor_along_last_dim from model.configuration_enc_dec import EncDecConfig from .layers import VocabParallelEmbedding from typing import Callable, Optional, List class LayerNorm(nn.Module): def __init__(self, hidden_size, eps=1e-6): """ Construct a layernorm module in the T5 style No bias and no subtraction of mean. """ super().__init__() self.weight = nn.Parameter(torch.ones(hidden_size)) # self.bias = nn.Parameter(torch.zeros(hidden_size)) self.eps = eps def forward(self, hidden_states): # layer norm should always be calculated in float32 variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True) hidden_states = hidden_states * torch.rsqrt(variance + self.eps) # convert into float16 if necessary if self.weight.dtype == torch.float16: hidden_states = hidden_states.to(torch.float16) return self.weight * hidden_states @torch.jit.script def gelu_impl(x): """OpenAI's gelu implementation.""" return 0.5 * x * (1.0 + torch.tanh(0.7978845608028654 * x * (1.0 + 0.044715 * x * x))) def gelu(x): return gelu_impl(x) def unscaled_init_method(sigma): """Init method based on N(0, sigma).""" def init_(tensor): return torch.nn.init.normal_(tensor, mean=0.0, std=sigma) return init_ def scaled_init_method(sigma, num_layers): """Init method based on N(0, sigma/sqrt(2*num_layers).""" std = sigma / math.sqrt(2.0 * num_layers) def init_(tensor): return torch.nn.init.normal_(tensor, mean=0.0, std=std) return init_ def init_method_normal(std): """Init method based on normal distribution. This is only used for embeddings. The transformer has its own initializer. """ def init_(tensor): return torch.nn.init.normal_(tensor, mean=0.0, std=std) return init_ class ParallelDenseReluDense(nn.Module): def __init__(self, config: EncDecConfig, init_method: Callable, output_layer_init_method: Optional[Callable] = None): super(ParallelDenseReluDense, self).__init__() self.wi_0 = ColumnParallelLinear( config.d_model, config.d_ff, gather_output=False, bias=False, init_method=init_method) self.wi_1 = ColumnParallelLinear( config.d_model, config.d_ff, gather_output=False, bias=False, init_method=init_method) self.wo = RowParallelLinear( config.d_ff, config.d_model, bias=False, input_is_parallel=True, init_method=output_layer_init_method) self.dropout = nn.Dropout(config.dropout_rate) # self.do_dim_trick = config.do_dim_trick # if torch.distributed.get_rank() % 5 == 4: # self.ff_mask = nn.Parameter(torch.tensor([1.0] * 13104 + [0.0] * 4), requires_grad=False) # else: # self.ff_mask = nn.Parameter(torch.tensor([1.0] * 13108), requires_grad=False) def forward(self, hidden_states): # hidden_states: [b, s, hp] hidden_gelu = gelu(self.wi_0(hidden_states)) hidden_linear = self.wi_1(hidden_states) hidden_states = hidden_gelu * hidden_linear # hidden_states: [b, s, d_ff_p] # if self.do_dim_trick: # ff_mask = self.ff_mask.view(1, 1, self.ff_mask.size(0)) # hidden_states = ff_mask * hidden_states # hidden_states = F.relu(hidden_states) hidden_states = self.dropout(hidden_states) hidden_states = self.wo(hidden_states) # hidden_states: [b, s, hp] return hidden_states class ParallelAttention(nn.Module): def __init__( self, config: EncDecConfig, init_method: Callable, is_decoder: bool = False, is_cross_attn: bool = False, output_layer_init_method: Optional[Callable] = None, has_relative_attention_bias: bool = False): super(ParallelAttention, self).__init__() self.is_decoder = is_decoder self.is_cross_attn = is_cross_attn self.output_attention = config.output_attention self.has_relative_attention_bias = has_relative_attention_bias self.relative_attention_num_buckets = config.relative_attention_num_buckets # Set output layer initialization if not provided. if output_layer_init_method is None: output_layer_init_method = init_method d_attn_out = config.d_kv * config.num_heads # h # Per attention head and per partition values. world_size = get_model_parallel_world_size() # p self.hidden_size_per_partition = divide(d_attn_out, world_size) # h_p self.hidden_size_per_attention_head = config.d_kv # h_i self.num_attention_heads_per_partition = divide(config.num_heads, world_size) # n_p # Strided linear layer. if is_cross_attn: self.project_q = ColumnParallelLinear(config.d_model, d_attn_out, stride=1, # NOTE: modify stride bias=False, gather_output=False, init_method=init_method) self.project_kv = ColumnParallelLinear(config.d_model, 2 * d_attn_out, stride=2, # NOTE: modify stride bias=False, gather_output=False, init_method=init_method) else: self.project = ColumnParallelLinear(config.d_model, 3 * d_attn_out, stride=3, bias=False, gather_output=False, init_method=init_method) if self.has_relative_attention_bias: self.relative_attention_bias = nn.Embedding(self.relative_attention_num_buckets, self.num_attention_heads_per_partition) # Dropout. Note that for a single iteration, this layer will generate # different outputs on different number of parallel partitions but # on average it should not be partition dependent. self.attention_dropout = nn.Dropout(config.dropout_rate) # Output. self.dense = RowParallelLinear(d_attn_out, config.d_model, input_is_parallel=True, bias=False, init_method=output_layer_init_method) self.output_dropout = nn.Dropout(config.dropout_rate) if deepspeed.checkpointing.is_configured(): global get_cuda_rng_tracker, checkpoint get_cuda_rng_tracker = deepspeed.checkpointing.get_cuda_rng_tracker checkpoint = deepspeed.checkpointing.checkpoint def _transpose_for_scores(self, tensor): """Transpose a 3D tensor [b, s, h_p=n_p*h_i] into a 4D tensor with size [b, np, s, hn]. """ new_tensor_shape = tensor.size()[:-1] + \ (self.num_attention_heads_per_partition, self.hidden_size_per_attention_head) # [b, s, n_p, h_i] tensor = tensor.view(*new_tensor_shape) # tensor: [b, n_p, s, h_i] return tensor.permute(0, 2, 1, 3) @staticmethod def _relative_position_bucket(relative_position, bidirectional=True, num_buckets=32, max_distance=128): """ Adapted from Mesh Tensorflow: https://github.com/tensorflow/mesh/blob/0cb87fe07da627bf0b7e60475d59f95ed6b5be3d/mesh_tensorflow/transformer/transformer_layers.py#L593 Translate relative position to a bucket number for relative attention. The relative position is defined as memory_position - query_position, i.e. the distance in tokens from the attending position to the attended-to position. If bidirectional=False, then positive relative positions are invalid. We use smaller buckets for small absolute relative_position and larger buckets for larger absolute relative_positions. All relative positions >=max_distance map to the same bucket. All relative positions <=-max_distance map to the same bucket. This should allow for more graceful generalization to longer sequences than the model has been trained on Args: relative_position: an int32 Tensor bidirectional: a boolean - whether the attention is bidirectional num_buckets: an integer max_distance: an integer Returns: a Tensor with the same shape as relative_position, containing int32 values in the range [0, num_buckets) """ relative_buckets = 0 if bidirectional: num_buckets //= 2 relative_buckets += (relative_position > 0).to(torch.long) * num_buckets relative_position = torch.abs(relative_position) else: relative_position = -torch.min(relative_position, torch.zeros_like(relative_position)) # now relative_position is in the range [0, inf) # half of the buckets are for exact increments in positions max_exact = num_buckets // 2 is_small = relative_position < max_exact # The other half of the buckets are for logarithmically bigger bins in positions up to max_distance relative_postion_if_large = max_exact + ( torch.log(relative_position.float() / max_exact) / math.log(max_distance / max_exact) * (num_buckets - max_exact) ).to(torch.long) relative_postion_if_large = torch.min( relative_postion_if_large, torch.full_like(relative_postion_if_large, num_buckets - 1) ) relative_buckets += torch.where(is_small, relative_position, relative_postion_if_large) return relative_buckets def compute_bias(self, query_length, key_length): """ Compute binned relative position bias """ context_position = torch.arange(query_length, dtype=torch.long)[:, None] memory_position = torch.arange(key_length, dtype=torch.long)[None, :] relative_position = memory_position - context_position # shape (query_length, key_length) relative_position_bucket = self._relative_position_bucket( relative_position, # shape (query_length, key_length) bidirectional=(not self.is_decoder), num_buckets=self.relative_attention_num_buckets, ) relative_position_bucket = relative_position_bucket.to(self.relative_attention_bias.weight.device) values = self.relative_attention_bias(relative_position_bucket) # shape (query_length, key_length, num_heads) values = values.permute([2, 0, 1]).unsqueeze(0) # shape (1, num_heads, query_length, key_length) return values def forward( self, hidden_states, attention_mask=None, key_value_states=None, position_bias=None, query_length=None, past_key_value=None,): batch_size, seq_length = hidden_states.shape[:2] real_seq_length = seq_length if past_key_value is not None: assert ( len(past_key_value) == 2 ), "past_key_value should have 2 past states: keys and values. Got {} past states".format( len(past_key_value) ) real_seq_length += past_key_value[0].shape[2] if query_length is None else query_length key_length = real_seq_length if key_value_states is None else key_value_states.shape[1] # hidden_states: [b, s, d_model] if key_value_states is not None: assert self.is_cross_attn is True # mixed_query_layer: [b, s, h_p] mixed_query_layer = self.project_q(hidden_states) # mixed_key_value_layer: [b, s, 2 * h_p] mixed_key_value_layer = self.project_kv(key_value_states) (mixed_key_layer, mixed_value_layer) = split_tensor_along_last_dim(mixed_key_value_layer, 2) else: assert self.is_cross_attn is False # hidden_states: [b, s, h] mixed_x_layer = self.project(hidden_states) # mixed_x_layer: [b, s, 3 * h_p] (mixed_query_layer, mixed_key_layer, mixed_value_layer) = split_tensor_along_last_dim(mixed_x_layer, 3) # mixed_***_layer: [b, s, h_p] # ***_layer [b, n_p, s, h_i] query_layer = self._transpose_for_scores(mixed_query_layer) key_layer = self._transpose_for_scores(mixed_key_layer) value_layer = self._transpose_for_scores(mixed_value_layer) if past_key_value is not None and not self.is_cross_attn: assert self.is_decoder is True # decoder # ***_layer: [b, n_p, 1, h_i] past_key_layer, past_value_layer = past_key_value # past_***_layer: [b, n_p, s-1, h_i] key_layer = torch.cat([past_key_layer, key_layer], dim=2) value_layer = torch.cat([past_value_layer, value_layer], dim=2) # ***_layer: [b, n_p, s_k, h_i] # Raw attention scores. [b, n_p, s_q, s_k] compute every head alone attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2)) # NOTE: We follow the implementation of Transformers to remove the scale of attention+acores # attention_scores = attention_scores / math.sqrt( # self.hidden_size_per_attention_head) # relative positional bias if position_bias is None: if not self.has_relative_attention_bias: position_bias = torch.zeros( (1, self.num_attention_heads_per_partition, real_seq_length, key_length), device=attention_scores.device, dtype=attention_scores.dtype ) else: position_bias = self.compute_bias(real_seq_length, key_length) # if key and values are already calculated # we want only the last query position bias if past_key_value is not None: position_bias = position_bias[:, :, -seq_length:, :] # if torch.distributed.get_rank() == 0: # print(real_seq_length, key_length, position_bias[0, 0, 0]) no_pos_bias_attn_probs = nn.Softmax(dim=-1)(attention_scores) # Apply the attention mask [b, 1, s_q, s_k] and relative position_bias # NOTE: 10000 can't be larger otherwise may cause fp16 overflow (max in fp16 = 65504) attention_scores = torch.mul(attention_scores, attention_mask) + (-10000.0 * (1.0 - attention_mask) + position_bias) # attention_scores = torch.mul(attention_scores, attention_mask) - 10000.0 * (1.0 - attention_mask) if hasattr(self, "score_storage"): if self.score_storage is None: self.score_storage = attention_scores[:, :, 0:1, :] # Attention probabilities. [b, n_p, s_q, s_k] attention_probs = nn.Softmax(dim=-1)(attention_scores) # This is actually dropping out entire tokens to attend to, which might # seem a bit unusual, but is taken from the original Transformer paper. with get_cuda_rng_tracker().fork(): attention_probs = self.attention_dropout(attention_probs) # Context layer. context_layer = torch.matmul(attention_probs, value_layer) # context_layer: [b, n_p, s, h_i] context_layer = context_layer.permute(0, 2, 1, 3).contiguous() # context_layer: [b, s, n_p, h_i] # if self.do_dim_trick: # head_mask = self.head_mask.view(1, 1, self.head_mask.size(0), 1).expand_as(context_layer) # context_layer = context_layer * head_mask new_context_layer_shape = context_layer.size()[:-2] + (self.hidden_size_per_partition,) context_layer = context_layer.view(*new_context_layer_shape) # context_layer: [b, s, h_p] attn_output = self.dense(context_layer) # attn_output: [b, s, d_model] attn_output = self.output_dropout(attn_output) present_key_value_state = torch.stack((key_layer, value_layer), dim=0) if self.is_decoder else None outputs = (attn_output,) + (present_key_value_state,) + (position_bias,) if self.output_attention: outputs += (no_pos_bias_attn_probs,) else: outputs += (None,) return outputs # attn_output, present_key_value_state, position_bias, attention_probs class ParallelSelfAttention(nn.Module): def __init__( self, config: EncDecConfig, init_method: Callable, is_decoder: bool = False, output_layer_init_method: Optional[Callable] = None, has_relative_attention_bias: bool = False): super(ParallelSelfAttention, self).__init__() self.self_attn = ParallelAttention( config, init_method, is_decoder=is_decoder, is_cross_attn=False, output_layer_init_method=output_layer_init_method, has_relative_attention_bias=has_relative_attention_bias) self.layer_norm = LayerNorm(config.d_model, eps=config.layer_norm_epsilon) self.dropout = nn.Dropout(config.dropout_rate) def forward( self, hidden_states, attention_mask=None, position_bias=None, past_key_value=None): normed_hidden_states = self.layer_norm(hidden_states) attention_output = self.self_attn( normed_hidden_states, attention_mask=attention_mask, position_bias=position_bias, past_key_value=past_key_value, ) hidden_states = hidden_states + self.dropout(attention_output[0]) # add attentions if we output them outputs = (hidden_states,) + attention_output[1:] return outputs # hidden_states, present_key_value_state, position_bias, (attention_probs) class ParallelCrossAttention(nn.Module): def __init__( self, config: EncDecConfig, init_method: Callable, is_decoder: bool = True, output_layer_init_method: Optional[Callable] = None): super(ParallelCrossAttention, self).__init__() self.cross_attn = ParallelAttention( config, init_method, is_decoder=is_decoder, is_cross_attn=True, output_layer_init_method=output_layer_init_method, has_relative_attention_bias=False) self.layer_norm = LayerNorm(config.d_model, eps=config.layer_norm_epsilon) self.dropout = nn.Dropout(config.dropout_rate) def forward( self, hidden_states, key_value_states, attention_mask=None, position_bias=None, query_length=None, past_key_value=None): normed_hidden_states = self.layer_norm(hidden_states) attention_output = self.cross_attn( normed_hidden_states, key_value_states=key_value_states, attention_mask=attention_mask, position_bias=position_bias, query_length=query_length, past_key_value=past_key_value ) hidden_states = hidden_states + self.dropout(attention_output[0]) # add attentions if we output them outputs = (hidden_states,) + attention_output[1:] return outputs # hidden_states, present_key_value_state, position_bias, (attention_probs) class ParallelFF(nn.Module): def __init__( self, config: EncDecConfig, init_method: Callable, output_layer_init_method: Callable = None): super(ParallelFF, self).__init__() self.dense_relu_dense = ParallelDenseReluDense(config, init_method, output_layer_init_method) self.layer_norm = LayerNorm(config.d_model, eps=config.layer_norm_epsilon) self.dropout = nn.Dropout(config.dropout_rate) def forward(self, hidden_states): # hidden_states [b, s, d_model] forwarded_states = self.layer_norm(hidden_states) forwarded_states = self.dense_relu_dense(forwarded_states) hidden_states = hidden_states + self.dropout(forwarded_states) return hidden_states class ParallelBlock(nn.Module): def __init__( self, config: EncDecConfig, init_method: Callable, output_layer_init_method: Optional[Callable] = None, has_relative_attention_bias: bool = False, is_decoder: bool = False): super(ParallelBlock, self).__init__() if output_layer_init_method is None: output_layer_init_method = init_method self.is_decoder = is_decoder self.self_attn = ParallelSelfAttention( config, init_method, is_decoder=is_decoder, output_layer_init_method=output_layer_init_method, has_relative_attention_bias=has_relative_attention_bias) if is_decoder: self.cross_attn = ParallelCrossAttention( config, init_method, is_decoder=is_decoder, output_layer_init_method=output_layer_init_method) self.ff = ParallelFF( config, init_method, output_layer_init_method=output_layer_init_method) def forward( self, hidden_states, attention_mask=None, position_bias=None, enc_hidden_states=None, cross_attention_mask=None, enc_dec_position_bias=None, past_key_value=None,): if past_key_value is not None: self_attn_past_key_value = past_key_value[0] cross_attn_past_key_value = past_key_value[1] else: self_attn_past_key_value, cross_attn_past_key_value = None, None self_attn_outputs = self.self_attn( hidden_states, attention_mask=attention_mask, position_bias=position_bias, past_key_value=self_attn_past_key_value, ) hidden_states, self_attn_present_key_value = self_attn_outputs[:2] position_bias = (self_attn_outputs[2],) attention_probs = (self_attn_outputs[3],) present_key_value = (self_attn_present_key_value,) # cross attn if self.is_decoder: if self_attn_present_key_value is not None: query_length = self_attn_present_key_value[0].shape[2] else: query_length = None cross_attn_outputs = self.cross_attn( hidden_states, key_value_states=enc_hidden_states, attention_mask=cross_attention_mask, position_bias=enc_dec_position_bias, past_key_value=cross_attn_past_key_value, query_length=query_length, ) hidden_states, cross_attn_present_key_value = cross_attn_outputs[:2] present_key_value += (cross_attn_present_key_value,) # Keep cross-attention outputs and relative position weights position_bias = position_bias + (cross_attn_outputs[2],) attention_probs = attention_probs + (cross_attn_outputs[3],) hidden_states = self.ff(hidden_states) outputs = (hidden_states,) outputs = outputs + (present_key_value,) + position_bias + attention_probs # (for encoder) hidden_states, present_key_value_states, self-attention position bias, attention_probs # (for decoder) hidden_states, present_key_value_states, self-attention position bias, cross-attention position bias, self_attention_probs, cross_attention_probs return outputs class ParallelTransformer(nn.Module): def __init__(self, config: EncDecConfig, word_embeds: VocabParallelEmbedding, prompt_config=None, is_decoder=False, checkpoint_activations=False, checkpoint_num_layers=1, args=None): super(ParallelTransformer, self).__init__() self.word_embeds = word_embeds self.config = config self.args = args self.prompt_config = prompt_config if self.prompt_config is not None and self.prompt_config["prompt_len"] > 0: prompt_dim = prompt_config.get("prompt_dim", config.d_model) self.prompt_embeds = nn.Embedding(prompt_config["prompt_len"], prompt_dim) self.dropout = nn.Dropout(config.dropout_rate) self.final_layernorm = LayerNorm(config.d_model, eps=config.layer_norm_epsilon) self.checkpoint_activations = checkpoint_activations self.checkpoint_num_layers = checkpoint_num_layers self.is_decoder = is_decoder output_layer_init_method = None if config.use_scaled_init_for_output_weights: output_layer_init_method = scaled_init_method(config.init_method_std, config.num_layers) self.blocks = nn.ModuleList( [ParallelBlock( config, unscaled_init_method(sigma=config.init_method_std), has_relative_attention_bias=bool(i == 0), output_layer_init_method=output_layer_init_method, is_decoder=is_decoder) for i in range(config.num_layers)] ) if deepspeed.checkpointing.is_configured(): global get_cuda_rng_tracker, checkpoint get_cuda_rng_tracker = deepspeed.checkpointing.get_cuda_rng_tracker checkpoint = deepspeed.checkpointing.checkpoint def init_prompt_embeds(self): if self.prompt_config is not None and self.prompt_config["prompt_len"] > 0: prompt_weights = self.word_embeds(self.prompt_config["init_ids"]).detach() self.prompt_embeds.weight.data = prompt_weights def load_prompt_embeds(self, prompt_embeds): if self.prompt_config is not None and self.prompt_config["prompt_len"] > 0: prompt_embeds = prompt_embeds.to(self.prompt_embeds.weight.data.device) print("loading prompts") self.prompt_embeds.weight.data = prompt_embeds def get_prompt(self): if self.prompt_config is not None and self.prompt_config["prompt_len"] > 0: return self.prompt_embeds.weight.data else: return None def get_input_embeds(self, input_ids): if self.prompt_config is None: return self.word_embeds(input_ids) p_embeds = None if self.prompt_config is not None and self.prompt_config["prompt_len"] > 0 and self.prompt_config.get("insert_input", True): prompt_mask = (input_ids < 0).long() prompt_ids = (-(input_ids * prompt_mask)) - prompt_mask p_embeds = self.prompt_embeds(prompt_ids) * prompt_mask.half().unsqueeze(-1) word_mask = (0 <= input_ids).long() word_ids = word_mask * input_ids w_embeds = self.word_embeds(word_ids) * word_mask.float().unsqueeze(-1) if p_embeds is not None: w_embeds = w_embeds + p_embeds return w_embeds # bs * seq_len * hidden def forward( self, input_ids=None, attention_mask=None, cross_attention_mask=None, enc_hidden_states=None, past_key_values=None,): bs = input_ids.size(0) inputs_embeds = self.get_input_embeds(input_ids) hidden_states = self.dropout(inputs_embeds) position_bias = None enc_dec_position_bias = None present_key_value_states = [] # initialize past_key_values with `None` if past does not exist if past_key_values is None: past_key_values = [None] * len(self.blocks) all_self_attention_probs = [] all_cross_attention_probs = [] def custom(start, end): def custom_forward(*inputs): layer_modules_ = self.blocks[start:end] past_key_values_ = past_key_values[start:end] self_attn_present_key_values_ = [] cross_attn_present_key_values_ = [] position_bias_, enc_dec_position_bias_ = None, None hidden_states_ = inputs[0] if len(inputs) > 2: position_bias_ = inputs[1] if len(inputs) > 3: enc_dec_position_bias_ = inputs[2] if enc_hidden_states is not None: enc_hidden_states_ = inputs[-1] else: enc_hidden_states_ = None _l = start for layer_, past_key_value_ in zip(layer_modules_, past_key_values_): attention_mask_ = attention_mask cross_attention_mask_ = cross_attention_mask layer_outputs_ = layer_(hidden_states_, attention_mask_, position_bias_, enc_hidden_states_, cross_attention_mask_, enc_dec_position_bias_, past_key_value=past_key_value_) hidden_states_, present_key_value_ = layer_outputs_[:2] if self.is_decoder: self_attn_present_key_values_.append(present_key_value_[0]) cross_attn_present_key_values_.append(present_key_value_[1]) all_self_attention_probs.append(layer_outputs_[-2]) all_cross_attention_probs.append(layer_outputs_[-1]) else: self_attn_present_key_values_.append(present_key_value_[0]) all_self_attention_probs.append(layer_outputs_[-1]) position_bias_ = layer_outputs_[2] if self.is_decoder and enc_hidden_states is not None: enc_dec_position_bias_ = layer_outputs_[3] _l += 1 outputs_ = (hidden_states_,) if position_bias_ is not None: outputs_ += (position_bias_,) if enc_dec_position_bias_ is not None: outputs_ += (enc_dec_position_bias_,) if self.is_decoder: self_attn_present_key_values_ = torch.stack(self_attn_present_key_values_, dim=0) cross_attn_present_key_values_ = torch.stack(cross_attn_present_key_values_, dim=0) outputs_ += (self_attn_present_key_values_, cross_attn_present_key_values_,) return outputs_ return custom_forward if self.checkpoint_activations: l = 0 num_layers = len(self.blocks) chunk_length = self.checkpoint_num_layers while l < num_layers: arg_list = (hidden_states,) if position_bias is not None: arg_list += (position_bias,) if enc_dec_position_bias is not None: arg_list += (enc_dec_position_bias,) if enc_hidden_states is not None: arg_list += (enc_hidden_states,) tmp_outputs = checkpoint(custom(l, l+chunk_length), *arg_list) else: arg_list += (attention_mask,) tmp_outputs = checkpoint(custom(l, l+chunk_length), *arg_list) hidden_states = tmp_outputs[0] if self.is_decoder: if len(tmp_outputs) > 3: position_bias = tmp_outputs[1] if len(tmp_outputs) > 4: enc_dec_position_bias = tmp_outputs[2] present_key_value_states.extend([(s, c) for s, c in zip(tmp_outputs[-2], tmp_outputs[-1])]) else: if len(tmp_outputs) > 1: position_bias = tmp_outputs[1] if len(tmp_outputs) > 2: enc_dec_position_bias = tmp_outputs[2] present_key_value_states.extend([None] * chunk_length) l += chunk_length else: for i, (layer_module, past_key_value) in enumerate(zip(self.blocks, past_key_values)): layer_outputs = layer_module( hidden_states, attention_mask=attention_mask, position_bias=position_bias, enc_hidden_states=enc_hidden_states, cross_attention_mask=cross_attention_mask, enc_dec_position_bias=enc_dec_position_bias, past_key_value=past_key_value ) # layer_outputs is a tuple with: # hidden-states, key-value-states, self-attention position bias, cross-attention position bias, attention_probs hidden_states, present_key_value_state = layer_outputs[:2] if self.is_decoder: all_self_attention_probs.append(layer_outputs[-2]) all_cross_attention_probs.append(layer_outputs[-1]) else: all_self_attention_probs.append(layer_outputs[-1]) position_bias = layer_outputs[2] if self.is_decoder and enc_hidden_states is not None: enc_dec_position_bias = layer_outputs[3] present_key_value_states.append(present_key_value_state) # We share the position biases between the layers - the first layer store them # layer_outputs = hidden-states, key-value-states (self-attention weights), # (self-attention position bias), (cross-attention weights), (cross-attention position bias) # position_bias = layer_outputs[2] hidden_states = self.final_layernorm(hidden_states) hidden_states = self.dropout(hidden_states) # exit(0) outputs = { "last_hidden_state": hidden_states, "past_key_values": present_key_value_states, "hidden_states": None, "attentions": all_self_attention_probs, "cross_attentions": all_cross_attention_probs } return outputs

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/LM/Flan-T5/model/enc_dec_modeling.py

import copy import torch import torch.nn as nn import torch.nn.functional as F import mpu from .configuration_enc_dec import EncDecConfig def init_method_normal(std): """Init method based on normal distribution. This is only used for embeddings. The transformer has its own initializer. """ def init_(tensor): return torch.nn.init.normal_(tensor, mean=0.0, std=std) return init_ class EncDecModel(nn.Module): def __init__( self, config: EncDecConfig, parallel_output=True, checkpoint_activations=False, checkpoint_num_layers=1, prompt_config=None, args=None): super(EncDecModel, self).__init__() if config.vocab_size is None: raise RuntimeError("Should set vocab size") self.enc_config = copy.deepcopy(config) self.dec_config = copy.deepcopy(config) self.parallel_output = parallel_output init_method = init_method_normal(std=config.init_method_std) # NOTE: good? self.word_embeds = mpu.VocabParallelEmbedding(config.vocab_size, config.d_model, init_method=init_method) self.prompt_config = prompt_config self.args = args self.lm_head = mpu.VocabParallelEmbedding(config.vocab_size, config.d_model, init_method=init_method) self.encoder = mpu.ParallelTransformer(self.enc_config, word_embeds=self.word_embeds, is_decoder=False, prompt_config=prompt_config["enc"] if prompt_config is not None else None, checkpoint_activations=checkpoint_activations, checkpoint_num_layers=checkpoint_num_layers, args=args) self.decoder = mpu.ParallelTransformer(self.dec_config, word_embeds=self.word_embeds, is_decoder=True, prompt_config=prompt_config["dec"] if prompt_config is not None else None, checkpoint_activations=checkpoint_activations, checkpoint_num_layers=checkpoint_num_layers, args=args) if config.tie_weights: self.tie_weights() def init_prompt_embeds(self): self.encoder.init_prompt_embeds() self.decoder.init_prompt_embeds() def load_prompt_embeds(self, prompt_embeds): self.encoder.load_prompt_embeds(prompt_embeds) self.decoder.load_prompt_embeds(prompt_embeds) def get_prompt_embeds(self): return { "encoder": self.encoder.get_prompt(), "decoder": self.decoder.get_prompt() } def tie_weights(self): self.lm_head.weight = self.word_embeds.weight def reset_score_storage(self): for mod in self.decoder.blocks: mod.cross_attn.cross_attn.score_storage = None def get_crossattention_scores(self, context_mask): scores = [] n_passages = context_mask.size(1) for mod in self.decoder.blocks: scores.append(mod.cross_attn.cross_attn.score_storage) scores = torch.cat(scores, dim=2) # FiD n_layers beacuse dec seq = 1, auto regressive bsz, n_heads, n_layers, _ = scores.size() # batch_size, n_head, n_layers, n_passages, text_maxlength scores = scores.view(bsz, n_heads, n_layers, n_passages, -1) # batch_size, 1, 1, n_passages, text_maxlength scores = scores.masked_fill(~context_mask[:, None, None], 0.).float() # batch_size, n_passages scores = scores.sum(dim=[1, 2, 4]) ntokens = context_mask.sum(dim=[2]) * n_layers * n_heads scores = scores / ntokens return scores def forward( self, enc_input_ids=None, enc_position_ids=None, enc_attention_mask=None, dec_input_ids=None, dec_position_ids=None, dec_attention_mask=None, cross_attention_mask=None, enc_hidden_states=None, past_key_values=None, only_encoder=False,): provided_hidden = (enc_hidden_states is not None) if enc_hidden_states is None: enc_outputs = self.encoder( input_ids=enc_input_ids, attention_mask=enc_attention_mask, ) enc_hidden_states = enc_outputs["last_hidden_state"] if only_encoder: outputs = { "encoder_last_hidden_state": enc_hidden_states, "encoder_hidden_states": enc_outputs["hidden_states"], "encoder_attentions": enc_outputs["attentions"], } return outputs dec_outputs = self.decoder( input_ids=dec_input_ids, attention_mask=dec_attention_mask, cross_attention_mask=cross_attention_mask, enc_hidden_states=enc_hidden_states, past_key_values=past_key_values, ) last_hidden_state_parallel = mpu.copy_to_model_parallel_region(dec_outputs["last_hidden_state"]) logits_parallel = F.linear(last_hidden_state_parallel, self.lm_head.weight) if self.parallel_output: lm_logits = logits_parallel else: lm_logits = mpu.gather_from_model_parallel_region(logits_parallel) outputs = { "lm_logits": lm_logits, "last_hidden_state": dec_outputs["last_hidden_state"], "past_key_values": dec_outputs["past_key_values"], "encoder_last_hidden_state": enc_hidden_states, "encoder_attentions": enc_outputs["attentions"] if not provided_hidden else None, "decoder_self_attentions": dec_outputs["attentions"], "decoder_cross_attentions": dec_outputs["cross_attentions"] } return outputs def enc_dec_get_params_for_weight_decay_optimization(module): weight_decay_params = {'params': []} no_weight_decay_params = {'params': [], 'weight_decay': 0.0} for module_ in module.modules(): if isinstance(module_, (mpu.LayerNorm, nn.LayerNorm, mpu.transformer_enc_dec.LayerNorm)): no_weight_decay_params['params'].extend( [p for p in list(module_._parameters.values()) if p is not None]) else: weight_decay_params['params'].extend( [p for n, p in list(module_._parameters.items()) if p is not None and n != 'bias']) no_weight_decay_params['params'].extend( [p for n, p in list(module_._parameters.items()) if p is not None and n == 'bias']) return weight_decay_params, no_weight_decay_params def enc_dec_get_params_for_prompt_optimization(module: nn.Module): params = [{'params': []}] for t in module.named_modules(): if "prompt" in t[0]: if torch.distributed.get_rank() == 0: print("Update params", t[0]) params[0]['params'].extend([p for p in list(t[1]._parameters.values()) if p is not None]) for t in module.named_parameters(): if "prompt" not in t[0]: t[1].requires_grad_(False) if torch.distributed.get_rank() == 0: print("print params", params) return params

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/LM/Flan-T5/model/distributed.py

# coding=utf-8 import torch from torch._utils import _flatten_dense_tensors, _unflatten_dense_tensors import torch.distributed as dist from torch.nn.modules import Module from torch.autograd import Variable import mpu class DistributedDataParallel(Module): def __init__(self, module): super(DistributedDataParallel, self).__init__() self.warn_on_half = True if dist._backend == dist.dist_backend.GLOO else False self.module = module self.data_parallel_group = mpu.get_data_parallel_group() src_rank = mpu.get_model_parallel_rank() for p in self.module.parameters(): if torch.is_tensor(p): dist.broadcast(p, src_rank, group=self.data_parallel_group) def allreduce_params(reduce_after=True, no_scale=False, fp32_allreduce=False): if(self.needs_reduction): self.needs_reduction = False buckets = {} for name, param in self.module.named_parameters(): if param.requires_grad and param.grad is not None: tp = (param.data.type()) if tp not in buckets: buckets[tp] = [] buckets[tp].append(param) if self.warn_on_half: if torch.cuda.HalfTensor in buckets: print("WARNING: gloo dist backend for half parameters may be extremely slow." + " It is recommended to use the NCCL backend in this case.") self.warn_on_half = False for tp in buckets: bucket = buckets[tp] grads = [param.grad.data for param in bucket] coalesced = _flatten_dense_tensors(grads) if fp32_allreduce: coalesced = coalesced.float() if not no_scale and not reduce_after: coalesced /= dist.get_world_size(group=self.data_parallel_group) dist.all_reduce(coalesced, group=self.data_parallel_group) torch.cuda.synchronize() if not no_scale and reduce_after: coalesced /= dist.get_world_size(group=self.data_parallel_group) for buf, synced in zip(grads, _unflatten_dense_tensors(coalesced, grads)): buf.copy_(synced) self.hook_handles = [] self.hooks = [] for param in list(self.module.parameters()): def allreduce_hook(*unused): Variable._execution_engine.queue_callback(allreduce_params) # handle = param.register_hook(allreduce_hook) #self.hooks.append(allreduce_hook) #self.hook_handles.append(handle) self.allreduce_params = allreduce_params def forward(self, *inputs, **kwargs): self.needs_reduction = True return self.module(*inputs, **kwargs) def state_dict(self, destination=None, prefix='', keep_vars=False): #[h.remove() for h in self.hook_handles] sd = self.module.state_dict(destination, prefix, keep_vars) # for handle, hook in zip(self.hook_handles, self.hooks): # d = handle.hooks_dict_ref() # d[handle.id] = hook return sd def load_state_dict(self, state_dict, strict=True): self.module.load_state_dict(state_dict, strict=strict) ''' def _sync_buffers(self): buffers = list(self.module._all_buffers()) if len(buffers) > 0: # cross-node buffer sync flat_buffers = _flatten_dense_tensors(buffers) dist.broadcast(flat_buffers, 0) for buf, synced in zip(buffers, _unflatten_dense_tensors(flat_buffers, buffers)): buf.copy_(synced) def train(self, mode=True): # Clear NCCL communicator and CUDA event cache of the default group ID, # These cache will be recreated at the later call. This is currently a # work-around for a potential NCCL deadlock. if dist._backend == dist.dist_backend.NCCL: dist._clear_group_cache() super(DistributedDataParallel, self).train(mode) self.module.train(mode) '''

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/LM/Flan-T5/data_utils/T0Datasets.py

import json import re import os import torch import math import numpy as np import pickle from torch.utils.data import Dataset from utils import print_rank_0, save_rank_0 from tokenization_t5 import EncDecTokenizer from .data_config import DATA_GROUP_CONFIG, DATA_CONFIG import datasets from promptsource.templates import TemplateCollection from datasets import load_dataset from .postprocess import OPTION_POST_FN datasets.disable_caching() class T0Dataset(Dataset): def __init__( self, args, tokenizer: EncDecTokenizer, data_prompts, split, ratio=1, few_data_names=None, num=-1, ): self.args = args self.tokenizer = tokenizer self.ratio = ratio self.data_prompts = data_prompts self.pad_id = tokenizer.pad_id self.split = split self.sample_num = num self.idx_size = 3 self.few_data_names = few_data_names self.selfsup_sample_num = {"train": 100000, "validation": 1000} self.all_data = {name: {} for name in data_prompts} self.all_enc_sizes = [] self.all_dec_sizes = [] self.all_cand_sizes = [] if self.args.FiD: self.all_passage_sizes = [] for name in self.data_prompts: if DATA_CONFIG[name].get("selfsup", False): data, enc_sizes, dec_sizes, cand_sizes = self.load_from_cache_self(name) self.all_data[name] = { "prompt_num": 1, "prompt_names": ["merged"], "data": data, } else: if DATA_CONFIG[name]["do_cache"]: ( data, enc_sizes, dec_sizes, cand_sizes, passage_sizes, ) = self.load_from_cache(name) else: ( data, enc_sizes, dec_sizes, cand_sizes, passage_sizes, ) = self.process_data(name) self.all_data[name] = { "prompt_num": len(data_prompts[name]), "prompt_names": [prompt.name for prompt in data_prompts[name]], "data": data, } print("len data", len(data)) self.all_enc_sizes.extend(enc_sizes) self.all_dec_sizes.extend(dec_sizes) self.all_cand_sizes.extend(cand_sizes) if self.args.FiD: self.all_passage_sizes.extend(passage_sizes) self.max_enc_len = max(self.all_enc_sizes) self.max_dec_len = max(self.all_dec_sizes) self.max_cand_len = max(self.all_cand_sizes) if self.args.FiD: self.max_passage_len = max(self.all_passage_sizes) self.max_enc_len = self.max_passage_len * self.args.passage_num self.flan_sample_num = { name: min( DATA_CONFIG[name].get("flan_sample_max", args.flan_sample_max) * d["prompt_num"], len(d["data"]), ) for name, d in self.all_data.items() } self.idxs = self.build_idx() self.cur_epoch = 0 print_str = "" for name in self.data_prompts: print_str += "Data: {}_{}".format(name, split) print_str += " | Ratio: {}".format(ratio) print_str += " | Max enc len: {}".format(self.max_enc_len) print_str += " | Max dec len: {}".format(self.max_dec_len) print_str += " | Max cand len: {}".format(self.max_cand_len) print_str += " | Prompt num: {}".format(self.all_data[name]["prompt_num"]) print_str += " | All data num: {}".format(len(self.all_data[name]["data"])) print_str += " | Sample num: {}".format(self.flan_sample_num[name]) print_str += " | Idx one epoch num: {}\n".format(len(self.idxs[0])) print_str = print_str[:-1] print_rank_0(print_str) save_rank_0(args, print_str) def set_epoch(self, e): self.cur_epoch = e def build_idx(self): epochs = self.args.epochs idx_repo = {} for (name, d), (name, sample_num) in zip( self.all_data.items(), self.flan_sample_num.items() ): data_idx = [i for i in range(len(d["data"]))] repeat_num = math.ceil(epochs * sample_num / len(data_idx)) tmp_data_idx = [] for i in range(repeat_num): if self.split == "train": np.random.shuffle(data_idx) tmp_data_idx.extend(data_idx) idx_repo[name] = tmp_data_idx print( name, "| repeat num:", repeat_num, "| sample num:", sample_num, "| data_idx len:", len(data_idx), "| tmp_data_idx:", len(tmp_data_idx), ) idxs = [] for e in range(epochs): samp_idx = [] for name, d in self.all_data.items(): sample_num = self.flan_sample_num[name] l = idx_repo[name][e * sample_num : (e + 1) * sample_num] l = [(name, x) for x in l] samp_idx.extend(l) idxs.append(samp_idx) first_len = len(idxs[0]) for e, x in enumerate(idxs): assert len(x) == first_len, (e, len(x), first_len) return idxs def load_from_cache_self(self, name): cache_path = os.path.join( DATA_CONFIG[name]["data_dir"], "cache_{}_{}.pkl".format(self.split, self.selfsup_sample_num[self.split]), ) with open(cache_path, "rb") as f: data, enc_sizes, dec_sizes, cand_sizes = pickle.load(f) return data, enc_sizes, dec_sizes, cand_sizes def load_from_cache(self, name): data_dir = DATA_CONFIG[name]["data_dir"] if self.split == "train": if self.args.few_data_num is not None: assert self.few_data_names is not None if name in self.few_data_names: sample_num = self.args.few_data_num else: sample_num = self.sample_num else: sample_num = self.sample_num cache_path = os.path.join( data_dir, "cache_{}_{}_{}.pkl".format(self.split, self.ratio, sample_num), ) else: prompt_name = self.data_prompts[name][0].name.replace("/", "_") cache_path = os.path.join( data_dir, "cache_{}_{}_{}_{}.pkl".format( self.split, self.ratio, self.sample_num, prompt_name ), ) print("cache path", cache_path) if os.path.exists(cache_path): with open(cache_path, "rb") as f: data, enc_sizes, dec_sizes, cand_sizes, passage_sizes = pickle.load(f) else: data, enc_sizes, dec_sizes, cand_sizes, passage_sizes = self.process_data( name ) with open(cache_path, "wb") as f: pickle.dump((data, enc_sizes, dec_sizes, cand_sizes, passage_sizes), f) return data, enc_sizes, dec_sizes, cand_sizes, passage_sizes def process_data(self, name): print_rank_0("Processing " + name) if self.split == "train": if self.args.few_data_num is not None: assert self.few_data_names is not None if name in self.few_data_names: sample_num = self.args.few_data_num else: sample_num = self.sample_num else: sample_num = DATA_CONFIG[name].get("train_num", self.sample_num) if self.args.data_aug is not None: sample_num += self.args.data_aug else: sample_num = DATA_CONFIG[name].get("dev_num", self.sample_num) data_dir = DATA_CONFIG[name]["data_dir"] data_files = {self.split: os.path.join(data_dir, "{}.jsonl".format(self.split))} dataset = load_dataset("json", data_files=data_files) data = [] enc_sizes = [] dec_sizes = [] cand_sizes = [] passage_sizes = [] sid, lid = 0, 0 skip = 0 for pid, prompt in enumerate(self.data_prompts[name]): print_rank_0(prompt.name) for sample in dataset[self.split]: if lid % 500 == 0: print_rank_0( "{}, {}, {}, {}, {}".format( name, self.split, prompt.name, lid, skip ) ) # genread_template = "{} Generate a background document from Wikipedia to help answer the given question:" answers = None if "popQA" in name: enc_str = sample["prompt"] # enc_str = genread_template.format(enc_str) dec_str = sample["answers"][0] answers = sample["answers"] else: applied_sample = prompt.apply(sample) if len(applied_sample) != 2: # print_rank_0("sample num out") skip += 1 continue enc_str, dec_str = applied_sample # enc_str = genread_template.format(enc_str) if "mmlu_demo" in sample: enc_str = sample["mmlu_demo"] + enc_str passages = None if "passages" in sample: passages = [] for i in range(self.args.passage_num): max_question_len = 1250 if self.split == "train" else 10000 max_passage_len = ( max(1250 - len(enc_str), 0) if self.split == "train" else 500 ) # Can last if self.args.prompt_tune: passage_str = enc_str[:max_question_len] passages.append( [-(i + 1)] + self.tokenizer.encode(passage_str) + [1] ) else: passage_str = ( sample["passages"][i][:max_passage_len] + enc_str[:max_question_len] ) passages.append(self.tokenizer.encode(passage_str) + [1]) if self.args.prompt_tune: context = ( [-(i + 1) for i in range(self.args.passage_num)] + self.tokenizer.encode(enc_str) + [1] ) else: context = self.tokenizer.encode(enc_str) + [1] target = [0] + self.tokenizer.encode(dec_str) + [1] # if len(enc_str) > 5000: # # print_rank_0("pre-check out " + str(len(enc_str))) # skip += 1 # continue # if len(context) > self.args.enc_seq_length: # skip += 1 # # print_rank_0("enc out " + str(len(context))) # continue # if len(target) > self.args.dec_seq_length: # skip += 1 # # print_rank_0("dec out " + str(len(target))) # continue options = prompt.get_answer_choices_list(sample) options = OPTION_POST_FN.get((name, prompt.name), lambda x: x)(options) if self.split != "train" and options is not None: cands = [ [0] + self.tokenizer.encode(option) + [1] for option in options ] else: cands = None if len(dec_str) == 0: # print_rank_0("dec str out " + str(len(dec_str))) skip += 1 continue if options is not None and dec_str not in options: print_rank_0(str(applied_sample)) print_rank_0( name + " " + prompt.name + " " + "Skip bug sample " + str(dec_str) + " " + str(options) ) continue data.append( { "idxs": [pid, lid, sid], "enc_input_ids": context, "dec_input_ids": target[:-1], "label_ids": target[1:], "answer": dec_str if answers is None else answers, "options": options, "cands": { "input_ids": [c[:-1] for c in cands], "target_ids": [c[1:] for c in cands], "label": options.index(dec_str), } if cands is not None else None, "passage_input_ids": passages, } ) enc_sizes.append(len(context)) dec_sizes.append(len(target) - 1) if cands is not None: cand_sizes.append(sum([len(c) - 1 for c in cands])) else: cand_sizes.append(0) if passages is not None: passage_sizes.extend([len(p) for p in passages]) else: passage_sizes.append(0) sid += 1 lid += 1 if sample_num > 0 and lid >= sample_num: break lid = 0 return data, enc_sizes, dec_sizes, cand_sizes, passage_sizes def __len__(self): return len(self.idxs[0]) def __getitem__(self, idx): name, sid = self.idxs[self.cur_epoch][idx] d = self.all_data[name]["data"][sid] return d, name def collate(self, samples): bs = len(samples) model_data = { "enc_input_ids": torch.ones(bs, self.max_enc_len, dtype=torch.long) * self.pad_id, "enc_attention_mask": torch.zeros( bs, 1, self.max_enc_len, self.max_enc_len ), "dec_attention_mask": torch.zeros( bs, 1, self.max_dec_len, self.max_dec_len ), "cross_attention_mask": torch.zeros( bs, 1, self.max_dec_len, self.max_enc_len ), "dec_input_ids": torch.ones(bs, self.max_dec_len, dtype=torch.long) * self.pad_id, } if self.args.FiD: model_data["passage_input_ids"] = ( torch.ones( bs, self.args.passage_num, self.max_passage_len, dtype=torch.long ) * self.pad_id ) model_data["passage_attention_mask"] = torch.zeros( bs, self.args.passage_num, 1, self.max_passage_len, self.max_passage_len ) no_model_data = { "idxs": torch.zeros(bs, self.idx_size, dtype=torch.long), "labels": torch.ones(bs, self.max_dec_len, dtype=torch.long) * self.pad_id, "loss_mask": torch.zeros(bs, self.max_dec_len), } name_list = [] for i, samp in enumerate(samples): samp, name = samp name_list.append(name) enc_len, dec_len = len(samp["enc_input_ids"]), len(samp["dec_input_ids"]) model_data["enc_input_ids"][i][:enc_len] = torch.tensor( samp["enc_input_ids"], dtype=torch.long ) model_data["enc_attention_mask"][i][0, :enc_len, :enc_len] = samp.get( "enc_attention_mask", 1.0 ) model_data["dec_input_ids"][i][:dec_len] = torch.tensor( samp["dec_input_ids"], dtype=torch.long ) model_data["dec_attention_mask"][i][0, :dec_len, :dec_len] = torch.tril( torch.ones(dec_len, dec_len) ) if self.args.FiD: enc_len = self.max_enc_len samp["cross_attention_mask"] = torch.zeros(enc_len) for j in range(self.args.passage_num): passage_len = len(samp["passage_input_ids"][j]) samp["cross_attention_mask"][ j * self.max_passage_len : j * self.max_passage_len + passage_len ] = 1.0 model_data["cross_attention_mask"][i][0, :dec_len, :enc_len] = samp.get( "cross_attention_mask", 1.0 ) if self.args.FiD: for j in range(self.args.passage_num): passage_len = len(samp["passage_input_ids"][j]) model_data["passage_input_ids"][i][j][:passage_len] = torch.tensor( samp["passage_input_ids"][j], dtype=torch.long ) model_data["passage_attention_mask"][i][j][ 0, :passage_len, :passage_len ] = 1.0 no_model_data["idxs"][i] = torch.tensor(samp["idxs"], dtype=torch.long) no_model_data["labels"][i][: len(samp["label_ids"])] = torch.tensor( samp["label_ids"], dtype=torch.long ) no_model_data["loss_mask"][i][: len(samp["label_ids"])] = 1.0 if self.args.fp16: model_data["enc_attention_mask"] = model_data["enc_attention_mask"].half() model_data["dec_attention_mask"] = model_data["dec_attention_mask"].half() model_data["cross_attention_mask"] = model_data[ "cross_attention_mask" ].half() if self.args.FiD: model_data["passage_attention_mask"] = model_data[ "passage_attention_mask" ].half() if samp["cands"] is not None: cand_model_data = { "dec_input_ids": torch.ones(bs, self.max_cand_len, dtype=torch.long) * self.pad_id, "dec_attention_mask": torch.zeros( bs, 1, self.max_cand_len, self.max_cand_len ), "cross_attention_mask": torch.zeros( bs, 1, self.max_cand_len, self.max_enc_len ), } cand_no_model_data = { "labels": torch.zeros(bs, dtype=torch.long), "target_ids": torch.ones(bs, self.max_cand_len, dtype=torch.long) * self.pad_id, "pos": torch.zeros(bs, self.max_cand_len, dtype=torch.bool), "loss_mask": torch.zeros(bs, self.max_cand_len), } for i, samp in enumerate(samples): samp, _ = samp start = 0 enc_len = len(samp["enc_input_ids"]) if self.args.FiD: enc_len = self.max_enc_len for input_ids, target_ids in zip( samp["cands"]["input_ids"], samp["cands"]["target_ids"] ): cand_model_data["dec_input_ids"][i][ start : start + len(input_ids) ] = torch.tensor(input_ids, dtype=torch.long) cand_no_model_data["target_ids"][i][ start : start + len(target_ids) ] = torch.tensor(target_ids, dtype=torch.long) cand_model_data["dec_attention_mask"][i][ 0, start : start + len(input_ids), start : start + len(input_ids), ] = torch.tril(torch.ones(len(input_ids), len(input_ids))) cand_model_data["cross_attention_mask"][i][ 0, start : start + len(input_ids), :enc_len ] = samp.get("cross_attention_mask", 1.0) cand_no_model_data["loss_mask"][i][ start : start + len(input_ids) ] = 1 start = start + len(input_ids) cand_no_model_data["pos"][i][start - 1] = True cand_no_model_data["labels"][i] = samp["cands"]["label"] if self.args.fp16: cand_model_data["dec_attention_mask"] = cand_model_data[ "dec_attention_mask" ].half() cand_model_data["cross_attention_mask"] = cand_model_data[ "cross_attention_mask" ].half() else: cand_model_data, cand_no_model_data = {}, {} # print(name_list) return model_data, no_model_data, cand_model_data, cand_no_model_data

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/Retriever/loss.py

import torch from torch import Tensor from torch.nn import functional as F from torch import distributed as dist class SimpleContrastiveLoss: def __call__(self, x: Tensor, y: Tensor, target: Tensor = None, reduction: str = 'mean'): if target is None: target_per_qry = y.size(0) // x.size(0) target = torch.arange( 0, x.size(0) * target_per_qry, target_per_qry, device=x.device, dtype=torch.long) logits = torch.matmul(x, y.transpose(0, 1)) return F.cross_entropy(logits, target, reduction=reduction) class DistributedContrastiveLoss(SimpleContrastiveLoss): def __init__(self, n_target: int = 0, scale_loss: bool = True): assert dist.is_initialized(), "Distributed training has not been properly initialized." super().__init__() self.word_size = dist.get_world_size() self.rank = dist.get_rank() self.scale_loss = scale_loss def __call__(self, x: Tensor, y: Tensor, **kwargs): dist_x = self.gather_tensor(x) dist_y = self.gather_tensor(y) loss = super().__call__(dist_x, dist_y, **kwargs) if self.scale_loss: loss = loss * self.word_size return loss def gather_tensor(self, t): gathered = [torch.empty_like(t) for _ in range(self.word_size)] dist.all_gather(gathered, t) gathered[self.rank] = t return torch.cat(gathered, dim=0) class MarginRankingLoss: def __init__(self, margin: float = 1.0): self.margin = margin def __call__(self, pos_scores: Tensor, neg_scores: Tensor): return torch.mean(F.relu(self.margin - pos_scores + neg_scores)) class SoftMarginRankingLoss: def __init__(self, margin: float = 1.0): self.margin = margin def __call__(self, pos_scores: Tensor, neg_scores: Tensor): return torch.mean(F.softplus(self.margin - pos_scores + neg_scores)) class BinaryCrossEntropyLoss: def __call__(self, pos_scores: Tensor, neg_scores: Tensor): return (F.binary_cross_entropy_with_logits(pos_scores, torch.ones_like(pos_scores)) + F.binary_cross_entropy_with_logits(neg_scores, torch.zeros_like(neg_scores))) class CrossEntropyLoss: def __call__(self, pos_scores: Tensor, neg_scores: Tensor): return (F.cross_entropy(pos_scores, torch.ones(pos_scores.shape[0], dtype=torch.long).to(pos_scores.device)) + F.cross_entropy(neg_scores, torch.zeros(neg_scores.shape[0], dtype=torch.long).to(pos_scores.device))) rr_loss_functions = { "mr": MarginRankingLoss, "smr": SoftMarginRankingLoss, "bce": BinaryCrossEntropyLoss, "ce": CrossEntropyLoss, }

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/Retriever/utils.py

# Adapted from Tevatron (https://github.com/texttron/tevatron) import csv import json import warnings from dataclasses import dataclass from typing import Dict, List import datasets import torch from transformers import PreTrainedTokenizer try: from opendelta import BitFitModel, AdapterModel, PrefixModel, LoraModel _opendelta_available = True except ModuleNotFoundError: _opendelta_available = False @dataclass class SimpleTrainPreProcessor: query_file: str collection_file: str tokenizer: PreTrainedTokenizer doc_max_len: int = 128 query_max_len: int = 32 columns = ['text_id', 'title', 'text'] title_field = 'title' text_field = 'text' query_field = 'text' doc_template: str = None query_template: str = None allow_not_found: bool = False def __post_init__(self): self.queries = self.read_queries(self.query_file) self.collection = datasets.load_dataset( 'csv', data_files=self.collection_file, column_names=self.columns, delimiter='\t', )['train'] @staticmethod def read_queries(queries): qmap = {} if queries[-3:] == "csv": with open(queries) as f: reader = csv.DictReader(f, fieldnames=["qid", "qry"], delimiter="\t") for row in reader: qid = row.pop("qid") qry = row.pop("qry") qmap[qid] = qry else: with open(queries) as f: for l in f: qid, qry = l.strip().split('\t') qmap[qid] = qry return qmap @staticmethod def read_qrel(relevance_file): qrel = {} with open(relevance_file, encoding='utf8') as f: tsvreader = csv.reader(f, delimiter="\t") for [topicid, _, docid, rel] in tsvreader: assert rel == "1" if topicid in qrel: qrel[topicid].append(docid) else: qrel[topicid] = [docid] return qrel def get_query(self, q): if self.query_template is None: query = self.queries[q] else: query = fill_template(self.query_template, data={self.query_field: self.queries[q]}, allow_not_found=self.allow_not_found) query_encoded = self.tokenizer.encode( query, add_special_tokens=False, max_length=self.query_max_len, truncation=True ) return query_encoded def get_passage(self, p): entry = self.collection[int(p)] title = entry[self.title_field] title = "" if title is None else title body = entry[self.text_field] if self.doc_template is None: content = title + self.tokenizer.sep_token + body else: content = fill_template(self.doc_template, data=entry, allow_not_found=self.allow_not_found) passage_encoded = self.tokenizer.encode( content, add_special_tokens=False, max_length=self.doc_max_len, truncation=True ) return passage_encoded def process_one(self, train): q, pp, nn = train train_example = { 'query': self.get_query(q), 'positives': [self.get_passage(p) for p in pp], 'negatives': [self.get_passage(n) for n in nn], } return json.dumps(train_example) @dataclass class SimpleCollectionPreProcessor: tokenizer: PreTrainedTokenizer separator: str = '\t' max_length: int = 128 def process_line(self, line: str): xx = line.strip().split(self.separator) text_id, text = xx[0], xx[1:] text_encoded = self.tokenizer.encode( self.tokenizer.sep_token.join(text), add_special_tokens=False, max_length=self.max_length, truncation=True ) encoded = { 'text_id': text_id, 'text': text_encoded } return json.dumps(encoded) def save_as_trec(rank_result: Dict[str, Dict[str, float]], output_path: str, run_id: str = "OpenMatch"): """ Save the rank result as TREC format: <query_id> Q0 <doc_id> <rank> <score> <run_id> """ with open(output_path, "w") as f: for qid in rank_result: # sort the results by score sorted_results = sorted(rank_result[qid].items(), key=lambda x: x[1], reverse=True) for i, (doc_id, score) in enumerate(sorted_results): f.write("{} Q0 {} {} {} {}\n".format(qid, doc_id, i + 1, score, run_id)) def load_from_trec(input_path: str, as_list: bool = False, max_len_per_q: int = None): """ Load the rank result from TREC format: <query_id> Q0 <doc_id> <rank> <score> <run_id> or <query_id> <doc_id> <score> """ rank_result = {} cnt = 0 with open(input_path, "r") as f: for line in f: content = line.strip().split() if len(content) == 6: qid, _, doc_id, _, score, _ = content elif len(content) == 3: qid, doc_id, score = content else: raise ValueError("Invalid run format") if not as_list: if qid not in rank_result: rank_result[qid] = {} cnt = 0 if max_len_per_q is None or cnt < max_len_per_q: rank_result[qid][doc_id] = float(score) else: if qid not in rank_result: rank_result[qid] = [] cnt = 0 if max_len_per_q is None or cnt < max_len_per_q: rank_result[qid].append((doc_id, float(score))) cnt += 1 return rank_result def find_all_markers(template: str): """ Find all markers' names (quoted in "<>") in a template. """ markers = [] start = 0 while True: start = template.find("<", start) if start == -1: break end = template.find(">", start) if end == -1: break markers.append(template[start + 1:end]) start = end + 1 return markers def fill_template(template: str, data: Dict, markers: List[str] = None, allow_not_found: bool = False): """ Fill a template with data. """ if markers is None: markers = find_all_markers(template) for marker in markers: marker_hierarchy = marker.split(".") found = True content = data for marker_level in marker_hierarchy: content = content.get(marker_level, None) if content is None: found = False break if not found: if allow_not_found: warnings.warn("Marker '{}' not found in data. Replacing it with an empty string.".format(marker), RuntimeWarning) content = "" else: raise ValueError("Cannot find the marker '{}' in the data".format(marker)) template = template.replace("<{}>".format(marker), str(content)) return template def merge_retrieval_results_by_score(results: List[Dict[str, Dict[str, float]]], topk: int = 100): """ Merge retrieval results from multiple partitions of document embeddings and keep topk. """ merged_results = {} for result in results: for qid in result: if qid not in merged_results: merged_results[qid] = {} for doc_id in result[qid]: if doc_id not in merged_results[qid]: merged_results[qid][doc_id] = result[qid][doc_id] for qid in merged_results: merged_results[qid] = {k: v for k, v in sorted(merged_results[qid].items(), key=lambda x: x[1], reverse=True)[:topk]} return merged_results # Mean Pooling - Take attention mask into account for correct averaging def mean_pooling(token_embeddings, attention_mask): input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float() return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9) def get_delta_model_class(model_type): if not _opendelta_available: raise ValueError( 'OpenDelta package not available. You can obtain it from https://github.com/thunlp/OpenDelta.') delta_models = { 'bitfit': BitFitModel, 'adapter': AdapterModel, 'prefix': PrefixModel, 'lora': LoraModel } return delta_models[model_type]

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/Retriever/trainer/dense_trainer.py

# Adapted from Tevatron (https://github.com/texttron/tevatron) import logging import os from itertools import repeat from typing import Any, Dict, List, Optional, Tuple, Union import datasets import torch import torch.distributed as dist from torch.utils.data import DataLoader from transformers.file_utils import is_datasets_available from transformers.trainer import Trainer, TRAINING_ARGS_NAME from transformers.trainer_pt_utils import IterableDatasetShard from ..loss import DistributedContrastiveLoss, SimpleContrastiveLoss logger = logging.getLogger(__name__) try: from grad_cache import GradCache _grad_cache_available = True except ModuleNotFoundError: _grad_cache_available = False class DRTrainer(Trainer): def __init__(self, delta_model=None, *args, **kwargs): super(DRTrainer, self).__init__(*args, **kwargs) self.delta_model = delta_model self._dist_loss_scale_factor = dist.get_world_size() if self.args.negatives_x_device else 1 def _save(self, output_dir: Optional[str] = None): output_dir = output_dir if output_dir is not None else self.args.output_dir os.makedirs(output_dir, exist_ok=True) logger.info("Saving model checkpoint to %s", output_dir) self.model.save(output_dir) if self.tokenizer is not None: self.tokenizer.save_pretrained(output_dir) # Good practice: save your training arguments together with the trained model torch.save(self.args, os.path.join(output_dir, TRAINING_ARGS_NAME)) if self.delta_model: logger.info("Saving delta model to %s", output_dir + "/delta_model") self.delta_model.save_finetuned(output_dir + "/delta_model") def _prepare_inputs( self, inputs: Tuple[Dict[str, Union[torch.Tensor, Any]], ...] ) -> List[Dict[str, Union[torch.Tensor, Any]]]: prepared = [] for x in inputs: if isinstance(x, torch.Tensor): prepared.append(x.to(self.args.device)) else: prepared.append(super()._prepare_inputs(x)) return prepared def get_train_dataloader(self) -> DataLoader: """ Returns the training [`~torch.utils.data.DataLoader`]. Will use no sampler if `self.train_dataset` does not implement `__len__`, a random sampler (adapted to distributed training if necessary) otherwise. Subclass and override this method if you want to inject some custom behavior. """ if self.train_dataset is None: raise ValueError("Trainer: training requires a train_dataset.") train_dataset = self.train_dataset if is_datasets_available() and isinstance(train_dataset, datasets.Dataset): train_dataset = self._remove_unused_columns(train_dataset, description="training") if isinstance(train_dataset, torch.utils.data.IterableDataset): if self.args.world_size > 1: train_dataset = IterableDatasetShard( train_dataset, batch_size=self.args.train_batch_size, drop_last=self.args.dataloader_drop_last, num_processes=self.args.world_size, process_index=self.args.process_index, ) return DataLoader( train_dataset, batch_size=self.args.per_device_train_batch_size, collate_fn=self.data_collator, drop_last=self.args.dataloader_drop_last, num_workers=self.args.dataloader_num_workers, pin_memory=self.args.dataloader_pin_memory, ) train_sampler = self._get_train_sampler() return DataLoader( train_dataset, batch_size=self.args.train_batch_size, sampler=train_sampler, collate_fn=self.data_collator, drop_last=self.args.dataloader_drop_last, num_workers=self.args.dataloader_num_workers, pin_memory=self.args.dataloader_pin_memory, ) def compute_loss(self, model, inputs, return_outputs=False): query, passage = inputs outputs = model(query=query, passage=passage) return (outputs.loss, outputs) if return_outputs else outputs.loss def training_step(self, *args): return super(DRTrainer, self).training_step(*args) / self._dist_loss_scale_factor def split_dense_inputs(model_input: dict, chunk_size: int): assert len(model_input) == 1 arg_key = list(model_input.keys())[0] arg_val = model_input[arg_key] keys = list(arg_val.keys()) chunked_tensors = [arg_val[k].split(chunk_size, dim=0) for k in keys] chunked_arg_val = [dict(zip(kk, tt)) for kk, tt in zip(repeat(keys), zip(*chunked_tensors))] return [{arg_key: c} for c in chunked_arg_val] def get_dense_rep(x): if x.q_reps is None: return x.p_reps else: return x.q_reps class GCDenseTrainer(DRTrainer): def __init__(self, *args, **kwargs): logger.info('Initializing Gradient Cache Trainer') if not _grad_cache_available: raise ValueError( 'Grad Cache package not available. You can obtain it from https://github.com/luyug/GradCache.') super(GCDenseTrainer, self).__init__(*args, **kwargs) loss_fn_cls = DistributedContrastiveLoss if self.args.negatives_x_device else SimpleContrastiveLoss loss_fn = loss_fn_cls() self.gc = GradCache( models=[self.model, self.model], chunk_sizes=[self.args.gc_q_chunk_size, self.args.gc_p_chunk_size], loss_fn=loss_fn, split_input_fn=split_dense_inputs, get_rep_fn=get_dense_rep, fp16=self.args.fp16, scaler=self.scaler ) def training_step(self, model, inputs) -> torch.Tensor: model.train() queries, passages = self._prepare_inputs(inputs) queries, passages = {'query': queries}, {'passage': passages} _distributed = self.args.local_rank > -1 self.gc.models = [model, model] loss = self.gc(queries, passages, no_sync_except_last=_distributed) return loss / self._dist_loss_scale_factor

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/Retriever/trainer/reranker_trainer.py

# Adapted from Tevatron (https://github.com/texttron/tevatron) import logging import os from typing import Any, Dict, List, Optional, Tuple, Union import torch import torch.nn as nn from transformers.trainer import Trainer from transformers.trainer_pt_utils import nested_detach logger = logging.getLogger(__name__) class RRTrainer(Trainer): def __init__(self, *args, **kwargs): super(RRTrainer, self).__init__(*args, **kwargs) def _save(self, output_dir: Optional[str] = None): output_dir = output_dir if output_dir is not None else self.args.output_dir os.makedirs(output_dir, exist_ok=True) logger.info("Saving model checkpoint to %s", output_dir) self.model.save(output_dir) def _prepare_inputs( self, inputs: Tuple[Dict[str, Union[torch.Tensor, Any]], ...] ) -> List[Dict[str, Union[torch.Tensor, Any]]]: prepared = [] for x in inputs: if isinstance(x, torch.Tensor): prepared.append(x.to(self.args.device)) else: prepared.append(super()._prepare_inputs(x)) return prepared def prediction_step( self, model: nn.Module, inputs, prediction_loss_only: bool, ignore_keys: Optional[List[str]] = None, ) -> Tuple[Optional[torch.Tensor], Optional[torch.Tensor], Optional[torch.Tensor]]: inputs = self._prepare_inputs(inputs) if ignore_keys is None: if hasattr(self.model, "config"): ignore_keys = getattr(self.model.config, "keys_to_ignore_at_inference", []) else: ignore_keys = [] with torch.no_grad(): with self.autocast_smart_context_manager(): loss, outputs = self.compute_loss(model, inputs, return_outputs=True) loss = loss.mean().detach() if isinstance(outputs, dict): logits = tuple(v for k, v in outputs.items() if k not in ignore_keys + ["loss"]) else: logits = outputs[1:] if prediction_loss_only: return (loss, None, None) logits = nested_detach(logits) if len(logits) == 1: logits = logits[0] return (loss, logits, None) def compute_loss(self, model, inputs, return_outputs=False): pos_pairs, neg_pairs = inputs outputs = model(pos_pairs=pos_pairs, neg_pairs=neg_pairs) return (outputs.loss, outputs) if return_outputs else outputs.loss

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/Retriever/dataset/inference_dataset.py

# Adapted from Tevatron (https://github.com/texttron/tevatron) import os from functools import lru_cache from typing import List, Union, Callable from datasets import load_dataset from torch.utils.data import Dataset, IterableDataset from transformers import PreTrainedTokenizer from ..arguments import DataArguments from ..utils import fill_template, find_all_markers def get_idx(obj): example_id = obj.get("_id", None) or obj.get("id", None) or obj.get("text_id", None) example_id = str(example_id) if example_id is not None else None return example_id class InferenceDataset(): def __init__( self, tokenizer: PreTrainedTokenizer, data_args: DataArguments, data_files: Union[str, List[str]], is_query: bool = False, full_tokenization: bool = True, mode: str = "processed", batch_size: int = 1, num_processes: int = 1, process_index: int = 0, filter_fn: Callable = lambda x: True, cache_dir: str = None ): self.cache_dir = cache_dir self.is_query = is_query self.data_files = data_files self.tokenizer = tokenizer self.max_len = data_args.q_max_len if self.is_query else data_args.p_max_len self.template = data_args.query_template if self.is_query else data_args.doc_template self.all_markers = find_all_markers(self.template) if data_args.all_markers is None else data_args.all_markers.split(",") self.full_tokenization = full_tokenization modes = ["raw", "dict_processed", "processed"] if mode not in modes: raise ValueError(f"mode must be one of {modes}") self.mode = mode self.batch_size = batch_size self.num_processes = num_processes self.process_index = process_index self.filter_fn = filter_fn self._prepare_data(data_args) def _prepare_data(self, data_args): raise NotImplementedError @classmethod def load( cls, tokenizer: PreTrainedTokenizer, data_args: DataArguments, data_files: Union[str, List[str]] = None, is_query: bool = False, full_tokenization: bool = True, mode: str = "processed", stream: bool = True, batch_size: int = 1, num_processes: int = 1, process_index: int = 0, filter_fn: Callable = lambda x: True, cache_dir: str = None ): if data_files is None: data_files = [data_args.query_path] if is_query else [data_args.corpus_path] else: data_files = [data_files] if isinstance(data_files, str) else data_files ext = os.path.splitext(data_files[0])[1] ext_to_cls = { ".json": StreamJsonlDataset if stream else MappingJsonlDataset, ".csv": StreamTsvDataset if stream else MappingTsvDataset, ".jsonl": StreamJsonlDataset if stream else MappingJsonlDataset, ".tsv": StreamTsvDataset if stream else MappingTsvDataset, ".txt": StreamTsvDataset if stream else MappingTsvDataset, } cls_ = ext_to_cls.get(ext, None) if cls_ is None: raise ValueError("Unsupported dataset file extension {}".format(ext)) return cls_( tokenizer=tokenizer, data_args=data_args, data_files=data_files, is_query=is_query, full_tokenization=full_tokenization, mode=mode, batch_size=batch_size, num_processes=num_processes, process_index=process_index, filter_fn=filter_fn, cache_dir=cache_dir ) def _tokenize(self, example: str): return self.tokenizer( example, add_special_tokens=self.full_tokenization, padding='max_length' if self.full_tokenization else False, truncation=True, max_length=self.max_len, return_attention_mask=self.full_tokenization, return_token_type_ids=False ) def process_one(self, example): if self.mode == "raw": return example elif self.mode == "dict_processed": example_id = get_idx(example) tokenized = {} for marker in self.all_markers: tokenized[marker] = dict(self._tokenize(example[marker])) if (marker in example and example[marker] is not None) else None return {"text_id": example_id, **tokenized} else: example_id = get_idx(example) full_text = fill_template(self.template, example, self.all_markers, allow_not_found=True) tokenized = self._tokenize(full_text) return {"text_id": example_id, **tokenized} class StreamInferenceDataset(InferenceDataset, IterableDataset): def __init__( self, tokenizer: PreTrainedTokenizer, data_args: DataArguments, data_files: Union[str, List[str]], **kwargs ): super(StreamInferenceDataset, self).__init__(tokenizer, data_args, data_files, **kwargs) def __iter__(self): real_batch_size = self.batch_size * self.num_processes process_slice = range(self.process_index * self.batch_size, (self.process_index + 1) * self.batch_size) current_batch = [] for element in self.dataset: current_batch.append(element) # Wait to have a full batch before yielding elements. if len(current_batch) == real_batch_size: for i in process_slice: yield self.process_one(current_batch[i]) current_batch = [] if len(current_batch) > 0: for i in process_slice: if i < len(current_batch): yield self.process_one(current_batch[i]) class StreamJsonlDataset(StreamInferenceDataset): def __init__( self, tokenizer: PreTrainedTokenizer, data_args: DataArguments, data_files: Union[str, List[str]], **kwargs ): super(StreamJsonlDataset, self).__init__(tokenizer, data_args, data_files, **kwargs) def _prepare_data(self, data_args): self.dataset = load_dataset( "json", data_files=self.data_files, streaming=True, cache_dir=self.cache_dir )["train"].filter(self.filter_fn) sample = list(self.dataset.take(1))[0] self.all_columns = sample.keys() class StreamTsvDataset(StreamInferenceDataset): def __init__( self, tokenizer: PreTrainedTokenizer, data_args: DataArguments, data_files: Union[str, List[str]], **kwargs ): super(StreamTsvDataset, self).__init__(tokenizer, data_args, data_files, **kwargs) def _prepare_data(self, data_args): self.all_columns = data_args.query_column_names if self.is_query else data_args.doc_column_names if self.all_columns is not None: self.all_columns = self.all_columns.split(',') self.dataset = load_dataset( "csv", data_files=self.data_files, streaming=True, column_names=self.all_columns, delimiter='\t', cache_dir=self.cache_dir )["train"].filter(self.filter_fn) class MappingInferenceDataset(InferenceDataset, Dataset): def __init__( self, tokenizer: PreTrainedTokenizer, data_args: DataArguments, data_files: Union[str, List[str]], **kwargs ): super(MappingInferenceDataset, self).__init__(tokenizer, data_args, data_files, **kwargs) @lru_cache(maxsize=None) def __getitem__(self, index): return self.process_one(self.dataset[index]) def get_raw(self, index): return self.dataset[index] def __len__(self): return len(self.dataset) class MappingJsonlDataset(MappingInferenceDataset): def __init__( self, tokenizer: PreTrainedTokenizer, data_args: DataArguments, data_files: Union[str, List[str]], **kwargs ): super(MappingJsonlDataset, self).__init__(tokenizer, data_args, data_files, **kwargs) def _prepare_data(self, data_args): hf_dataset = load_dataset( "json", data_files=self.data_files, streaming=True, cache_dir=self.cache_dir )["train"].filter(self.filter_fn) sample = list(self.dataset.take(1))[0] self.all_columns = sample.keys() self.dataset = {} for item in hf_dataset: self.dataset[get_idx(item)] = item class MappingTsvDataset(MappingInferenceDataset): def __init__( self, tokenizer: PreTrainedTokenizer, data_args: DataArguments, data_files: Union[str, List[str]], **kwargs ): super(MappingTsvDataset, self).__init__(tokenizer, data_args, data_files, **kwargs) def _prepare_data(self, data_args): self.all_columns = data_args.query_column_names if self.is_query else data_args.doc_column_names if self.all_columns is not None: self.all_columns = self.all_columns.split(',') hf_dataset = load_dataset( "csv", data_files=self.data_files, streaming=True, column_names=self.all_columns, delimiter='\t', cache_dir=self.cache_dir )["train"].filter(self.filter_fn) self.dataset = {} for item in hf_dataset: self.dataset[get_idx(item)] = item

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/Retriever/dataset/train_dataset.py

# Adapted from Tevatron (https://github.com/texttron/tevatron) import glob import logging import os import random from typing import Callable, Dict, List, Union from datasets import load_dataset from torch.utils.data import Dataset, IterableDataset from transformers import BatchEncoding, PreTrainedTokenizer from ..arguments import DataArguments, DRPretrainingDataArguments from ..data_augmentation_strategy import Cropping, NullStrategy, SequentialStrategies from ..trainer import DRTrainer logger = logging.getLogger(__name__) class TrainDatasetBase: """ Abstract base class for all train datasets in Openmatch.\n This implants arguments and data preparation, but should be mostly used for identifying an OpenMatch Train Dataset.\n All future dataset ABCs would subclass this and `(Iterable)Dataset`. """ def __init__( self, tokenizer: PreTrainedTokenizer, data_args: DataArguments, trainer: DRTrainer = None, is_eval: bool = False, shuffle_seed: int = None, cache_dir: str = None, ) -> None: self.tokenizer = tokenizer self.data_args = data_args self.q_max_len = data_args.q_max_len self.p_max_len = data_args.p_max_len self.trainer = trainer self.is_eval = is_eval self._prepare_data(data_args, shuffle_seed, cache_dir) def _prepare_data(self, data_args, shuffle_seed, cache_dir): if not self.is_eval: self.data_files = ( [data_args.train_path] if data_args.train_dir is None else glob.glob(os.path.join(data_args.train_dir, "*.jsonl")) ) else: self.data_files = [data_args.eval_path] def get_process_fn(self, epoch, hashed_seed): raise NotImplementedError class StreamTrainDatasetMixin(IterableDataset): def _prepare_data(self, data_args, shuffle_seed, cache_dir): super()._prepare_data(data_args, shuffle_seed, cache_dir) self.dataset = load_dataset( "json", data_files=self.data_files, streaming=True, cache_dir=cache_dir )["train"] self.dataset = ( self.dataset.shuffle(seed=shuffle_seed, buffer_size=10_000) if shuffle_seed is not None else self.dataset ) sample = list(self.dataset.take(1))[0] self.all_columns = sample.keys() def __len__(self): concat_filenames = " ".join(self.data_files) count = 0 with os.popen("wc -l {}".format(concat_filenames)) as f: for line in f: lc, filename = line.strip().split() lc = int(lc) if filename != "total": count += lc return count def __iter__(self): if not self.is_eval: epoch = int(self.trainer.state.epoch) _hashed_seed = hash(self.trainer.args.seed) self.dataset.set_epoch(epoch) return iter( self.dataset.map( self.get_process_fn(epoch, _hashed_seed), remove_columns=self.all_columns, ) ) return iter( self.dataset.map( self.get_process_fn(0, None), remove_columns=self.all_columns ) ) class MappingTrainDatasetMixin(Dataset): def _prepare_data(self, data_args, shuffle_seed, cache_dir): super()._prepare_data(data_args, shuffle_seed, cache_dir) self.dataset = load_dataset( "json", data_files=self.data_files, streaming=False, cache_dir=cache_dir )["train"] sample = self.dataset[0] self.all_columns = sample.keys() def __len__(self): return len(self.dataset) def __getitem__(self, index): group = self.dataset[index] if not self.is_eval: epoch = int(self.trainer.state.epoch) _hashed_seed = hash(index + self.trainer.args.seed) return self.get_process_fn(epoch, _hashed_seed)(group) return self.get_process_fn(0, None)(group) class DRTrainDataset(TrainDatasetBase): def create_one_example( self, text_encoding: List[int], is_query=False ) -> BatchEncoding: item = self.tokenizer.encode_plus( text_encoding, truncation="only_first", max_length=self.data_args.q_max_len if is_query else self.data_args.p_max_len, padding=False, return_attention_mask=False, return_token_type_ids=False, ) return item def get_process_fn(self, epoch, hashed_seed): def process_fn(example): qry = example["query"] encoded_query = self.create_one_example(qry, is_query=True) encoded_passages = [] group_positives = example["positives"] group_negatives = example["negatives"] if not self.data_args.use_all_positive_passages: if self.data_args.positive_passage_no_shuffle or hashed_seed is None: pos_psg = group_positives[0] else: pos_psg = group_positives[ (hashed_seed + epoch) % len(group_positives) ] encoded_passages.append(self.create_one_example(pos_psg)) else: for pos_psg in group_positives: encoded_passages.append(self.create_one_example(pos_psg)) negative_size = self.data_args.train_n_passages - 1 if len(group_negatives) < negative_size: if hashed_seed is not None: negs = random.choices(group_negatives, k=negative_size) else: negs = [x for x in group_negatives] negs = negs * 2 negs = negs[:negative_size] elif self.data_args.train_n_passages == 1: negs = [] elif self.data_args.negative_passage_no_shuffle: negs = group_negatives[:negative_size] else: _offset = epoch * negative_size % len(group_negatives) negs = [x for x in group_negatives] if hashed_seed is not None: random.Random(hashed_seed).shuffle(negs) negs = negs * 2 negs = negs[_offset : _offset + negative_size] for neg_psg in negs: encoded_passages.append(self.create_one_example(neg_psg)) if not self.data_args.use_all_positive_passages: assert len(encoded_passages) == self.data_args.train_n_passages else: assert ( len(encoded_passages) == self.data_args.train_n_passages + len(group_positives) - 1 ) return { "query_": encoded_query, "passages": encoded_passages, } # Avoid name conflict with query in the original dataset return process_fn class StreamDRTrainDataset(StreamTrainDatasetMixin, DRTrainDataset): pass class MappingDRTrainDataset(MappingTrainDatasetMixin, DRTrainDataset): pass class DRPretrainDataset(TrainDatasetBase): def __init__( self, tokenizer: PreTrainedTokenizer, data_args: DRPretrainingDataArguments, trainer: DRTrainer = None, is_eval: bool = False, shuffle_seed: int = None, cache_dir: str = None, ) -> None: super(DRPretrainDataset, self).__init__( tokenizer, data_args, trainer, is_eval, shuffle_seed, cache_dir ) pretrain_strategies_str = ( data_args.pretrain_strategies.split(",") if data_args.pretrain_strategies is not None else [] ) strategies = [] for strategy_str in pretrain_strategies_str: if strategy_str == "null": strategies.append(NullStrategy()) logger.info("Adding NullStrategy") elif strategy_str == "crop": strategies.append( Cropping( ratio_min=data_args.cropping_ratio_min, ratio_max=data_args.cropping_ratio_max, ) ) logger.info( "Adding Cropping, ratio_min={}, ratio_max={}".format( data_args.cropping_ratio_min, data_args.cropping_ratio_max ) ) else: raise ValueError( "Unknown pretraining strategy: {}".format(strategy_str) ) self.apply_strategy = SequentialStrategies(*strategies) def create_one_example( self, text_encoding: List[int], is_query=False ) -> BatchEncoding: text_encoding = self.apply_strategy(text_encoding) item = self.tokenizer.encode_plus( text_encoding, truncation="only_first", max_length=self.data_args.q_max_len if is_query else self.data_args.p_max_len, padding=False, return_attention_mask=False, return_token_type_ids=False, ) return item def get_process_fn(self, epoch, hashed_seed): def process_fn(example): content = example[self.data_args.pretrain_target_field] encoded_query = self.create_one_example(content, is_query=True) encoded_passages = [self.create_one_example(content)] return {"query": encoded_query, "passages": encoded_passages} return process_fn class StreamDRPretrainDataset(StreamTrainDatasetMixin, DRPretrainDataset): pass class MappingDRPretrainDataset(MappingTrainDatasetMixin, DRPretrainDataset): pass class RRTrainDataset(TrainDatasetBase): def create_one_example(self, qry_encoding, psg_encoding) -> BatchEncoding: item = self.tokenizer.encode_plus( qry_encoding + psg_encoding, truncation="longest_first", max_length=self.data_args.q_max_len + self.data_args.p_max_len + 2, padding=False, return_attention_mask=False, return_token_type_ids=False, ) return item def get_process_fn(self, epoch, hashed_seed): def process_fn(example): qry = example["query"] group_positives = example["positives"] group_negatives = example["negatives"] if self.data_args.positive_passage_no_shuffle or hashed_seed is None: pos_psg = group_positives[0] else: pos_psg = group_positives[(hashed_seed + epoch) % len(group_positives)] encoded_pos_pair = self.create_one_example(qry, pos_psg) if hashed_seed is None: neg_psg = group_negatives[0] else: neg_psg = group_negatives[(hashed_seed + epoch) % len(group_negatives)] encoded_neg_pair = self.create_one_example(qry, neg_psg) return {"pos_pair": encoded_pos_pair, "neg_pair": encoded_neg_pair} return process_fn class StreamRRTrainDataset(StreamTrainDatasetMixin, RRTrainDataset): pass class MappingRRTrainDataset(MappingTrainDatasetMixin, RRTrainDataset): pass

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/Retriever/modeling/reranking_model.py

import copy import json import logging import os from dataclasses import dataclass from typing import Dict, Optional import torch import torch.distributed as dist import torch.nn as nn import torch.nn.functional as F from torch import Tensor from transformers import (AutoModel, BatchEncoding, PreTrainedModel, T5EncoderModel, PreTrainedTokenizer, AutoConfig, T5ForConditionalGeneration) from transformers.modeling_outputs import ModelOutput from ..arguments import DataArguments from ..arguments import RRTrainingArguments as TrainingArguments from ..arguments import ModelArguments from ..loss import rr_loss_functions, CrossEntropyLoss from ..utils import mean_pooling from .linear import LinearHead logger = logging.getLogger(__name__) @dataclass class RROutput(ModelOutput): pos_pair_scores: Tensor = None neg_pair_scores: Tensor = None loss: Tensor = None class RRModel(nn.Module): def __init__( self, lm: PreTrainedModel, head: nn.Module, feature: str = "last_hidden_state", pooling: str = "first", pos_token: str = None, neg_token: str = None, tokenizer: PreTrainedTokenizer = None, model_args: ModelArguments = None, data_args: DataArguments = None, train_args: TrainingArguments = None, ): super().__init__() self.lm = lm self.head = head self.feature = feature self.pooling = pooling self.pos_token = pos_token self.neg_token = neg_token self.tokenizer = tokenizer self.pos_token_id = tokenizer.encode(self.pos_token, add_special_tokens=False)[0] if self.pos_token else None self.neg_token_id = tokenizer.encode(self.neg_token, add_special_tokens=False)[0] if self.neg_token else None self.model_args = model_args self.data_args = data_args self.train_args = train_args if train_args is not None: self.loss_fn_str = train_args.loss_fn self.loss_fn = rr_loss_functions[self.loss_fn_str]() self.margin = train_args.margin if "T5" in type(self.lm).__name__ and not self.model_args.encoder_only: self.loss_fn_str = "ce" self.loss_fn = CrossEntropyLoss() def _get_config_dict(self): config = { "plm_backbone": { "type": type(self.lm).__name__, "feature": self.feature, }, "pooling": self.pooling, "pos_token": self.pos_token, "neg_token": self.neg_token, } return config def forward( self, pos_pairs: Dict[str, Tensor] = None, neg_pairs: Dict[str, Tensor] = None, ): pos_pair_scores = self.encode(pos_pairs) neg_pair_scores = self.encode(neg_pairs) if self.loss_fn_str in ["mr", "smr"]: loss = self.loss_fn(pos_pair_scores, neg_pair_scores, margin=self.margin) else: loss = self.loss_fn(pos_pair_scores, neg_pair_scores) return RROutput( loss=loss, pos_pair_scores=pos_pair_scores, neg_pair_scores=neg_pair_scores, ) def encode(self, items): if items is None: return None, None items = BatchEncoding(items) if "T5" in type(self.lm).__name__ and not self.model_args.encoder_only: decoder_input_ids = torch.zeros((items.input_ids.shape[0], 1), dtype=torch.long).to(items.input_ids.device) items_out = self.lm(**items, decoder_input_ids=decoder_input_ids, return_dict=True) logits = items_out.logits scores = logits[:, 0, [self.neg_token_id, self.pos_token_id]] # batch_size * 2 else: items_out = self.lm(**items, return_dict=True) hidden = getattr(items_out, self.feature) if self.pooling == "first": reps = hidden[:, 0, :] elif self.pooling == "mean": reps = mean_pooling(hidden, items.attention_mask) else: raise ValueError("Unknown pooling type: {}".format(self.pooling)) scores = self.head(reps) # batch_size * 1 return scores @classmethod def build( cls, model_args: ModelArguments, data_args: DataArguments = None, train_args: TrainingArguments = None, tokenizer: PreTrainedTokenizer = None, **hf_kwargs, ): # load local config = None model_class = None hf_config = AutoConfig.from_pretrained(model_args.model_name_or_path, **hf_kwargs) if model_args.encoder_only: model_class = T5EncoderModel elif "T5" in hf_config.architectures[0]: # Pre-trained T5 model model_class = T5ForConditionalGeneration else: model_class = AutoModel if os.path.exists(os.path.join(model_args.model_name_or_path, "openmatch_config.json")): with open(os.path.join(model_args.model_name_or_path, "openmatch_config.json")) as f: config = json.load(f) if os.path.isdir(model_args.model_name_or_path) and config is not None: # not a raw Huggingface model logger.info(f'loading reranking model weight from {model_args.model_name_or_path}') lm = model_class.from_pretrained( model_args.model_name_or_path, **hf_kwargs ) head = LinearHead.load(ckpt_dir=model_args.model_name_or_path) else: # a Huggingface model lm = model_class.from_pretrained(model_args.model_name_or_path, **hf_kwargs) head = LinearHead(model_args.projection_in_dim, 1) model = cls( lm=lm, head=head, feature=model_args.feature if config is None else config["plm_backbone"]["feature"], pooling=model_args.pooling if config is None else config["pooling"], pos_token=model_args.pos_token if config is None else config["pos_token"], neg_token=model_args.neg_token if config is None else config["neg_token"], tokenizer=tokenizer, model_args=model_args, data_args=data_args, train_args=train_args, ) return model def save(self, output_dir: str): self.lm.save_pretrained(output_dir) self.head.save(output_dir) with open(os.path.join(output_dir, 'openmatch_config.json'), 'w') as f: json.dump(self._get_config_dict(), f, indent=4)

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/Retriever/modeling/linear.py

import logging import os import json import torch import torch.nn as nn from torch import Tensor logger = logging.getLogger(__name__) class LinearHead(nn.Module): def __init__( self, input_dim: int = 768, output_dim: int = 768, ): super(LinearHead, self).__init__() self.linear = nn.Linear(input_dim, output_dim, bias=False) self.config = {'input_dim': input_dim, 'output_dim': output_dim} def forward(self, rep: Tensor = None): return self.linear(rep) @classmethod def load(cls, ckpt_dir: str): logger.info(f'Loading linear head from {ckpt_dir}') model_path = os.path.join(ckpt_dir, 'linear.pt') config_path = os.path.join(ckpt_dir, 'head_config.json') with open(config_path, 'r') as f: config = json.load(f) model = cls(**config) model.load_state_dict(torch.load(model_path)) return model def save(self, save_path): torch.save(self.state_dict(), os.path.join(save_path, 'linear.pt')) with open(os.path.join(save_path, 'head_config.json'), 'w') as f: json.dump(self.config, f, indent=4)

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/Retriever/modeling/dense_retrieval_model.py

# Adapted from Tevatron (https://github.com/texttron/tevatron) import copy import importlib import json import logging import os from dataclasses import dataclass from typing import Dict, Optional import torch import torch.distributed as dist import torch.nn as nn import torch.nn.functional as F from torch import Tensor from transformers import ( AutoConfig, AutoModel, BatchEncoding, PreTrainedModel, T5EncoderModel, ) from transformers.modeling_outputs import ModelOutput from ..arguments import DataArguments from ..arguments import DRTrainingArguments as TrainingArguments from ..arguments import ModelArguments from ..utils import mean_pooling from .linear import LinearHead logger = logging.getLogger(__name__) @dataclass class DROutput(ModelOutput): q_reps: Tensor = None p_reps: Tensor = None loss: Tensor = None scores: Tensor = None class DRModel(nn.Module): def __init__( self, lm_q: PreTrainedModel, lm_p: PreTrainedModel, tied: bool = True, feature: str = "last_hidden_state", pooling: str = "first", head_q: nn.Module = None, head_p: nn.Module = None, normalize: bool = False, model_args: ModelArguments = None, data_args: DataArguments = None, train_args: TrainingArguments = None, ): super().__init__() self.tied = tied self.lm_q = lm_q self.lm_p = lm_p self.head_q = head_q self.head_p = head_p self.loss_fn = nn.CrossEntropyLoss(reduction="mean") self.feature = feature self.pooling = pooling self.normalize = normalize self.model_args = model_args self.train_args = train_args self.data_args = data_args if train_args is not None and train_args.negatives_x_device: if not dist.is_initialized(): raise ValueError( "Distributed training has not been initialized for representation all gather." ) self.process_rank = dist.get_rank() self.world_size = dist.get_world_size() def _get_config_dict(self): config = { "tied": self.tied, "plm_backbone": { "type": type(self.lm_q).__name__, "feature": self.feature, }, "pooling": self.pooling, "linear_head": bool(self.head_q), "normalize": self.normalize, } return config def forward( self, query: Dict[str, Tensor] = None, passage: Dict[str, Tensor] = None, ): q_hidden, q_reps = self.encode_query(query) p_hidden, p_reps = self.encode_passage(passage) if q_reps is None or p_reps is None: return DROutput(q_reps=q_reps, p_reps=p_reps) # if self.training: if self.train_args.negatives_x_device: q_reps = self.dist_gather_tensor(q_reps) p_reps = self.dist_gather_tensor(p_reps) effective_bsz = ( self.train_args.per_device_train_batch_size * self.world_size if self.train_args.negatives_x_device else self.train_args.per_device_train_batch_size ) scores = torch.matmul(q_reps, p_reps.transpose(0, 1)) # scores = torch.matmul(q_reps, p_reps.transpose(0, 1)) / 0.05 # contriever if not self.data_args.use_all_positive_passages: target = torch.arange( scores.size(0), device=scores.device, dtype=torch.long ) target = target * self.data_args.train_n_passages loss = self.loss_fn(scores, target) else: batch_size = scores.size(0) n_total_passages = int(scores.size(1) / scores.size(0)) n_positive_passages = n_total_passages - ( self.data_args.train_n_passages - 1 ) losses = None num = 0 target = torch.arange(1, device=scores.device, dtype=torch.long) for i in range(batch_size): indices = [0] positive_indices = [ i * n_total_passages + j for j in range(n_positive_passages) ] for j in range(scores.size(1)): if j not in positive_indices: indices.append(j) for j in range(n_positive_passages): indices[0] = i * n_total_passages + j now_scores = scores[i][indices].unsqueeze(0) loss = self.loss_fn(now_scores, target) losses = losses + loss if losses != None else loss num += 1 loss = losses / num if self.training and self.train_args.negatives_x_device: loss = loss * self.world_size # counter average weight reduction return DROutput(loss=loss, scores=scores, q_reps=q_reps, p_reps=p_reps) def encode(self, items, model, head): if items is None: return None, None items = BatchEncoding(items) if "T5" in type(model).__name__ and not self.model_args.encoder_only: decoder_input_ids = torch.zeros( (items.input_ids.shape[0], 1), dtype=torch.long ).to(items.input_ids.device) items_out = model( **items, decoder_input_ids=decoder_input_ids, return_dict=True ) hidden = items_out.last_hidden_state reps = hidden[:, 0, :] else: items_out = model(**items, return_dict=True) hidden = getattr(items_out, self.feature) if self.pooling == "first": reps = hidden[:, 0, :] elif self.pooling == "mean": reps = mean_pooling(hidden, items.attention_mask) elif self.pooling == "no": reps = hidden else: raise ValueError("Unknown pooling type: {}".format(self.pooling)) if head is not None: reps = head(reps) # D * d if self.normalize: reps = F.normalize(reps, dim=1) return hidden, reps def encode_passage(self, psg): return self.encode(psg, self.lm_p, self.head_p) def encode_query(self, qry): return self.encode(qry, self.lm_q, self.head_q) @classmethod def build( cls, model_args: ModelArguments, data_args: DataArguments = None, train_args: TrainingArguments = None, **hf_kwargs, ): # load local config = None head_q = head_p = None if os.path.exists( os.path.join(model_args.model_name_or_path, "openmatch_config.json") ): with open( os.path.join(model_args.model_name_or_path, "openmatch_config.json") ) as f: config = json.load(f) if ( os.path.isdir(model_args.model_name_or_path) and config is not None ): # an OpenMatch model tied = config["tied"] if tied: logger.info( f"loading query model weight from {model_args.model_name_or_path}" ) model_name = config["plm_backbone"]["type"] model_class = getattr( importlib.import_module("transformers"), model_name ) lm_q = lm_p = model_class.from_pretrained( model_args.model_name_or_path, **hf_kwargs ) if config["linear_head"]: head_q = head_p = LinearHead.load(model_args.model_name_or_path) else: _qry_model_path = os.path.join( model_args.model_name_or_path, "query_model" ) _psg_model_path = os.path.join( model_args.model_name_or_path, "passage_model" ) _qry_head_path = os.path.join( model_args.model_name_or_path, "query_head" ) _psg_head_path = os.path.join( model_args.model_name_or_path, "passage_head" ) logger.info(f"loading query model weight from {_qry_model_path}") model_name = config["plm_backbone"]["lm_q_type"] model_class = getattr( importlib.import_module("transformers"), model_name ) if os.path.exists(os.path.join(_qry_model_path, "config.json")): logger.info(f"loading query model config from {_qry_model_path}") qry_model_config = AutoConfig.from_pretrained(_qry_model_path) hf_kwargs["config"] = qry_model_config lm_q = model_class.from_pretrained(_qry_model_path, **hf_kwargs) logger.info(f"loading passage model weight from {_psg_model_path}") model_name = config["plm_backbone"]["lm_p_type"] model_class = getattr( importlib.import_module("transformers"), model_name ) if os.path.exists(os.path.join(_psg_model_path, "config.json")): logger.info(f"loading passage model config from {_psg_model_path}") psg_model_config = AutoConfig.from_pretrained(_psg_model_path) hf_kwargs["config"] = psg_model_config lm_p = model_class.from_pretrained(_psg_model_path, **hf_kwargs) if config["linear_head"]: head_q = LinearHead.load(_qry_head_path) head_p = LinearHead.load(_psg_head_path) else: # a Huggingface model tied = not model_args.untie_encoder model_class = T5EncoderModel if model_args.encoder_only else AutoModel lm_q = model_class.from_pretrained( model_args.model_name_or_path, **hf_kwargs ) lm_p = copy.deepcopy(lm_q) if not tied else lm_q if model_args.add_linear_head: head_q = LinearHead( model_args.projection_in_dim, model_args.projection_out_dim ) head_p = copy.deepcopy(head_q) if not tied else head_q model = cls( lm_q=lm_q, lm_p=lm_p, tied=tied, feature=model_args.feature if config is None else config["plm_backbone"]["feature"], pooling=model_args.pooling if config is None else config["pooling"], head_q=head_q, head_p=head_p, normalize=model_args.normalize if config is None else config["normalize"], model_args=model_args, data_args=data_args, train_args=train_args, ) return model def save(self, output_dir: str): if not self.tied: os.makedirs(os.path.join(output_dir, "query_model")) os.makedirs(os.path.join(output_dir, "passage_model")) self.lm_q.save_pretrained(os.path.join(output_dir, "query_model")) self.lm_p.save_pretrained(os.path.join(output_dir, "passage_model")) if self.head_q is not None: self.head_q.save(os.path.join(output_dir, "query_head")) self.head_p.save(os.path.join(output_dir, "passage_head")) else: self.lm_q.save_pretrained(output_dir) if self.head_q is not None: self.head_q.save(output_dir) with open(os.path.join(output_dir, "openmatch_config.json"), "w") as f: json.dump(self._get_config_dict(), f, indent=4) def dist_gather_tensor(self, t: Optional[torch.Tensor]): if t is None: return None t = t.contiguous() all_tensors = [torch.empty_like(t) for _ in range(self.world_size)] dist.all_gather(all_tensors, t) all_tensors[self.process_rank] = t all_tensors = torch.cat(all_tensors, dim=0) return all_tensors class DRModelForInference(DRModel): def __init__(self, *args, **kwargs): super().__init__(*args, **kwargs) # self.eval() @torch.no_grad() def encode_passage(self, psg): return super(DRModelForInference, self).encode_passage(psg) @torch.no_grad() def encode_query(self, qry): return super(DRModelForInference, self).encode_query(qry) def forward( self, query: Dict[str, Tensor] = None, passage: Dict[str, Tensor] = None, ): q_hidden, q_reps = self.encode_query(query) p_hidden, p_reps = self.encode_passage(passage) return DROutput(q_reps=q_reps, p_reps=p_reps)

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/Retriever/retriever/reranker.py

import logging import os from contextlib import nullcontext from typing import Dict import torch from torch.cuda import amp from torch.nn import functional as F from torch.utils.data import DataLoader, Dataset, IterableDataset from tqdm import tqdm from transformers import PreTrainedTokenizer from transformers.trainer_pt_utils import IterableDatasetShard from ..arguments import InferenceArguments as EncodingArguments from ..dataset import InferenceDataset, RRInferenceCollator from ..modeling import RRModel from ..utils import (load_from_trec, merge_retrieval_results_by_score, save_as_trec) logger = logging.getLogger(__name__) def encode_pair(tokenizer, item1, item2, max_len_1=32, max_len_2=128): return tokenizer.encode_plus( item1 + item2, truncation='longest_first', padding='max_length', max_length=max_len_1 + max_len_2 + 2, ) def add_to_result_dict(result_dicts, qids, dids, scores): for qid, did, score in zip(qids, dids, scores): if qid not in result_dicts: result_dicts[qid] = {} result_dicts[qid][did] = float(score) class RRPredictDataset(IterableDataset): def __init__( self, tokenizer: PreTrainedTokenizer, query_dataset: InferenceDataset, corpus_dataset: InferenceDataset, run: Dict[str, Dict[str, float]] ): super(RRPredictDataset, self).__init__() self.tokenizer = tokenizer self.query_dataset = query_dataset self.corpus_dataset = corpus_dataset self.run = run def __iter__(self): def gen_q_d_pair(): for qid, did_and_scores in self.run.items(): for did, _ in did_and_scores.items(): yield { "query_id": qid, "doc_id": did, **encode_pair( self.tokenizer, self.query_dataset[qid]["input_ids"], self.corpus_dataset[did]["input_ids"], self.query_dataset.max_len, self.corpus_dataset.max_len ), } return gen_q_d_pair() class Reranker: def __init__( self, model: RRModel, tokenizer: PreTrainedTokenizer, corpus_dataset: Dataset, args: EncodingArguments ): logger.info("Initializing reranker") self.model = model self.tokenizer = tokenizer self.corpus_dataset = corpus_dataset self.args = args self.model = model.to(self.args.device) self.model.eval() def rerank(self, query_dataset: InferenceDataset, run: Dict[str, Dict[str, float]]): return_dict = {} dataset = RRPredictDataset(self.tokenizer, query_dataset, self.corpus_dataset, run) if self.args.world_size > 1: dataset = IterableDatasetShard( dataset, batch_size=self.args.per_device_eval_batch_size, drop_last=False, num_processes=self.args.world_size, process_index=self.args.process_index ) dataloader = DataLoader( dataset, batch_size=self.args.eval_batch_size, collate_fn=RRInferenceCollator(), num_workers=self.args.dataloader_num_workers, pin_memory=self.args.dataloader_pin_memory, ) with torch.no_grad(): for qids, dids, batch in tqdm(dataloader, desc="Reranking", disable=self.args.local_process_index > 0): with amp.autocast() if self.args.fp16 else nullcontext(): for k, v in batch.items(): batch[k] = v.to(self.args.device) outputs = self.model.encode(batch) if len(outputs.shape) == 2 and outputs.shape[1] == 2: outputs = F.log_softmax(outputs, dim=1)[:, 1] scores = outputs.cpu().numpy() add_to_result_dict(return_dict, qids, dids, scores) if self.args.world_size > 1: save_as_trec(return_dict, self.args.trec_save_path + ".rank.{}".format(self.args.process_index)) torch.distributed.barrier() if self.args.process_index == 0: # aggregate results all_results = [] for i in range(self.args.world_size): all_results.append(load_from_trec(self.args.trec_save_path + ".rank.{}".format(i))) return_dict = merge_retrieval_results_by_score(all_results) # remove temp files for i in range(self.args.world_size): os.remove(self.args.trec_save_path + ".rank.{}".format(i)) torch.distributed.barrier() return return_dict

py

Augmentation-Adapted-Retriever

Augmentation-Adapted-Retriever-main/src/Retriever/retriever/dense_retriever.py

import gc import glob import logging import os import pickle from contextlib import nullcontext from typing import Dict, List import faiss import numpy as np import torch from torch.cuda import amp from torch.utils.data import DataLoader, IterableDataset from tqdm import tqdm from ..arguments import InferenceArguments as EncodingArguments from ..dataset import DRInferenceCollator from ..modeling import DRModelForInference, DROutput from ..utils import merge_retrieval_results_by_score logger = logging.getLogger(__name__) class Retriever: def __init__(self, model: DRModelForInference, corpus_dataset: IterableDataset, args: EncodingArguments): logger.info("Initializing retriever") self.model = model self.corpus_dataset = corpus_dataset self.args = args self.doc_lookup = [] self.query_lookup = [] self.model.to(self.args.device) self.model.eval() def _initialize_faiss_index(self, dim: int): self.index = None cpu_index = faiss.IndexFlatIP(dim) self.index = cpu_index def _move_index_to_gpu(self): logger.info("Moving index to GPU(s)") ngpu = faiss.get_num_gpus() gpu_resources = [] for i in range(ngpu): res = faiss.StandardGpuResources() gpu_resources.append(res) co = faiss.GpuMultipleClonerOptions() co.shard = True co.usePrecomputed = False vres = faiss.GpuResourcesVector() vdev = faiss.IntVector() for i in range(0, ngpu): vdev.push_back(i) vres.push_back(gpu_resources[i]) self.index = faiss.index_cpu_to_gpu_multiple( vres, vdev, self.index, co) def doc_embedding_inference(self): # Note: during evaluation, there's no point in wrapping the model # inside a DistributedDataParallel as we'll be under `no_grad` anyways. if self.corpus_dataset is None: raise ValueError("No corpus dataset provided") dataloader = DataLoader( self.corpus_dataset, # Note that we do not support DataParallel here batch_size=self.args.per_device_eval_batch_size, collate_fn=DRInferenceCollator(), num_workers=self.args.dataloader_num_workers, pin_memory=self.args.dataloader_pin_memory, ) os.makedirs(self.args.output_dir, exist_ok=True) encoded = [] lookup_indices = [] idx = 0 prev_idx = 0 for (batch_ids, batch) in tqdm(dataloader, disable=self.args.process_index > 0): lookup_indices.extend(batch_ids) idx += len(batch_ids) with amp.autocast() if self.args.fp16 else nullcontext(): with torch.no_grad(): for k, v in batch.items(): batch[k] = v.to(self.args.device) model_output: DROutput = self.model(passage=batch) encoded.append(model_output.p_reps.cpu().detach().numpy()) if len(lookup_indices) >= self.args.max_inmem_docs // self.args.world_size: encoded = np.concatenate(encoded) with open(os.path.join(self.args.output_dir, "embeddings.corpus.rank.{}.{}-{}".format(self.args.process_index, prev_idx, idx)), 'wb') as f: pickle.dump((encoded, lookup_indices), f, protocol=4) encoded = [] lookup_indices = [] prev_idx = idx gc.collect() if len(lookup_indices) > 0: encoded = np.concatenate(encoded) with open(os.path.join(self.args.output_dir, "embeddings.corpus.rank.{}.{}-{}".format(self.args.process_index, prev_idx, idx)), 'wb') as f: pickle.dump((encoded, lookup_indices), f, protocol=4) del encoded del lookup_indices if self.args.world_size > 1: torch.distributed.barrier() def init_index_and_add(self, partition: str = None): logger.info( "Initializing Faiss index from pre-computed document embeddings") partitions = [partition] if partition is not None else glob.glob( os.path.join(self.args.output_dir, "embeddings.corpus.rank.*")) for i, part in enumerate(partitions): with open(part, 'rb') as f: data = pickle.load(f) encoded = data[0] lookup_indices = data[1] if i == 0: dim = encoded.shape[1] self._initialize_faiss_index(dim) self.index.add(encoded) self.doc_lookup.extend(lookup_indices) @classmethod def build_all(cls, model: DRModelForInference, corpus_dataset: IterableDataset, args: EncodingArguments): retriever = cls(model, corpus_dataset, args) retriever.doc_embedding_inference() if args.process_index == 0: retriever.init_index_and_add() if args.world_size > 1: torch.distributed.barrier() return retriever @classmethod def build_embeddings(cls, model: DRModelForInference, corpus_dataset: IterableDataset, args: EncodingArguments): retriever = cls(model, corpus_dataset, args) retriever.doc_embedding_inference() return retriever @classmethod def from_embeddings(cls, model: DRModelForInference, args: EncodingArguments): retriever = cls(model, None, args) if args.process_index == 0: retriever.init_index_and_add() if args.world_size > 1: torch.distributed.barrier() return retriever def reset_index(self): if self.index: self.index.reset() self.doc_lookup = [] self.query_lookup = [] def query_embedding_inference(self, query_dataset: IterableDataset): dataloader = DataLoader( query_dataset, batch_size=self.args.per_device_eval_batch_size, collate_fn=DRInferenceCollator(), num_workers=self.args.dataloader_num_workers, pin_memory=self.args.dataloader_pin_memory, ) encoded = [] lookup_indices = [] for (batch_ids, batch) in tqdm(dataloader, disable=self.args.process_index > 0): lookup_indices.extend(batch_ids) with amp.autocast() if self.args.fp16 else nullcontext(): with torch.no_grad(): for k, v in batch.items(): batch[k] = v.to(self.args.device) if not self.args.encode_query_as_passage: model_output: DROutput = self.model(query=batch) encoded.append( model_output.q_reps.cpu().detach().numpy()) else: model_output: DROutput = self.model(passage=batch) encoded.append( model_output.p_reps.cpu().detach().numpy()) if len(encoded) > 0: # If there is no data in the process, we don't do anything encoded = np.concatenate(encoded) with open(os.path.join(self.args.output_dir, "embeddings.query.rank.{}".format(self.args.process_index)), 'wb') as f: pickle.dump((encoded, lookup_indices), f, protocol=4) if self.args.world_size > 1: torch.distributed.barrier() def search(self, topk: int = 100): logger.info("Searching") if self.index is None: raise ValueError("Index is not initialized") encoded = [] for i in range(self.args.world_size): with open(os.path.join(self.args.output_dir, "embeddings.query.rank.{}".format(i)), 'rb') as f: data = pickle.load(f) lookup_indices = data[1] if len(lookup_indices) == 0: # No data continue encoded.append(data[0]) self.query_lookup.extend(lookup_indices) encoded = np.concatenate(encoded) return_dict = {} D, I = self.index.search(encoded, topk) original_indices = np.array(self.doc_lookup)[I] q = 0 for scores_per_q, doc_indices_per_q in zip(D, original_indices): qid = str(self.query_lookup[q]) return_dict[qid] = {} for doc_index, score in zip(doc_indices_per_q, scores_per_q): doc_index = str(doc_index) if self.args.remove_identical and qid == doc_index: continue return_dict[qid][doc_index] = float(score) q += 1 logger.info("End searching with {} queries".format(len(return_dict))) return return_dict def retrieve(self, query_dataset: IterableDataset, topk: int = 100): self.query_embedding_inference(query_dataset) self.model.cpu() del self.model torch.cuda.empty_cache() results = {} if self.args.process_index == 0: if self.args.use_gpu: self._move_index_to_gpu() results = self.search(topk) if self.args.world_size > 1: torch.distributed.barrier() return results class SuccessiveRetriever(Retriever): def __init__(self, model: DRModelForInference, corpus_dataset: IterableDataset, args: EncodingArguments): super().__init__(model, corpus_dataset, args) @classmethod def from_embeddings(cls, model: DRModelForInference, args: EncodingArguments): retriever = cls(model, None, args) return retriever def retrieve(self, query_dataset: IterableDataset, topk: int = 100): self.query_embedding_inference(query_dataset) del self.model torch.cuda.empty_cache() final_result = {} if self.args.process_index == 0: all_partitions = glob.glob(os.path.join( self.args.output_dir, "embeddings.corpus.rank.*")) for partition in all_partitions: logger.info("Loading partition {}".format(partition)) self.init_index_and_add(partition) if self.args.use_gpu: self._move_index_to_gpu() cur_result = self.search(topk) self.reset_index() final_result = merge_retrieval_results_by_score( [final_result, cur_result], topk) if self.args.world_size > 1: torch.distributed.barrier() return final_result

py

ccc_mse_ser

ccc_mse_ser-master/code/ser_iemocap_gemaps_ccc.py

# Dimensional speech emotion recognition # To evaluate loss function (MSE vs CCC) # Coded by Bagus Tris Atmaja (bagus@ep.its.ac.id) # changelog # 2020-02-13: Modified from gemaps-paa hfs # 2020-02-14: Use 'tanh' activation to lock the output range in [-1, 1] # with RMSprop optimizer import numpy as np import pickle import pandas as pd import keras.backend as K from keras.models import Model from keras.layers import Input, Dense, CuDNNLSTM, Flatten, \ Embedding, Dropout, BatchNormalization, \ RNN, concatenate, Activation from keras.callbacks import EarlyStopping from sklearn.preprocessing import StandardScaler, MinMaxScaler from keras.preprocessing.text import Tokenizer from keras.preprocessing import sequence import random as rn import tensorflow as tf rn.seed(123) np.random.seed(99) tf.set_random_seed(1234) # load feature and labels feat = np.load('/home/s1820002/spro2020/data/feat_ws_3.npy') vad = np.load('/home/s1820002/IEMOCAP-Emotion-Detection/y_egemaps.npy') # use only mean and std feat = feat[:,:-1] # for LSTM input shape (batch, steps, features/channel) #feat = feat.reshape(feat.shape[0], 1, feat.shape[1]) # remove outlier, < 1, > 5 vad = np.where(vad==5.5, 5.0, vad) vad = np.where(vad==0.5, 1.0, vad) # standardization scaled_feature = True # set Dropout do = 0.3 if scaled_feature == True: scaler = StandardScaler() scaler = scaler.fit(feat) #.reshape(feat.shape[0]*feat.shape[1], feat.shape[2])) scaled_feat = scaler.transform(feat) #.reshape(feat.shape[0]*feat.shape[1], feat.shape[2])) #scaled_feat = scaled_feat.reshape(feat.shape[0], feat.shape[1], feat.shape[2]) feat = scaled_feat else: feat = feat scaled_vad = True # standardization if scaled_vad: scaler = MinMaxScaler(feature_range=(-1, 1)) scaler = scaler.fit(vad) #.reshape(vad.shape[0]*vad.shape[1], vad.shape[2])) scaled_vad = scaler.transform(vad) #.reshape(vad.shape[0]*vad.shape[1], vad.shape[2])) vad = scaled_vad else: vad = vad # Concordance correlation coefficient (CCC)-based loss function - using non-inductive statistics def ccc(gold, pred): gold = K.squeeze(gold, axis=-1) pred = K.squeeze(pred, axis=-1) gold_mean = K.mean(gold, axis=-1, keepdims=True) pred_mean = K.mean(pred, axis=-1, keepdims=True) covariance = (gold-gold_mean)*(pred-pred_mean) gold_var = K.mean(K.square(gold-gold_mean), axis=-1, keepdims=True) pred_var = K.mean(K.square(pred-pred_mean), axis=-1, keepdims=True) ccc = K.constant(2.) * covariance / (gold_var + pred_var + K.square(gold_mean - pred_mean) + K.common.epsilon()) return ccc def ccc_loss(gold, pred): # input (num_batches, seq_len, 1) ccc_loss = K.constant(1.) - ccc(gold, pred) return ccc_loss # reshape input feature for LSTM feat = feat.reshape(feat.shape[0], 1, feat.shape[1]) # API model, if use RNN, first two rnn layer must return_sequences=True def api_model(alpha, beta, gamma): # speech network input_speech = Input(shape=(feat.shape[1], feat.shape[2]), name='speech_input') net_speech = BatchNormalization()(input_speech) net_speech = CuDNNLSTM(feat.shape[2], return_sequences=True)(net_speech) net_speech = CuDNNLSTM(256, return_sequences=True)(net_speech) net_speech = CuDNNLSTM(256, return_sequences=False)(net_speech) #net_speech = Flatten()(net_speech) net_speech = Dense(64)(net_speech) #net_speech = Dropout(0.1)(net_speech) target_names = ('v', 'a', 'd') model_combined = [Dense(1, name=name, activation='tanh')(net_speech) for name in target_names] model = Model(input_speech, model_combined) #model.compile(loss=ccc_loss, optimizer='rmsprop', metrics=[ccc]) model.compile(loss=ccc_loss, loss_weights={'v': alpha, 'a': beta, 'd': gamma}, optimizer='rmsprop', metrics=[ccc, 'mse']) return model #def main(alpha, beta, gamma): model = api_model(0.1, 0.5, 0.4) model.summary() # 7869 first data of session 5 (for LOSO) earlystop = EarlyStopping(monitor='val_loss', mode='min', patience=10, restore_best_weights=True) hist = model.fit(feat[:7869], vad[:7869].T.tolist(), batch_size=64, #best:8 validation_split=0.2, epochs=200, verbose=1, shuffle=True, callbacks=[earlystop]) metrik = model.evaluate(feat[7869:], vad[7869:].T.tolist()) print('CCC= ', np.array(metrik)[[-6,-4,-2]]) print('MSE= ', np.array(metrik)[[-5,-3,-1]]) # Plot scatter #va = vad[7869:, :-1] #predik_vad = model.predict(feat[7869:], batch_size=64) #predik_va = np.array(predik_vad).T.reshape(2170,3)[:,:-1] #import matplotlib.pyplot as plt #plt.scatter(va[:,0], va[:,1]) #plt.scatter(predik_va[:,0], predik_va[:,1]) #plt.savefig('scatter_gemaps_mse.pdf') ## check max min #predik_va.max() #predik_va.min()

py

ccc_mse_ser

ccc_mse_ser-master/code/ser_iemocap_gemaps_mse.py

# Dimensional speech emotion recognition # To evaluate loss function (MSE vs CCC) # Coded by Bagus Tris Atmaja (bagus@ep.its.ac.id) # changelog # 2020-02-13: Modified from gemaps-paa hfs import numpy as np import pickle import pandas as pd import keras.backend as K from keras.models import Model from keras.layers import Input, Dense, CuDNNLSTM, Flatten, \ Embedding, Dropout, BatchNormalization, \ RNN, concatenate, Activation from keras.callbacks import EarlyStopping from sklearn.preprocessing import StandardScaler, MinMaxScaler from keras.preprocessing.text import Tokenizer from keras.preprocessing import sequence import random as rn import tensorflow as tf rn.seed(123) np.random.seed(99) tf.set_random_seed(1234) # load feature and labels feat = np.load('/home/s1820002/spro2020/data/feat_ws_3.npy') vad = np.load('/home/s1820002/IEMOCAP-Emotion-Detection/y_egemaps.npy') # use only mean and std feat = feat[:,:-1] # for LSTM input shape (batch, steps, features/channel) #feat = feat.reshape(feat.shape[0], 1, feat.shape[1]) # remove outlier, < 1, > 5 vad = np.where(vad==5.5, 5.0, vad) vad = np.where(vad==0.5, 1.0, vad) # standardization scaled_feature = True if scaled_feature == True: scaler = StandardScaler() scaler = scaler.fit(feat) #.reshape(feat.shape[0]*feat.shape[1], feat.shape[2])) scaled_feat = scaler.transform(feat)#.reshape(feat.shape[0]*feat.shape[1], feat.shape[2])) #scaled_feat = scaled_feat.reshape(feat.shape[0], feat.shape[1], feat.shape[2]) feat = scaled_feat else: feat = feat scaled_vad = True # standardization if scaled_vad: scaler = MinMaxScaler(feature_range=(-1, 1)) scaler = scaler.fit(vad) #.reshape(vad.shape[0]*vad.shape[1], vad.shape[2])) scaled_vad = scaler.transform(vad) #.reshape(vad.shape[0]*vad.shape[1], vad.shape[2])) vad = scaled_vad else: vad = vad # Concordance correlation coefficient (CCC)-based loss function - using non-inductive statistics def ccc(gold, pred): gold = K.squeeze(gold, axis=-1) pred = K.squeeze(pred, axis=-1) gold_mean = K.mean(gold, axis=-1, keepdims=True) pred_mean = K.mean(pred, axis=-1, keepdims=True) covariance = (gold-gold_mean)*(pred-pred_mean) gold_var = K.mean(K.square(gold-gold_mean), axis=-1, keepdims=True) pred_var = K.mean(K.square(pred-pred_mean), axis=-1, keepdims=True) ccc = K.constant(2.) * covariance / (gold_var + pred_var + K.square(gold_mean - pred_mean) + K.common.epsilon()) return ccc def ccc_loss(gold, pred): # input (num_batches, seq_len, 1) ccc_loss = K.constant(1.) - ccc(gold, pred) return ccc_loss # reshape input feature for LSTM feat = feat.reshape(feat.shape[0], 1, feat.shape[1]) # API model, if use RNN, first two rnn layer must return_sequences=True def api_model(alpha, beta, gamma): # speech network input_speech = Input(shape=(feat.shape[1], feat.shape[2]), name='speech_input') net_speech = BatchNormalization()(input_speech) net_speech = CuDNNLSTM(feat.shape[2], return_sequences=True)(net_speech) net_speech = CuDNNLSTM(256, return_sequences=True)(net_speech) net_speech = CuDNNLSTM(256, return_sequences=False)(net_speech) #net_speech = Flatten()(net_speech) net_speech = Dense(64)(net_speech) #net_speech = Dropout(0.1)(net_speech) target_names = ('v', 'a', 'd') model_combined = [Dense(1, name=name, activation='tanh')(net_speech) for name in target_names] model = Model(input_speech, model_combined) #model.compile(loss=ccc_loss, optimizer='rmsprop', metrics=[ccc]) model.compile(loss='mse', loss_weights={'v': alpha, 'a': beta, 'd': gamma}, optimizer='rmsprop', metrics=[ccc, 'mse']) return model #def main(alpha, beta, gamma): model = api_model(0.1, 0.5, 0.4) model.summary() # 7869 first data of session 5 (for LOSO) earlystop = EarlyStopping(monitor='val_loss', mode='min', patience=10, restore_best_weights=True) hist = model.fit(feat[:7869], vad[:7869].T.tolist(), batch_size=64, #best:8 validation_split=0.2, epochs=200, verbose=1, shuffle=True, callbacks=[earlystop]) metrik = model.evaluate(feat[7869:], vad[7869:].T.tolist()) print('CCC= ', np.array(metrik)[[-6,-4,-2]]) print('MSE= ', np.array(metrik)[[-5,-3,-1]]) # Plot scatter #va = vad[7869:, :-1] #predik_vad = model.predict(feat[7869:], batch_size=64) #predik_va = np.array(predik_vad).T.reshape(2170,3)[:,:-1] #import matplotlib.pyplot as plt #plt.scatter(va[:,0], va[:,1]) #plt.scatter(predik_va[:,0], predik_va[:,1]) #plt.savefig('scatter_gemaps_mse.pdf') ## check max min #predik_va.max() #predik_va.min()

py

ccc_mse_ser

ccc_mse_ser-master/code/ser_improv_gemaps_mse.py

# ser_improv_paa_ccc.py # speech emotion recognition for MSP-IMPROV dataset with pyAudioAnalysis # HFS features using CCC-based loss function # coded by Bagus Tris Atmaja (bagus@ep.its.ac.id) # changelog: # 2020-02-13: Inital code, modified from deepMLP repo import numpy as np import pickle import pandas as pd import keras.backend as K from keras.models import Model from keras.layers import Input, Dense, CuDNNLSTM, Flatten, \ Embedding, Dropout, BatchNormalization, \ RNN, concatenate, Activation from keras.callbacks import EarlyStopping from sklearn.preprocessing import StandardScaler, MinMaxScaler from keras.preprocessing.text import Tokenizer from keras.preprocessing import sequence import random as rn import tensorflow as tf rn.seed(123) np.random.seed(99) tf.set_random_seed(1234) # loading file and label feat_train = np.load('/home/s1820002/ccc_mse/data/feat_hfs_gemaps_msp_train.npy') feat_test = np.load('/home/s1820002/ccc_mse/data/feat_hfs_gemaps_msp_test.npy') feat = np.vstack([feat_train, feat_test]) list_path = '/home/s1820002/msp-improv/helper/improv_data.csv' list_file = pd.read_csv(list_path, index_col=None) list_file = pd.DataFrame(list_file) data = list_file.sort_values(by=['wavfile']) vad_train = [] vad_test = [] for index, row in data.iterrows(): #print(row['wavfile'], row['v'], row['a'], row['d']) if int(row['wavfile'][18]) in range(1,6): #print("Process vad..", row['wavfile']) vad_train.append([row['v'], row['a'], row['d']]) else: #print("Process..", row['wavfile']) vad_test.append([row['v'], row['a'], row['d']]) vad = np.vstack([vad_train, vad_test]) # standardization scaled_feature = True if scaled_feature: scaler = StandardScaler() scaler = scaler.fit(feat) scaled_feat = scaler.transform(feat) feat = scaled_feat else: feat = feat scaled_vad = True # standardization if scaled_vad: scaler = MinMaxScaler(feature_range=(-1, 1)) scaler = scaler.fit(vad) #.reshape(vad.shape[0]*vad.shape[1], vad.shape[2])) scaled_vad = scaler.transform(vad) #.reshape(vad.shape[0]*vad.shape[1], vad.shape[2])) vad = scaled_vad else: vad = vad # reshape feat size to match LSTM config feat = feat.reshape(feat.shape[0], 1, feat.shape[1]) # train/test split, LOSO X_train = feat[:len(feat_train)] X_test = feat[len(feat_train):] y_train = vad[:len(vad_train)] y_test = vad[len(vad_train):] # Concordance correlation coefficient (CCC)-based loss function - using non-inductive statistics def ccc(gold, pred): gold = K.squeeze(gold, axis=-1) pred = K.squeeze(pred, axis=-1) gold_mean = K.mean(gold, axis=-1, keepdims=True) pred_mean = K.mean(pred, axis=-1, keepdims=True) covariance = (gold-gold_mean)*(pred-pred_mean) gold_var = K.mean(K.square(gold-gold_mean), axis=-1, keepdims=True) pred_var = K.mean(K.square(pred-pred_mean), axis=-1, keepdims=True) ccc = K.constant(2.) * covariance / (gold_var + pred_var + K.square(gold_mean - pred_mean) + K.common.epsilon()) return ccc def ccc_loss(gold, pred): # input (num_batches, seq_len, 1) ccc_loss = K.constant(1.) - ccc(gold, pred) return ccc_loss # API model, if use RNN, first two rnn layer must return_sequences=True def api_model(): inputs = Input(shape=(feat.shape[1], feat.shape[2]), name='feat_input') net = BatchNormalization()(inputs) #net = Bidirectional(LSTM(64, return_sequences=True, dropout=do, recurrent_dropout=do))(net) net = CuDNNLSTM(feat.shape[2], return_sequences=True)(net) net = CuDNNLSTM(256, return_sequences=True)(net) net = CuDNNLSTM(256, return_sequences=False)(net) net = Dense(64)(net) target_names = ('v', 'a', 'd') outputs = [Dense(1, name=name, activation='tanh')(net) for name in target_names] model = Model(inputs=inputs, outputs=outputs) #=[out1, out2, out3]) model.compile(loss='mse', #{'v': ccc_loss, 'a': ccc_loss, 'd': ccc_loss}, loss_weights={'v': 0.3, 'a': 0.6, 'd': 0.1}, optimizer='rmsprop', metrics=[ccc, 'mse']) return model model2 = api_model() model2.summary() earlystop = EarlyStopping(monitor='val_loss', mode='min', patience=10, restore_best_weights=True) hist = model2.fit(X_train, y_train.T.tolist(), batch_size=64, validation_split=0.2, epochs=50, verbose=1, shuffle=True, callbacks=[earlystop]) metrik = model2.evaluate(X_test, y_test.T.tolist()) print('CCC= ', np.array(metrik)[[-6,-4,-2]]) print('MSE= ', np.array(metrik)[[-5,-3,-1]])

py

ccc_mse_ser

ccc_mse_ser-master/code/ser_iemocap_paa_ccc.py

# Dimensional speech emotion recognition from acoustic # Changelog: # 2019-09-01: initial version # 2019-10-06: optimizer MTL parameters with linear search (in progress) # 2020-12-25: modified fot ser_iemocap_loso_hfs.py # feature is either std+mean or std+mean+silence (uncomment line 44) # 2020-02-13: Modified to evaluate loss function (MSE vs CCC) for EUSIPCO import numpy as np import pickle import pandas as pd import keras.backend as K from keras.models import Model from keras.layers import Input, Dense, CuDNNLSTM, Flatten, \ Embedding, Dropout, BatchNormalization, \ RNN, concatenate, Activation from keras.callbacks import EarlyStopping from sklearn.preprocessing import StandardScaler, MinMaxScaler from keras.preprocessing.text import Tokenizer from keras.preprocessing import sequence import random as rn import tensorflow as tf rn.seed(123) np.random.seed(99) tf.set_random_seed(1234) # load feature and labels feat = np.load('/home/s1820002/atsit/data/feat_34_hfs.npy') vad = np.load('/home/s1820002/IEMOCAP-Emotion-Detection/y_egemaps.npy') # for LSTM input shape (batch, steps, features/channel) #feat = feat.reshape(feat.shape[0], 1, feat.shape[1]) # remove outlier, < 1, > 5 vad = np.where(vad==5.5, 5.0, vad) vad = np.where(vad==0.5, 1.0, vad) # standardization scaled_feature = False # set Dropout do = 0.3 if scaled_feature == True: scaler = StandardScaler() scaler = scaler.fit(feat.reshape(feat.shape[0]*feat.shape[1], feat.shape[2])) scaled_feat = scaler.transform(feat.reshape(feat.shape[0]*feat.shape[1], feat.shape[2])) scaled_feat = scaled_feat.reshape(feat.shape[0], feat.shape[1], feat.shape[2]) feat = scaled_feat else: feat = feat scaled_vad = True # standardization if scaled_vad: scaler = MinMaxScaler(feature_range=(-1, 1)) scaler = scaler.fit(vad) #.reshape(vad.shape[0]*vad.shape[1], vad.shape[2])) scaled_vad = scaler.transform(vad) #.reshape(vad.shape[0]*vad.shape[1], vad.shape[2])) vad = scaled_vad else: vad = vad # Concordance correlation coefficient (CCC)-based loss function - using non-inductive statistics def ccc(gold, pred): gold = K.squeeze(gold, axis=-1) pred = K.squeeze(pred, axis=-1) gold_mean = K.mean(gold, axis=-1, keepdims=True) pred_mean = K.mean(pred, axis=-1, keepdims=True) covariance = (gold-gold_mean)*(pred-pred_mean) gold_var = K.mean(K.square(gold-gold_mean), axis=-1, keepdims=True) pred_var = K.mean(K.square(pred-pred_mean), axis=-1, keepdims=True) ccc = K.constant(2.) * covariance / (gold_var + pred_var + K.square(gold_mean - pred_mean) + K.common.epsilon()) return ccc def ccc_loss(gold, pred): # input (num_batches, seq_len, 1) ccc_loss = K.constant(1.) - ccc(gold, pred) return ccc_loss # API model, if use RNN, first two rnn layer must return_sequences=True def api_model(alpha, beta, gamma): # speech network input_speech = Input(shape=(feat.shape[1], feat.shape[2]), name='speech_input') net_speech = BatchNormalization()(input_speech) net_speech = CuDNNLSTM(feat.shape[2], return_sequences=True)(net_speech) net_speech = CuDNNLSTM(256, return_sequences=True)(net_speech) net_speech = CuDNNLSTM(256, return_sequences=False)(net_speech) #net_speech = Flatten()(net_speech) net_speech = Dense(64)(net_speech) #net_speech = Dropout(0.1)(net_speech) target_names = ('v', 'a', 'd') model_combined = [Dense(1, name=name, activation='tanh')(net_speech) for name in target_names] model = Model(input_speech, model_combined) #model.compile(loss=ccc_loss, optimizer='rmsprop', metrics=[ccc]) model.compile(loss=ccc_loss, loss_weights={'v': alpha, 'a': beta, 'd': gamma}, optimizer='rmsprop', metrics=[ccc, 'mse']) return model #def main(alpha, beta, gamma): model = api_model(0.1, 0.5, 0.4) model.summary() # 7869 first data of session 5 (for LOSO) earlystop = EarlyStopping(monitor='val_loss', mode='min', patience=10, restore_best_weights=True) hist = model.fit(feat[:7869], vad[:7869].T.tolist(), batch_size=64, #best:8 validation_split=0.2, epochs=200, verbose=1, shuffle=True, callbacks=[earlystop]) metrik = model.evaluate(feat[7869:], vad[7869:].T.tolist()) print('CCC= ', np.array(metrik)[[-6,-4,-2]]) print('MSE= ', np.array(metrik)[[-5,-3,-1]])

py

ccc_mse_ser

ccc_mse_ser-master/code/ser_improv_paa_ccc.py

# ser_improv_paa_ccc.py # speech emotion recognition for MSP-IMPROV dataset with pyAudioAnalysis # HFS features using CCC-based loss function # coded by Bagus Tris Atmaja (bagus@ep.its.ac.id) # changelog: # 2020-02-13: Inital code, modified from deepMLP repo import numpy as np import pickle import pandas as pd import keras.backend as K from keras.models import Model from keras.layers import Input, Dense, CuDNNLSTM, Flatten, \ Embedding, Dropout, BatchNormalization, \ RNN, concatenate, Activation from keras.callbacks import EarlyStopping from sklearn.preprocessing import StandardScaler, MinMaxScaler from keras.preprocessing.text import Tokenizer from keras.preprocessing import sequence import random as rn import tensorflow as tf rn.seed(123) np.random.seed(99) tf.set_random_seed(1234) # loading file and label feat_train = np.load('/home/s1820002/ccc_mse/data/feat_hfs_paa_msp_train.npy') feat_test = np.load('/home/s1820002/ccc_mse/data/feat_hfs_paa_msp_test.npy') feat = np.vstack([feat_train, feat_test]) list_path = '/home/s1820002/msp-improv/helper/improv_data.csv' list_file = pd.read_csv(list_path, index_col=None) list_file = pd.DataFrame(list_file) data = list_file.sort_values(by=['wavfile']) vad_train = [] vad_test = [] for index, row in data.iterrows(): #print(row['wavfile'], row['v'], row['a'], row['d']) if int(row['wavfile'][18]) in range(1,6): #print("Process vad..", row['wavfile']) vad_train.append([row['v'], row['a'], row['d']]) else: #print("Process..", row['wavfile']) vad_test.append([row['v'], row['a'], row['d']]) vad = np.vstack([vad_train, vad_test]) # standardization scaled_feature = False if scaled_feature: scaler = StandardScaler() scaler = scaler.fit(feat) scaled_feat = scaler.transform(feat) feat = scaled_feat else: feat = feat scaled_vad = True # standardization if scaled_vad: scaler = MinMaxScaler(feature_range=(-1, 1)) scaler = scaler.fit(vad) #.reshape(vad.shape[0]*vad.shape[1], vad.shape[2])) scaled_vad = scaler.transform(vad) #.reshape(vad.shape[0]*vad.shape[1], vad.shape[2])) vad = scaled_vad else: vad = vad # reshape feat size to match LSTM config feat = feat.reshape(feat.shape[0], 1, feat.shape[1]) # train/test split, LOSO X_train = feat[:len(feat_train)] X_test = feat[len(feat_train):] y_train = vad[:len(vad_train)] y_test = vad[len(vad_train):] # Concordance correlation coefficient (CCC)-based loss function - using non-inductive statistics def ccc(gold, pred): gold = K.squeeze(gold, axis=-1) pred = K.squeeze(pred, axis=-1) gold_mean = K.mean(gold, axis=-1, keepdims=True) pred_mean = K.mean(pred, axis=-1, keepdims=True) covariance = (gold-gold_mean)*(pred-pred_mean) gold_var = K.mean(K.square(gold-gold_mean), axis=-1, keepdims=True) pred_var = K.mean(K.square(pred-pred_mean), axis=-1, keepdims=True) ccc = K.constant(2.) * covariance / (gold_var + pred_var + K.square(gold_mean - pred_mean) + K.common.epsilon()) return ccc def ccc_loss(gold, pred): # input (num_batches, seq_len, 1) ccc_loss = K.constant(1.) - ccc(gold, pred) return ccc_loss # API model, if use RNN, first two rnn layer must return_sequences=True def api_model(): inputs = Input(shape=(feat.shape[1], feat.shape[2]), name='feat_input') net = BatchNormalization()(inputs) #net = Bidirectional(LSTM(64, return_sequences=True, dropout=do, recurrent_dropout=do))(net) net = CuDNNLSTM(feat.shape[2], return_sequences=True)(net) net = CuDNNLSTM(256, return_sequences=True)(net) net = CuDNNLSTM(256, return_sequences=False)(net) net = Dense(64)(net) #net = Dropout(0.1)(net) target_names = ('v', 'a', 'd') outputs = [Dense(1, name=name, activation='tanh')(net) for name in target_names] model = Model(inputs=inputs, outputs=outputs) #=[out1, out2, out3]) model.compile(loss=ccc_loss, #{'v': ccc_loss, 'a': ccc_loss, 'd': ccc_loss}, loss_weights={'v': 0.3, 'a': 0.6, 'd': 0.1}, optimizer='rmsprop', metrics=[ccc, 'mse']) return model model2 = api_model() model2.summary() earlystop = EarlyStopping(monitor='val_loss', mode='min', patience=10, restore_best_weights=True) hist = model2.fit(X_train, y_train.T.tolist(), batch_size=64, validation_split=0.2, epochs=50, verbose=1, shuffle=True, callbacks=[earlystop]) metrik = model2.evaluate(X_test, y_test.T.tolist()) print('CCC= ', np.array(metrik)[[-6,-4,-2]]) print('MSE= ', np.array(metrik)[[-5,-3,-1]])

py

ccc_mse_ser

ccc_mse_ser-master/code/ser_improv_paa_mse.py

# ser_improv_paa_ccc.py # speech emotion recognition for MSP-IMPROV dataset with pyAudioAnalysis # HFS features using CCC-based loss function # coded by Bagus Tris Atmaja (bagus@ep.its.ac.id) # changelog: # 2020-02-13: Inital code, modified from deepMLP repo import numpy as np import pickle import pandas as pd import keras.backend as K from keras.models import Model from keras.layers import Input, Dense, CuDNNLSTM, Flatten, \ Embedding, Dropout, BatchNormalization, \ RNN, concatenate, Activation from keras.callbacks import EarlyStopping from sklearn.preprocessing import StandardScaler, MinMaxScaler from keras.preprocessing.text import Tokenizer from keras.preprocessing import sequence import random as rn import tensorflow as tf rn.seed(123) np.random.seed(99) tf.set_random_seed(1234) # loading file and label feat_train = np.load('/home/s1820002/ccc_mse/data/feat_hfs_paa_msp_train.npy') feat_test = np.load('/home/s1820002/ccc_mse/data/feat_hfs_paa_msp_test.npy') feat = np.vstack([feat_train, feat_test]) list_path = '/home/s1820002/msp-improv/helper/improv_data.csv' list_file = pd.read_csv(list_path, index_col=None) list_file = pd.DataFrame(list_file) data = list_file.sort_values(by=['wavfile']) vad_train = [] vad_test = [] for index, row in data.iterrows(): #print(row['wavfile'], row['v'], row['a'], row['d']) if int(row['wavfile'][18]) in range(1,6): #print("Process vad..", row['wavfile']) vad_train.append([row['v'], row['a'], row['d']]) else: #print("Process..", row['wavfile']) vad_test.append([row['v'], row['a'], row['d']]) vad = np.vstack([vad_train, vad_test]) # standardization scaled_feature = False if scaled_feature: scaler = StandardScaler() scaler = scaler.fit(feat) scaled_feat = scaler.transform(feat) feat = scaled_feat else: feat = feat scaled_vad = True # standardization if scaled_vad: scaler = MinMaxScaler(feature_range=(-1, 1)) scaler = scaler.fit(vad) #.reshape(vad.shape[0]*vad.shape[1], vad.shape[2])) scaled_vad = scaler.transform(vad) #.reshape(vad.shape[0]*vad.shape[1], vad.shape[2])) vad = scaled_vad else: vad = vad # reshape feat size to match LSTM config feat = feat.reshape(feat.shape[0], 1, feat.shape[1]) # train/test split, LOSO X_train = feat[:len(feat_train)] X_test = feat[len(feat_train):] y_train = vad[:len(vad_train)] y_test = vad[len(vad_train):] # Concordance correlation coefficient (CCC)-based loss function - using non-inductive statistics def ccc(gold, pred): gold = K.squeeze(gold, axis=-1) pred = K.squeeze(pred, axis=-1) gold_mean = K.mean(gold, axis=-1, keepdims=True) pred_mean = K.mean(pred, axis=-1, keepdims=True) covariance = (gold-gold_mean)*(pred-pred_mean) gold_var = K.mean(K.square(gold-gold_mean), axis=-1, keepdims=True) pred_var = K.mean(K.square(pred-pred_mean), axis=-1, keepdims=True) ccc = K.constant(2.) * covariance / (gold_var + pred_var + K.square(gold_mean - pred_mean) + K.common.epsilon()) return ccc def ccc_loss(gold, pred): # input (num_batches, seq_len, 1) ccc_loss = K.constant(1.) - ccc(gold, pred) return ccc_loss # API model, if use RNN, first two rnn layer must return_sequences=True def api_model(): inputs = Input(shape=(feat.shape[1], feat.shape[2]), name='feat_input') net = BatchNormalization()(inputs) #net = Bidirectional(LSTM(64, return_sequences=True, dropout=do, recurrent_dropout=do))(net) net = CuDNNLSTM(feat.shape[2], return_sequences=True)(net) net = CuDNNLSTM(256, return_sequences=True)(net) net = CuDNNLSTM(256, return_sequences=False)(net) net = Dense(64)(net) #net = Dropout(0.1)(net) target_names = ('v', 'a', 'd') outputs = [Dense(1, name=name, activation='tanh')(net) for name in target_names] model = Model(inputs=inputs, outputs=outputs) #=[out1, out2, out3]) model.compile(loss='mse', #{'v': ccc_loss, 'a': ccc_loss, 'd': ccc_loss}, loss_weights={'v': 0.3, 'a': 0.6, 'd': 0.1}, optimizer='rmsprop', metrics=[ccc, 'mse']) return model model2 = api_model() model2.summary() earlystop = EarlyStopping(monitor='val_loss', mode='min', patience=10, restore_best_weights=True) hist = model2.fit(X_train, y_train.T.tolist(), batch_size=64, validation_split=0.2, epochs=50, verbose=1, shuffle=True, callbacks=[earlystop]) metrik = model2.evaluate(X_test, y_test.T.tolist()) print('CCC= ', np.array(metrik)[[-6,-4,-2]]) print('MSE= ', np.array(metrik)[[-5,-3,-1]])

py

ccc_mse_ser

ccc_mse_ser-master/code/ser_improv_gemaps_ccc.py

# ser_improv_paa_ccc.py # speech emotion recognition for MSP-IMPROV dataset with pyAudioAnalysis # HFS features using CCC-based loss function # coded by Bagus Tris Atmaja (bagus@ep.its.ac.id) # changelog: # 2020-02-13: Inital code, modified from deepMLP repo import numpy as np import pickle import pandas as pd import keras.backend as K from keras.models import Model from keras.layers import Input, Dense, CuDNNLSTM, Flatten, \ Embedding, Dropout, BatchNormalization, \ RNN, concatenate, Activation from keras.callbacks import EarlyStopping from sklearn.preprocessing import StandardScaler, MinMaxScaler from keras.preprocessing.text import Tokenizer from keras.preprocessing import sequence import random as rn import tensorflow as tf rn.seed(123) np.random.seed(99) tf.set_random_seed(1234) # loading file and label feat_train = np.load('/home/s1820002/ccc_mse/data/feat_hfs_gemaps_msp_train.npy') feat_test = np.load('/home/s1820002/ccc_mse/data/feat_hfs_gemaps_msp_test.npy') feat = np.vstack([feat_train, feat_test]) list_path = '/home/s1820002/msp-improv/helper/improv_data.csv' list_file = pd.read_csv(list_path, index_col=None) list_file = pd.DataFrame(list_file) data = list_file.sort_values(by=['wavfile']) vad_train = [] vad_test = [] for index, row in data.iterrows(): #print(row['wavfile'], row['v'], row['a'], row['d']) if int(row['wavfile'][18]) in range(1,6): #print("Process vad..", row['wavfile']) vad_train.append([row['v'], row['a'], row['d']]) else: #print("Process..", row['wavfile']) vad_test.append([row['v'], row['a'], row['d']]) vad = np.vstack([vad_train, vad_test]) # standardization scaled_feature = True if scaled_feature: scaler = StandardScaler() scaler = scaler.fit(feat) scaled_feat = scaler.transform(feat) feat = scaled_feat else: feat = feat scaled_vad = True # standardization if scaled_vad: scaler = MinMaxScaler(feature_range=(-1, 1)) scaler = scaler.fit(vad) #.reshape(vad.shape[0]*vad.shape[1], vad.shape[2])) scaled_vad = scaler.transform(vad) #.reshape(vad.shape[0]*vad.shape[1], vad.shape[2])) vad = scaled_vad else: vad = vad # reshape feat size to match LSTM config feat = feat.reshape(feat.shape[0], 1, feat.shape[1]) # train/test split, LOSO X_train = feat[:len(feat_train)] X_test = feat[len(feat_train):] y_train = vad[:len(vad_train)] y_test = vad[len(vad_train):] # Concordance correlation coefficient (CCC)-based loss function - using non-inductive statistics def ccc(gold, pred): gold = K.squeeze(gold, axis=-1) pred = K.squeeze(pred, axis=-1) gold_mean = K.mean(gold, axis=-1, keepdims=True) pred_mean = K.mean(pred, axis=-1, keepdims=True) covariance = (gold-gold_mean)*(pred-pred_mean) gold_var = K.mean(K.square(gold-gold_mean), axis=-1, keepdims=True) pred_var = K.mean(K.square(pred-pred_mean), axis=-1, keepdims=True) ccc = K.constant(2.) * covariance / (gold_var + pred_var + K.square(gold_mean - pred_mean) + K.common.epsilon()) return ccc def ccc_loss(gold, pred): # input (num_batches, seq_len, 1) ccc_loss = K.constant(1.) - ccc(gold, pred) return ccc_loss # API model, if use RNN, first two rnn layer must return_sequences=True def api_model(): inputs = Input(shape=(feat.shape[1], feat.shape[2]), name='feat_input') net = BatchNormalization()(inputs) net = CuDNNLSTM(feat.shape[2], return_sequences=True)(net) net = CuDNNLSTM(256, return_sequences=True)(net) net = CuDNNLSTM(256, return_sequences=False)(net) net = Dense(64)(net) target_names = ('v', 'a', 'd') outputs = [Dense(1, name=name, activation='tanh')(net) for name in target_names] model = Model(inputs=inputs, outputs=outputs) #=[out1, out2, out3]) model.compile(loss=ccc_loss, #{'v': ccc_loss, 'a': ccc_loss, 'd': ccc_loss}, loss_weights={'v': 0.3, 'a': 0.6, 'd': 0.1}, optimizer='rmsprop', metrics=[ccc, 'mse']) return model model2 = api_model() model2.summary() earlystop = EarlyStopping(monitor='val_loss', mode='min', patience=10, restore_best_weights=True) hist = model2.fit(X_train, y_train.T.tolist(), batch_size=64, validation_split=0.2, epochs=50, verbose=1, shuffle=True, callbacks=[earlystop]) metrik = model2.evaluate(X_test, y_test.T.tolist()) print('CCC= ', np.array(metrik)[[-6,-4,-2]]) print('MSE= ', np.array(metrik)[[-5,-3,-1]])

py

ccc_mse_ser

ccc_mse_ser-master/code/ser_iemocap_paa_mse.py

# CSL Paper: Dimensional speech emotion recognition from acoustic and text # Changelog: # 2019-09-01: initial version # 2019-10-06: optimizer MTL parameters with linear search (in progress) # 2012-12-25: modified fot ser_iemocap_loso_hfs.py # feature is either std+mean or std+mean+silence (uncomment line 44) import numpy as np import pickle import pandas as pd import keras.backend as K from keras.models import Model from keras.layers import Input, Dense, CuDNNLSTM, Flatten, \ Embedding, Dropout, BatchNormalization, \ RNN, concatenate, Activation from keras.callbacks import EarlyStopping from sklearn.preprocessing import StandardScaler, MinMaxScaler from keras.preprocessing.text import Tokenizer from keras.preprocessing import sequence import random as rn import tensorflow as tf rn.seed(123) np.random.seed(99) tf.set_random_seed(1234) # load feature and labels feat = np.load('/home/s1820002/atsit/data/feat_34_hfs.npy') vad = np.load('/home/s1820002/IEMOCAP-Emotion-Detection/y_egemaps.npy') # for LSTM input shape (batch, steps, features/channel) #feat = feat.reshape(feat.shape[0], 1, feat.shape[1]) # remove outlier, < 1, > 5 vad = np.where(vad==5.5, 5.0, vad) vad = np.where(vad==0.5, 1.0, vad) # standardization scaled_feature = False # set Dropout do = 0.3 if scaled_feature == True: scaler = StandardScaler() scaler = scaler.fit(feat.reshape(feat.shape[0]*feat.shape[1], feat.shape[2])) scaled_feat = scaler.transform(feat.reshape(feat.shape[0]*feat.shape[1], feat.shape[2])) scaled_feat = scaled_feat.reshape(feat.shape[0], feat.shape[1], feat.shape[2]) feat = scaled_feat else: feat = feat scaled_vad = True # standardization if scaled_vad: scaler = MinMaxScaler(feature_range=(-1, 1)) scaler = scaler.fit(vad) #.reshape(vad.shape[0]*vad.shape[1], vad.shape[2])) scaled_vad = scaler.transform(vad) #.reshape(vad.shape[0]*vad.shape[1], vad.shape[2])) vad = scaled_vad else: vad = vad # Concordance correlation coefficient (CCC)-based loss function - using non-inductive statistics def ccc(gold, pred): gold = K.squeeze(gold, axis=-1) pred = K.squeeze(pred, axis=-1) gold_mean = K.mean(gold, axis=-1, keepdims=True) pred_mean = K.mean(pred, axis=-1, keepdims=True) covariance = (gold-gold_mean)*(pred-pred_mean) gold_var = K.mean(K.square(gold-gold_mean), axis=-1, keepdims=True) pred_var = K.mean(K.square(pred-pred_mean), axis=-1, keepdims=True) ccc = K.constant(2.) * covariance / (gold_var + pred_var + K.square(gold_mean - pred_mean) + K.common.epsilon()) return ccc def ccc_loss(gold, pred): # input (num_batches, seq_len, 1) ccc_loss = K.constant(1.) - ccc(gold, pred) return ccc_loss # API model, if use RNN, first two rnn layer must return_sequences=True def api_model(alpha, beta, gamma): # speech network input_speech = Input(shape=(feat.shape[1], feat.shape[2]), name='speech_input') net_speech = BatchNormalization()(input_speech) net_speech = CuDNNLSTM(feat.shape[2], return_sequences=True)(net_speech) net_speech = CuDNNLSTM(256, return_sequences=True)(net_speech) net_speech = CuDNNLSTM(256, return_sequences=False)(net_speech) #net_speech = Flatten()(net_speech) net_speech = Dense(64)(net_speech) #net_speech = Dropout(0.1)(net_speech) target_names = ('v', 'a', 'd') model_combined = [Dense(1, name=name, activation='tanh')(net_speech) for name in target_names] model = Model(input_speech, model_combined) #model.compile(loss=ccc_loss, optimizer='rmsprop', metrics=[ccc]) model.compile(loss='mse', loss_weights={'v': alpha, 'a': beta, 'd': gamma}, optimizer='rmsprop', metrics=[ccc, 'mse']) return model #def main(alpha, beta, gamma): model = api_model(0.1, 0.5, 0.4) model.summary() # 7869 first data of session 5 (for LOSO) earlystop = EarlyStopping(monitor='val_loss', mode='min', patience=10, restore_best_weights=True) hist = model.fit(feat[:7869], vad[:7869].T.tolist(), batch_size=64, #best:8 validation_split=0.2, epochs=200, verbose=1, shuffle=True, callbacks=[earlystop]) metrik = model.evaluate(feat[7869:], vad[7869:].T.tolist()) print(metrik) print('CCC= ', np.array(metrik)[[-6,-4,-2]]) print('MSE= ', np.array(metrik)[[-5,-3,-1]])

py

LDU

LDU-main/monocular_depth_estimation/pytorch/distributed_sampler_no_evenly_divisible.py

import math import torch from torch.utils.data import Sampler import torch.distributed as dist class DistributedSamplerNoEvenlyDivisible(Sampler): """Sampler that restricts data loading to a subset of the dataset. It is especially useful in conjunction with :class:`torch.nn.parallel.DistributedDataParallel`. In such case, each process can pass a DistributedSampler instance as a DataLoader sampler, and load a subset of the original dataset that is exclusive to it. .. note:: Dataset is assumed to be of constant size. Arguments: dataset: Dataset used for sampling. num_replicas (optional): Number of processes participating in distributed training. rank (optional): Rank of the current process within num_replicas. shuffle (optional): If true (default), sampler will shuffle the indices """ def __init__(self, dataset, num_replicas=None, rank=None, shuffle=True): if num_replicas is None: if not dist.is_available(): raise RuntimeError("Requires distributed package to be available") num_replicas = dist.get_world_size() if rank is None: if not dist.is_available(): raise RuntimeError("Requires distributed package to be available") rank = dist.get_rank() self.dataset = dataset self.num_replicas = num_replicas self.rank = rank self.epoch = 0 num_samples = int(math.floor(len(self.dataset) * 1.0 / self.num_replicas)) rest = len(self.dataset) - num_samples * self.num_replicas if self.rank < rest: num_samples += 1 self.num_samples = num_samples self.total_size = len(dataset) # self.total_size = self.num_samples * self.num_replicas self.shuffle = shuffle def __iter__(self): # deterministically shuffle based on epoch g = torch.Generator() g.manual_seed(self.epoch) if self.shuffle: indices = torch.randperm(len(self.dataset), generator=g).tolist() else: indices = list(range(len(self.dataset))) # add extra samples to make it evenly divisible # indices += indices[:(self.total_size - len(indices))] # assert len(indices) == self.total_size # subsample indices = indices[self.rank:self.total_size:self.num_replicas] self.num_samples = len(indices) # assert len(indices) == self.num_samples return iter(indices) def __len__(self): return self.num_samples def set_epoch(self, epoch): self.epoch = epoch

py

LDU

LDU-main/monocular_depth_estimation/pytorch/bts_live_3d.py

# Copyright (C) 2019 Jin Han Lee # # This file is a part of BTS. # This program is free software: you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation, either version 3 of the License, or # (at your option) any later version. # # This program is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with this program. If not, see <http://www.gnu.org/licenses/> from __future__ import absolute_import, division, print_function import os os.environ['TF_CPP_MIN_LOG_LEVEL'] = '1' import os import sys import time import argparse import numpy as np # Computer Vision import cv2 from scipy import ndimage from skimage.transform import resize # Visualization import matplotlib.pyplot as plt plasma = plt.get_cmap('plasma') greys = plt.get_cmap('Greys') # UI and OpenGL from PySide2 import QtCore, QtGui, QtWidgets, QtOpenGL from OpenGL import GL, GLU from OpenGL.arrays import vbo from OpenGL.GL import shaders import glm # Argument Parser parser = argparse.ArgumentParser(description='BTS Live 3D') parser.add_argument('--model_name', type=str, help='model name', default='bts_nyu_v2') parser.add_argument('--encoder', type=str, help='type of encoder, densenet121_bts or densenet161_bts', default='densenet161_bts') parser.add_argument('--max_depth', type=float, help='maximum depth in estimation', default=10) parser.add_argument('--checkpoint_path', type=str, help='path to a checkpoint to load', required=True) parser.add_argument('--input_height', type=int, help='input height', default=480) parser.add_argument('--input_width', type=int, help='input width', default=640) parser.add_argument('--dataset', type=str, help='dataset this model trained on', default='nyu') args = parser.parse_args() model_dir = os.path.join("./models", args.model_name) sys.path.append(model_dir) for key, val in vars(__import__(args.model_name)).items(): if key.startswith('__') and key.endswith('__'): continue vars()[key] = val # Image shapes height_rgb, width_rgb = 480, 640 height_depth, width_depth = height_rgb, width_rgb height_rgb = height_rgb import torch import torch.nn as nn import torch.backends.cudnn as cudnn from torch.autograd import Variable # Intrinsic parameters for your own webcam camera camera_matrix = np.zeros(shape=(3, 3)) camera_matrix[0, 0] = 5.4765313594010649e+02 camera_matrix[0, 2] = 3.2516069906172453e+02 camera_matrix[1, 1] = 5.4801781476172562e+02 camera_matrix[1, 2] = 2.4794113960783835e+02 camera_matrix[2, 2] = 1 dist_coeffs = np.array([ 3.7230261423972011e-02, -1.6171708069773008e-01, -3.5260752900266357e-04, 1.7161234226767313e-04, 1.0192711400840315e-01 ]) # Parameters for a model trained on NYU Depth V2 new_camera_matrix = np.zeros(shape=(3, 3)) new_camera_matrix[0, 0] = 518.8579 new_camera_matrix[0, 2] = 320 new_camera_matrix[1, 1] = 518.8579 new_camera_matrix[1, 2] = 240 new_camera_matrix[2, 2] = 1 R = np.identity(3, dtype=np.float) map1, map2 = cv2.initUndistortRectifyMap(camera_matrix, dist_coeffs, R, new_camera_matrix, (640, 480), cv2.CV_32FC1) def load_model(): args.mode = 'test' model = BtsModel(params=args) model = torch.nn.DataParallel(model) checkpoint = torch.load(args.checkpoint_path) model.load_state_dict(checkpoint['model']) model.eval() model.cuda() return model # Function timing ticTime = time.time() def tic(): global ticTime; ticTime = time.time() def toc(): print('{0} seconds.'.format(time.time() - ticTime)) # Conversion from Numpy to QImage and back def np_to_qimage(a): im = a.copy() return QtGui.QImage(im.data, im.shape[1], im.shape[0], im.strides[0], QtGui.QImage.Format_RGB888).copy() def qimage_to_np(img): img = img.convertToFormat(QtGui.QImage.Format.Format_ARGB32) return np.array(img.constBits()).reshape(img.height(), img.width(), 4) # Compute edge magnitudes def edges(d): dx = ndimage.sobel(d, 0) # horizontal derivative dy = ndimage.sobel(d, 1) # vertical derivative return np.abs(dx) + np.abs(dy) # Main window class Window(QtWidgets.QWidget): updateInput = QtCore.Signal() def __init__(self, parent=None): QtWidgets.QWidget.__init__(self, parent) self.model = None self.capture = None self.glWidget = GLWidget() mainLayout = QtWidgets.QVBoxLayout() # Input / output views viewsLayout = QtWidgets.QGridLayout() self.inputViewer = QtWidgets.QLabel("[Click to start]") self.inputViewer.setPixmap(QtGui.QPixmap(width_rgb, height_rgb)) self.outputViewer = QtWidgets.QLabel("[Click to start]") self.outputViewer.setPixmap(QtGui.QPixmap(width_rgb, height_rgb)) imgsFrame = QtWidgets.QFrame() inputsLayout = QtWidgets.QVBoxLayout() imgsFrame.setLayout(inputsLayout) inputsLayout.addWidget(self.inputViewer) inputsLayout.addWidget(self.outputViewer) viewsLayout.addWidget(imgsFrame, 0, 0) viewsLayout.addWidget(self.glWidget, 0, 1) viewsLayout.setColumnStretch(1, 10) mainLayout.addLayout(viewsLayout) # Load depth estimation model toolsLayout = QtWidgets.QHBoxLayout() self.button2 = QtWidgets.QPushButton("Webcam") self.button2.clicked.connect(self.loadCamera) toolsLayout.addWidget(self.button2) self.button4 = QtWidgets.QPushButton("Pause") self.button4.clicked.connect(self.loadImage) toolsLayout.addWidget(self.button4) self.button6 = QtWidgets.QPushButton("Refresh") self.button6.clicked.connect(self.updateCloud) toolsLayout.addWidget(self.button6) mainLayout.addLayout(toolsLayout) self.setLayout(mainLayout) self.setWindowTitle(self.tr("BTS Live")) # Signals self.updateInput.connect(self.update_input) # Default example if self.glWidget.rgb.any() and self.glWidget.depth.any(): img = (self.glWidget.rgb * 255).astype('uint8') self.inputViewer.setPixmap(QtGui.QPixmap.fromImage(np_to_qimage(img))) coloredDepth = (plasma(self.glWidget.depth[:, :, 0])[:, :, :3] * 255).astype('uint8') self.outputViewer.setPixmap(QtGui.QPixmap.fromImage(np_to_qimage(coloredDepth))) def loadModel(self): QtGui.QGuiApplication.setOverrideCursor(QtCore.Qt.WaitCursor) tic() self.model = load_model() print('Model loaded.') toc() self.updateCloud() QtGui.QGuiApplication.restoreOverrideCursor() def loadCamera(self): tic() self.model = load_model() print('Model loaded.') toc() self.capture = cv2.VideoCapture(0) self.updateInput.emit() def loadVideoFile(self): self.capture = cv2.VideoCapture('video.mp4') self.updateInput.emit() def loadImage(self): self.capture = None img = (self.glWidget.rgb * 255).astype('uint8') self.inputViewer.setPixmap(QtGui.QPixmap.fromImage(np_to_qimage(img))) self.updateCloud() def loadImageFile(self): self.capture = None filename = \ QtWidgets.QFileDialog.getOpenFileName(None, 'Select image', '', self.tr('Image files (*.jpg *.png)'))[0] img = QtGui.QImage(filename).scaledToHeight(height_rgb) xstart = 0 if img.width() > width_rgb: xstart = (img.width() - width_rgb) // 2 img = img.copy(xstart, 0, xstart + width_rgb, height_rgb) self.inputViewer.setPixmap(QtGui.QPixmap.fromImage(img)) self.updateCloud() def update_input(self): # Don't update anymore if no capture device is set if self.capture == None: return # Capture a frame ret, frame = self.capture.read() # Loop video playback if current stream is video file if not ret: self.capture.set(cv2.CAP_PROP_POS_FRAMES, 0) ret, frame = self.capture.read() # Prepare image and show in UI frame_ud = cv2.remap(frame, map1, map2, interpolation=cv2.INTER_LINEAR) frame = cv2.cvtColor(frame_ud, cv2.COLOR_BGR2RGB) image = np_to_qimage(frame) self.inputViewer.setPixmap(QtGui.QPixmap.fromImage(image)) # Update the point cloud self.updateCloud() def updateCloud(self): rgb8 = qimage_to_np(self.inputViewer.pixmap().toImage()) self.glWidget.rgb = (rgb8[:, :, :3] / 255)[:, :, ::-1] if self.model: input_image = rgb8[:, :, :3].astype(np.float32) # Normalize image input_image[:, :, 0] = (input_image[:, :, 0] - 123.68) * 0.017 input_image[:, :, 1] = (input_image[:, :, 1] - 116.78) * 0.017 input_image[:, :, 2] = (input_image[:, :, 2] - 103.94) * 0.017 input_image_cropped = input_image[32:-1 - 31, 32:-1 - 31, :] input_images = np.expand_dims(input_image_cropped, axis=0) input_images = np.transpose(input_images, (0, 3, 1, 2)) with torch.no_grad(): image = Variable(torch.from_numpy(input_images)).cuda() focal = Variable(torch.tensor([518.8579])).cuda() # Predict lpg8x8, lpg4x4, lpg2x2, reduc1x1, depth_cropped = self.model(image, focal) depth = np.zeros((480, 640), dtype=np.float32) depth[32:-1-31, 32:-1-31] = depth_cropped[0].cpu().squeeze() / args.max_depth coloredDepth = (greys(np.log10(depth * args.max_depth))[:, :, :3] * 255).astype('uint8') self.outputViewer.setPixmap(QtGui.QPixmap.fromImage(np_to_qimage(coloredDepth))) self.glWidget.depth = depth else: self.glWidget.depth = 0.5 + np.zeros((height_rgb // 2, width_rgb // 2, 1)) self.glWidget.updateRGBD() self.glWidget.updateGL() # Update to next frame if we are live QtCore.QTimer.singleShot(10, self.updateInput) class GLWidget(QtOpenGL.QGLWidget): def __init__(self, parent=None): QtOpenGL.QGLWidget.__init__(self, parent) self.object = 0 self.xRot = 5040 self.yRot = 40 self.zRot = 0 self.zoomLevel = 9 self.lastPos = QtCore.QPoint() self.green = QtGui.QColor.fromCmykF(0.0, 0.0, 0.0, 1.0) self.black = QtGui.QColor.fromCmykF(0.0, 0.0, 0.0, 1.0) # Precompute for world coordinates self.xx, self.yy = self.worldCoords(width=width_rgb, height=height_rgb) self.rgb = np.zeros((480, 640, 3), dtype=np.uint8) self.depth = np.zeros((480, 640), dtype=np.float32) self.col_vbo = None self.pos_vbo = None if self.rgb.any() and self.detph.any(): self.updateRGBD() def xRotation(self): return self.xRot def yRotation(self): return self.yRot def zRotation(self): return self.zRot def minimumSizeHint(self): return QtCore.QSize(640, 480) def sizeHint(self): return QtCore.QSize(640, 480) def setXRotation(self, angle): if angle != self.xRot: self.xRot = angle self.emit(QtCore.SIGNAL("xRotationChanged(int)"), angle) self.updateGL() def setYRotation(self, angle): if angle != self.yRot: self.yRot = angle self.emit(QtCore.SIGNAL("yRotationChanged(int)"), angle) self.updateGL() def setZRotation(self, angle): if angle != self.zRot: self.zRot = angle self.emit(QtCore.SIGNAL("zRotationChanged(int)"), angle) self.updateGL() def resizeGL(self, width, height): GL.glViewport(0, 0, width, height) def mousePressEvent(self, event): self.lastPos = QtCore.QPoint(event.pos()) def mouseMoveEvent(self, event): dx = -(event.x() - self.lastPos.x()) dy = (event.y() - self.lastPos.y()) if event.buttons() & QtCore.Qt.LeftButton: self.setXRotation(self.xRot + dy) self.setYRotation(self.yRot + dx) elif event.buttons() & QtCore.Qt.RightButton: self.setXRotation(self.xRot + dy) self.setZRotation(self.zRot + dx) self.lastPos = QtCore.QPoint(event.pos()) def wheelEvent(self, event): numDegrees = event.delta() / 8 numSteps = numDegrees / 15 self.zoomLevel = self.zoomLevel + numSteps event.accept() self.updateGL() def initializeGL(self): self.qglClearColor(self.black.darker()) GL.glShadeModel(GL.GL_FLAT) GL.glEnable(GL.GL_DEPTH_TEST) GL.glEnable(GL.GL_CULL_FACE) VERTEX_SHADER = shaders.compileShader("""#version 330 layout(location = 0) in vec3 position; layout(location = 1) in vec3 color; uniform mat4 mvp; out vec4 frag_color; void main() {gl_Position = mvp * vec4(position, 1.0);frag_color = vec4(color, 1.0);}""", GL.GL_VERTEX_SHADER) FRAGMENT_SHADER = shaders.compileShader("""#version 330 in vec4 frag_color; out vec4 out_color; void main() {out_color = frag_color;}""", GL.GL_FRAGMENT_SHADER) self.shaderProgram = shaders.compileProgram(VERTEX_SHADER, FRAGMENT_SHADER) self.UNIFORM_LOCATIONS = { 'position': GL.glGetAttribLocation(self.shaderProgram, 'position'), 'color': GL.glGetAttribLocation(self.shaderProgram, 'color'), 'mvp': GL.glGetUniformLocation(self.shaderProgram, 'mvp'), } shaders.glUseProgram(self.shaderProgram) def paintGL(self): if self.rgb.any() and self.depth.any(): GL.glClear(GL.GL_COLOR_BUFFER_BIT | GL.GL_DEPTH_BUFFER_BIT) self.drawObject() def worldCoords(self, width, height): cx, cy = width / 2, height / 2 fx = 518.8579 fy = 518.8579 xx, yy = np.tile(range(width), height), np.repeat(range(height), width) xx = (xx - cx) / fx yy = (yy - cy) / fy return xx, yy def posFromDepth(self, depth): length = depth.shape[0] * depth.shape[1] depth[edges(depth) > 0.3] = 1e6 # Hide depth edges z = depth.reshape(length) return np.dstack((self.xx * z, self.yy * z, z)).reshape((length, 3)) def createPointCloudVBOfromRGBD(self): # Create position and color VBOs self.pos_vbo = vbo.VBO(data=self.pos, usage=GL.GL_DYNAMIC_DRAW, target=GL.GL_ARRAY_BUFFER) self.col_vbo = vbo.VBO(data=self.col, usage=GL.GL_DYNAMIC_DRAW, target=GL.GL_ARRAY_BUFFER) def updateRGBD(self): # RGBD dimensions width, height = self.depth.shape[1], self.depth.shape[0] # Reshape points = self.posFromDepth(self.depth.copy()) colors = resize(self.rgb, (height, width)).reshape((height * width, 3)) # Flatten and convert to float32 self.pos = points.astype('float32') self.col = colors.reshape(height * width, 3).astype('float32') # Move center of scene self.pos = self.pos + glm.vec3(0, -0.06, -0.3) # Create VBOs if not self.col_vbo: self.createPointCloudVBOfromRGBD() def drawObject(self): # Update camera model, view, proj = glm.mat4(1), glm.mat4(1), glm.perspective(45, self.width() / self.height(), 0.01, 100) center, up, eye = glm.vec3(0, -0.075, 0), glm.vec3(0, -1, 0), glm.vec3(0, 0, -0.4 * (self.zoomLevel / 10)) view = glm.lookAt(eye, center, up) model = glm.rotate(model, self.xRot / 160.0, glm.vec3(1, 0, 0)) model = glm.rotate(model, self.yRot / 160.0, glm.vec3(0, 1, 0)) model = glm.rotate(model, self.zRot / 160.0, glm.vec3(0, 0, 1)) mvp = proj * view * model GL.glUniformMatrix4fv(self.UNIFORM_LOCATIONS['mvp'], 1, False, glm.value_ptr(mvp)) # Update data self.pos_vbo.set_array(self.pos) self.col_vbo.set_array(self.col) # Point size GL.glPointSize(2) # Position self.pos_vbo.bind() GL.glEnableVertexAttribArray(0) GL.glVertexAttribPointer(0, 3, GL.GL_FLOAT, GL.GL_FALSE, 0, None) # Color self.col_vbo.bind() GL.glEnableVertexAttribArray(1) GL.glVertexAttribPointer(1, 3, GL.GL_FLOAT, GL.GL_FALSE, 0, None) # Draw GL.glDrawArrays(GL.GL_POINTS, 0, self.pos.shape[0]) if __name__ == '__main__': app = QtWidgets.QApplication(sys.argv) window = Window() window.show() res = app.exec_()

py

LDU

LDU-main/monocular_depth_estimation/pytorch/bts_main.py

# Copyright (C) 2019 Jin Han Lee # # This file is a part of BTS. # This program is free software: you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation, either version 3 of the License, or # (at your option) any later version. # # This program is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with this program. If not, see <http://www.gnu.org/licenses/> import time import argparse import sys import os import torch import torch.backends.cudnn as cudnn import torch.distributed as dist import torch.multiprocessing as mp from tensorboardX import SummaryWriter import matplotlib import matplotlib.cm from tqdm import tqdm from bts_dataloader import * from bts_ldu import * def convert_arg_line_to_args(arg_line): for arg in arg_line.split(): if not arg.strip(): continue yield arg parser = argparse.ArgumentParser(description='BTS PyTorch implementation.', fromfile_prefix_chars='@') parser.convert_arg_line_to_args = convert_arg_line_to_args parser.add_argument('--mode', type=str, help='train or test', default='train') parser.add_argument('--model_name', type=str, help='model name', default='bts_eigen_v2') parser.add_argument('--encoder', type=str, help='type of encoder, desenet121_bts, densenet161_bts, ' 'resnet101_bts, resnet50_bts, resnext50_bts or resnext101_bts', default='densenet161_bts') # Dataset parser.add_argument('--dataset', type=str, help='dataset to train on, kitti or nyu', default='nyu') parser.add_argument('--data_path', type=str, help='path to the data', required=True) parser.add_argument('--gt_path', type=str, help='path to the groundtruth data', required=True) parser.add_argument('--filenames_file', type=str, help='path to the filenames text file', required=True) parser.add_argument('--input_height', type=int, help='input height', default=480) parser.add_argument('--input_width', type=int, help='input width', default=640) parser.add_argument('--max_depth', type=float, help='maximum depth in estimation', default=10) # Log and save parser.add_argument('--log_directory', type=str, help='directory to save checkpoints and summaries', default='') parser.add_argument('--checkpoint_path', type=str, help='path to a checkpoint to load', default='') parser.add_argument('--log_freq', type=int, help='Logging frequency in global steps', default=100) parser.add_argument('--save_freq', type=int, help='Checkpoint saving frequency in global steps', default=500) # Training parser.add_argument('--fix_first_conv_blocks', help='if set, will fix the first two conv blocks', action='store_true') parser.add_argument('--fix_first_conv_block', help='if set, will fix the first conv block', action='store_true') parser.add_argument('--bn_no_track_stats', help='if set, will not track running stats in batch norm layers', action='store_true') parser.add_argument('--weight_decay', type=float, help='weight decay factor for optimization', default=1e-2) parser.add_argument('--bts_size', type=int, help='initial num_filters in bts', default=512) parser.add_argument('--retrain', help='if used with checkpoint_path, will restart training from step zero', action='store_true') parser.add_argument('--adam_eps', type=float, help='epsilon in Adam optimizer', default=1e-6) parser.add_argument('--batch_size', type=int, help='batch size', default=4) parser.add_argument('--num_epochs', type=int, help='number of epochs', default=50) parser.add_argument('--learning_rate', type=float, help='initial learning rate', default=1e-4) parser.add_argument('--end_learning_rate', type=float, help='end learning rate', default=-1) parser.add_argument('--variance_focus', type=float, help='lambda in paper: [0, 1], higher value more focus on minimizing variance of error', default=0.85) parser.add_argument('--nb_proto', type=int, help='initial num_proto in bts', default=30) parser.add_argument('--loss_lambda', type=float, help='weight of the additional losses', default=0.1) # Preprocessing parser.add_argument('--do_random_rotate', help='if set, will perform random rotation for augmentation', action='store_true') parser.add_argument('--degree', type=float, help='random rotation maximum degree', default=2.5) parser.add_argument('--do_kb_crop', help='if set, crop input images as kitti benchmark images', action='store_true') parser.add_argument('--use_right', help='if set, will randomly use right images when train on KITTI', action='store_true') # Multi-gpu training parser.add_argument('--num_threads', type=int, help='number of threads to use for data loading', default=1) parser.add_argument('--world_size', type=int, help='number of nodes for distributed training', default=1) parser.add_argument('--rank', type=int, help='node rank for distributed training', default=0) parser.add_argument('--dist_url', type=str, help='url used to set up distributed training', default='tcp://127.0.0.1:1234') parser.add_argument('--dist_backend', type=str, help='distributed backend', default='nccl') parser.add_argument('--gpu', type=int, help='GPU id to use.', default=None) parser.add_argument('--multiprocessing_distributed', help='Use multi-processing distributed training to launch ' 'N processes per node, which has N GPUs. This is the ' 'fastest way to use PyTorch for either single node or ' 'multi node data parallel training', action='store_true',) # Online eval parser.add_argument('--do_online_eval', help='if set, perform online eval in every eval_freq steps', action='store_true') parser.add_argument('--data_path_eval', type=str, help='path to the data for online evaluation', required=False) parser.add_argument('--gt_path_eval', type=str, help='path to the groundtruth data for online evaluation', required=False) parser.add_argument('--filenames_file_eval', type=str, help='path to the filenames text file for online evaluation', required=False) parser.add_argument('--min_depth_eval', type=float, help='minimum depth for evaluation', default=1e-3) parser.add_argument('--max_depth_eval', type=float, help='maximum depth for evaluation', default=80) parser.add_argument('--eigen_crop', help='if set, crops according to Eigen NIPS14', action='store_true') parser.add_argument('--garg_crop', help='if set, crops according to Garg ECCV16', action='store_true') parser.add_argument('--eval_freq', type=int, help='Online evaluation frequency in global steps', default=500) parser.add_argument('--eval_summary_directory', type=str, help='output directory for eval summary,' 'if empty outputs to checkpoint folder', default='') if sys.argv.__len__() == 2: arg_filename_with_prefix = '@' + sys.argv[1] args = parser.parse_args([arg_filename_with_prefix]) else: args = parser.parse_args() inv_normalize = transforms.Normalize( mean=[-0.485/0.229, -0.456/0.224, -0.406/0.225], std=[1/0.229, 1/0.224, 1/0.225] ) eval_metrics = ['silog', 'abs_rel', 'log10', 'rms', 'sq_rel', 'log_rms', 'd1', 'd2', 'd3'] def compute_errors(gt, pred): thresh = np.maximum((gt / pred), (pred / gt)) d1 = (thresh < 1.25).mean() d2 = (thresh < 1.25 ** 2).mean() d3 = (thresh < 1.25 ** 3).mean() rms = (gt - pred) ** 2 rms = np.sqrt(rms.mean()) log_rms = (np.log(gt) - np.log(pred)) ** 2 log_rms = np.sqrt(log_rms.mean()) abs_rel = np.mean(np.abs(gt - pred) / gt) sq_rel = np.mean(((gt - pred) ** 2) / gt) err = np.log(pred) - np.log(gt) silog = np.sqrt(np.mean(err ** 2) - np.mean(err) ** 2) * 100 err = np.abs(np.log10(pred) - np.log10(gt)) log10 = np.mean(err) return [silog, abs_rel, log10, rms, sq_rel, log_rms, d1, d2, d3] def block_print(): sys.stdout = open(os.devnull, 'w') def enable_print(): sys.stdout = sys.__stdout__ def get_num_lines(file_path): f = open(file_path, 'r') lines = f.readlines() f.close() return len(lines) def colorize(value, vmin=None, vmax=None, cmap='Greys'): value = value.cpu().numpy()[:, :, :] value = np.log10(value) vmin = value.min() if vmin is None else vmin vmax = value.max() if vmax is None else vmax if vmin != vmax: value = (value - vmin) / (vmax - vmin) else: value = value*0. cmapper = matplotlib.cm.get_cmap(cmap) value = cmapper(value, bytes=True) img = value[:, :, :3] return img.transpose((2, 0, 1)) def normalize_result(value, vmin=None, vmax=None): value = value.cpu().numpy()[0, :, :] vmin = value.min() if vmin is None else vmin vmax = value.max() if vmax is None else vmax if vmin != vmax: value = (value - vmin) / (vmax - vmin) else: value = value * 0. return np.expand_dims(value, 0) def set_misc(model): if args.bn_no_track_stats: print("Disabling tracking running stats in batch norm layers") model.apply(bn_init_as_tf) if args.fix_first_conv_blocks: if 'resne' in args.encoder: fixing_layers = ['base_model.conv1', 'base_model.layer1.0', 'base_model.layer1.1', '.bn'] else: fixing_layers = ['conv0', 'denseblock1.denselayer1', 'denseblock1.denselayer2', 'norm'] print("Fixing first two conv blocks") elif args.fix_first_conv_block: if 'resne' in args.encoder: fixing_layers = ['base_model.conv1', 'base_model.layer1.0', '.bn'] else: fixing_layers = ['conv0', 'denseblock1.denselayer1', 'norm'] print("Fixing first conv block") else: if 'resne' in args.encoder: fixing_layers = ['base_model.conv1', '.bn'] else: fixing_layers = ['conv0', 'norm'] print("Fixing first conv layer") for name, child in model.named_children(): if not 'encoder' in name: continue for name2, parameters in child.named_parameters(): # print(name, name2) if any(x in name2 for x in fixing_layers): parameters.requires_grad = False def online_eval(model, dataloader_eval, gpu, ngpus): eval_measures = torch.zeros(10).cuda(device=gpu) for _, eval_sample_batched in enumerate(tqdm(dataloader_eval.data)): with torch.no_grad(): image = torch.autograd.Variable(eval_sample_batched['image'].cuda(gpu, non_blocking=True)) focal = torch.autograd.Variable(eval_sample_batched['focal'].cuda(gpu, non_blocking=True)) gt_depth = eval_sample_batched['depth'] has_valid_depth = eval_sample_batched['has_valid_depth'] if not has_valid_depth: continue pred_depth, _ = model(image, focal) pred_depth = pred_depth.cpu().numpy().squeeze() gt_depth = gt_depth.cpu().numpy().squeeze() if args.do_kb_crop: height, width = gt_depth.shape top_margin = int(height - 352) left_margin = int((width - 1216) / 2) pred_depth_uncropped = np.zeros((height, width), dtype=np.float32) pred_depth_uncropped[top_margin:top_margin + 352, left_margin:left_margin + 1216] = pred_depth pred_depth = pred_depth_uncropped pred_depth[pred_depth < args.min_depth_eval] = args.min_depth_eval pred_depth[pred_depth > args.max_depth_eval] = args.max_depth_eval pred_depth[np.isinf(pred_depth)] = args.max_depth_eval pred_depth[np.isnan(pred_depth)] = args.min_depth_eval valid_mask = np.logical_and(gt_depth > args.min_depth_eval, gt_depth < args.max_depth_eval) if args.garg_crop or args.eigen_crop: gt_height, gt_width = gt_depth.shape eval_mask = np.zeros(valid_mask.shape) if args.garg_crop: eval_mask[int(0.40810811 * gt_height):int(0.99189189 * gt_height), int(0.03594771 * gt_width):int(0.96405229 * gt_width)] = 1 elif args.eigen_crop: if args.dataset == 'kitti': eval_mask[int(0.3324324 * gt_height):int(0.91351351 * gt_height), int(0.0359477 * gt_width):int(0.96405229 * gt_width)] = 1 else: eval_mask[45:471, 41:601] = 1 valid_mask = np.logical_and(valid_mask, eval_mask) measures = compute_errors(gt_depth[valid_mask], pred_depth[valid_mask]) eval_measures[:9] += torch.tensor(measures).cuda(device=gpu) eval_measures[9] += 1 if args.multiprocessing_distributed: group = dist.new_group([i for i in range(ngpus)]) dist.all_reduce(tensor=eval_measures, op=dist.ReduceOp.SUM, group=group) if not args.multiprocessing_distributed or gpu == 0: eval_measures_cpu = eval_measures.cpu() cnt = eval_measures_cpu[9].item() eval_measures_cpu /= cnt print('Computing errors for {} eval samples'.format(int(cnt))) print("{:>7}, {:>7}, {:>7}, {:>7}, {:>7}, {:>7}, {:>7}, {:>7}, {:>7}".format('silog', 'abs_rel', 'log10', 'rms', 'sq_rel', 'log_rms', 'd1', 'd2', 'd3')) for i in range(8): print('{:7.3f}, '.format(eval_measures_cpu[i]), end='') print('{:7.3f}'.format(eval_measures_cpu[8])) return eval_measures_cpu return None def main_worker(gpu, ngpus_per_node, args): args.gpu = gpu if args.gpu is not None: print("Use GPU: {} for training".format(args.gpu)) if args.distributed: if args.dist_url == "env://" and args.rank == -1: args.rank = int(os.environ["RANK"]) if args.multiprocessing_distributed: args.rank = args.rank * ngpus_per_node + gpu dist.init_process_group(backend=args.dist_backend, init_method=args.dist_url, world_size=args.world_size, rank=args.rank) # Create model model = BtsModel(args) model.train() model.decoder.apply(weights_init_xavier) set_misc(model) num_params = sum([np.prod(p.size()) for p in model.parameters()]) print("Total number of parameters: {}".format(num_params)) num_params_update = sum([np.prod(p.shape) for p in model.parameters() if p.requires_grad]) print("Total number of learning parameters: {}".format(num_params_update)) if args.distributed: if args.gpu is not None: torch.cuda.set_device(args.gpu) model.cuda(args.gpu) args.batch_size = int(args.batch_size / ngpus_per_node) model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu], find_unused_parameters=True) else: model.cuda() model = torch.nn.parallel.DistributedDataParallel(model, find_unused_parameters=True) else: model = torch.nn.DataParallel(model) model.cuda() if args.distributed: print("Model Initialized on GPU: {}".format(args.gpu)) else: print("Model Initialized") global_step = 0 best_eval_measures_lower_better = torch.zeros(6).cpu() + 1e3 best_eval_measures_higher_better = torch.zeros(3).cpu() best_eval_steps = np.zeros(9, dtype=np.int32) # Training parameters optimizer = torch.optim.AdamW([{'params': model.module.encoder.parameters(), 'weight_decay': args.weight_decay}, {'params': model.module.decoder.parameters(), 'weight_decay': 0}], lr=args.learning_rate, eps=args.adam_eps) model_just_loaded = False if args.checkpoint_path != '': if os.path.isfile(args.checkpoint_path): print("Loading checkpoint '{}'".format(args.checkpoint_path)) if args.gpu is None: checkpoint = torch.load(args.checkpoint_path) else: loc = 'cuda:{}'.format(args.gpu) checkpoint = torch.load(args.checkpoint_path, map_location=loc) global_step = checkpoint['global_step'] model.load_state_dict(checkpoint['model']) optimizer.load_state_dict(checkpoint['optimizer']) try: best_eval_measures_higher_better = checkpoint['best_eval_measures_higher_better'].cpu() best_eval_measures_lower_better = checkpoint['best_eval_measures_lower_better'].cpu() best_eval_steps = checkpoint['best_eval_steps'] except KeyError: print("Could not load values for online evaluation") print("Loaded checkpoint '{}' (global_step {})".format(args.checkpoint_path, checkpoint['global_step'])) else: print("No checkpoint found at '{}'".format(args.checkpoint_path)) model_just_loaded = True if args.retrain: global_step = 0 cudnn.benchmark = True dataloader = BtsDataLoader(args, 'train') dataloader_eval = BtsDataLoader(args, 'online_eval') # Logging if not args.multiprocessing_distributed or (args.multiprocessing_distributed and args.rank % ngpus_per_node == 0): writer = SummaryWriter(args.log_directory + '/' + args.model_name + '/summaries', flush_secs=30) if args.do_online_eval: if args.eval_summary_directory != '': eval_summary_path = os.path.join(args.eval_summary_directory, args.model_name) else: eval_summary_path = os.path.join(args.log_directory, 'eval') eval_summary_writer = SummaryWriter(eval_summary_path, flush_secs=30) criterion = silog_loss(args.variance_focus) criterion_uncer = uncertainty_loss(args) criterion_entro = entropy_loss() criterion_dissi = dissimilar_loss() start_time = time.time() duration = 0 num_log_images = args.batch_size end_learning_rate = args.end_learning_rate if args.end_learning_rate != -1 else 0.1 * args.learning_rate steps_per_epoch = len(dataloader.data) num_total_steps = args.num_epochs * steps_per_epoch epoch = global_step // steps_per_epoch while epoch < args.num_epochs: if args.distributed: dataloader.train_sampler.set_epoch(epoch) for step, sample_batched in enumerate(dataloader.data): optimizer.zero_grad() before_op_time = time.time() image = torch.autograd.Variable(sample_batched['image'].cuda(args.gpu, non_blocking=True)) focal = torch.autograd.Variable(sample_batched['focal'].cuda(args.gpu, non_blocking=True)) depth_gt = torch.autograd.Variable(sample_batched['depth'].cuda(args.gpu, non_blocking=True)) final_depth, final_uncer, omega, embedding_ = model(image, focal) if args.dataset == 'nyu': mask = depth_gt > 0.1 else: mask = depth_gt > 1.0 mask = mask.to(torch.bool) loss_depth = criterion.forward(final_depth, depth_gt, mask) loss_uncer = criterion_uncer.forward(final_uncer, final_depth, depth_gt, mask) loss_omega = criterion_entro.forward(embedding_) loss_dissi = criterion_dissi.forward(omega) loss = loss_depth + (loss_uncer + loss_omega + loss_dissi) * args.loss_lambda loss.backward() for param_group in optimizer.param_groups: current_lr = (args.learning_rate - end_learning_rate) * (1 - global_step / num_total_steps) ** 0.9 + end_learning_rate param_group['lr'] = current_lr optimizer.step() if not args.multiprocessing_distributed or (args.multiprocessing_distributed and args.rank % ngpus_per_node == 0): print('[epoch][s/s_per_e/gs]: [{}][{}/{}/{}], lr: {:.12f}, loss: {:.12f}'.format(epoch, step, steps_per_epoch, global_step, current_lr, loss)) if np.isnan(loss.cpu().item()): print('NaN in loss occurred. Aborting training.') return -1 duration += time.time() - before_op_time if global_step and global_step % args.log_freq == 0 and not model_just_loaded: examples_per_sec = args.batch_size / duration * args.log_freq duration = 0 time_sofar = (time.time() - start_time) / 3600 training_time_left = (num_total_steps / global_step - 1.0) * time_sofar if not args.multiprocessing_distributed or (args.multiprocessing_distributed and args.rank % ngpus_per_node == 0): print("{}".format(args.model_name)) print_string = 'GPU: {} | examples/s: {:4.2f} | loss: {:.5f} | time elapsed: {:.2f}h | time left: {:.2f}h' print(print_string.format(args.gpu, examples_per_sec, loss, time_sofar, training_time_left)) if not args.multiprocessing_distributed or (args.multiprocessing_distributed and args.rank % ngpus_per_node == 0): writer.add_scalar('total_loss', loss, global_step) writer.add_scalar('loss_depth', loss_depth, global_step) writer.add_scalar('loss_uncer', loss_uncer, global_step) writer.add_scalar('learning_rate', current_lr, global_step) for i in range(num_log_images): writer.add_image('depth_mean/image/{}'.format(i), normalize_result(1/(final_depth)[i, :, :, :].data), global_step) writer.add_image('depth_var/image/{}'.format(i), normalize_result((final_uncer.detach().sigmoid())[i, :, :, :].data), global_step) writer.add_image('image/image/{}'.format(i), inv_normalize(image[i, :, :, :]).data, global_step) writer.flush() if not args.do_online_eval and global_step and global_step % args.save_freq == 0: if not args.multiprocessing_distributed or (args.multiprocessing_distributed and args.rank % ngpus_per_node == 0): checkpoint = {'global_step': global_step, 'model': model.state_dict(), 'optimizer': optimizer.state_dict()} torch.save(checkpoint, args.log_directory + '/' + args.model_name + '/model-{}'.format(global_step)) if args.do_online_eval and global_step and global_step % args.eval_freq == 0 and not model_just_loaded: time.sleep(0.1) model.eval() eval_measures = online_eval(model, dataloader_eval, gpu, ngpus_per_node) if eval_measures is not None: for i in range(9): eval_summary_writer.add_scalar(eval_metrics[i], eval_measures[i].cpu(), int(global_step)) measure = eval_measures[i] is_best = False if i < 6 and measure < best_eval_measures_lower_better[i]: old_best = best_eval_measures_lower_better[i].item() best_eval_measures_lower_better[i] = measure.item() is_best = True elif i >= 6 and measure > best_eval_measures_higher_better[i-6]: old_best = best_eval_measures_higher_better[i-6].item() best_eval_measures_higher_better[i-6] = measure.item() is_best = True if is_best: old_best_step = best_eval_steps[i] old_best_name = '/model-{}-best_{}_{:.5f}'.format(old_best_step, eval_metrics[i], old_best) model_path = args.log_directory + '/' + args.model_name + old_best_name if os.path.exists(model_path): command = 'rm {}'.format(model_path) os.system(command) best_eval_steps[i] = global_step model_save_name = '/model-{}-best_{}_{:.5f}'.format(global_step, eval_metrics[i], measure) print('New best for {}. Saving model: {}'.format(eval_metrics[i], model_save_name)) checkpoint = {'global_step': global_step, 'model': model.state_dict(), 'best_eval_measures_higher_better': best_eval_measures_higher_better, 'best_eval_measures_lower_better': best_eval_measures_lower_better, 'best_eval_steps': best_eval_steps } torch.save(checkpoint, args.log_directory + '/' + args.model_name + model_save_name) eval_summary_writer.flush() model.train() block_print() set_misc(model) enable_print() model_just_loaded = False global_step += 1 epoch += 1 if not args.multiprocessing_distributed or (args.multiprocessing_distributed and args.rank % ngpus_per_node == 0): writer.close() if args.do_online_eval: eval_summary_writer.close() def main(): if args.mode != 'train': print('bts_main.py is only for training. Use bts_test.py instead.') return -1 model_filename = args.model_name + '.py' command = 'mkdir ' + args.log_directory + '/' + args.model_name os.system(command) args_out_path = args.log_directory + '/' + args.model_name + '/' + sys.argv[1] command = 'cp ' + sys.argv[1] + ' ' + args_out_path os.system(command) if args.checkpoint_path == '': model_out_path = args.log_directory + '/' + args.model_name + '/' + model_filename command = 'cp bts.py ' + model_out_path os.system(command) aux_out_path = args.log_directory + '/' + args.model_name + '/.' command = 'cp bts_main.py ' + aux_out_path os.system(command) command = 'cp bts_dataloader.py ' + aux_out_path os.system(command) else: loaded_model_dir = os.path.dirname(args.checkpoint_path) loaded_model_name = os.path.basename(loaded_model_dir) loaded_model_filename = loaded_model_name + '.py' model_out_path = args.log_directory + '/' + args.model_name + '/' + model_filename command = 'cp ' + loaded_model_dir + '/' + loaded_model_filename + ' ' + model_out_path os.system(command) torch.cuda.empty_cache() args.distributed = args.world_size > 1 or args.multiprocessing_distributed ngpus_per_node = torch.cuda.device_count() if ngpus_per_node > 1 and not args.multiprocessing_distributed: print("This machine has more than 1 gpu. Please specify --multiprocessing_distributed, or set \'CUDA_VISIBLE_DEVICES=0\'") return -1 if args.do_online_eval: print("You have specified --do_online_eval.") print("This will evaluate the model every eval_freq {} steps and save best models for individual eval metrics." .format(args.eval_freq)) if args.multiprocessing_distributed: args.world_size = ngpus_per_node * args.world_size mp.spawn(main_worker, nprocs=ngpus_per_node, args=(ngpus_per_node, args)) else: main_worker(args.gpu, ngpus_per_node, args) if __name__ == '__main__': main()

py

LDU

LDU-main/monocular_depth_estimation/pytorch/bts_ldu.py

# Copyright (C) 2019 Jin Han Lee # # This file is a part of BTS. # This program is free software: you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation, either version 3 of the License, or # (at your option) any later version. # # This program is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with this program. If not, see <http://www.gnu.org/licenses/> import torch import torch.nn as nn import torch.nn.functional as torch_nn_func import math import numpy as np # This sets the batch norm layers in pytorch as if {'is_training': False, 'scale': True} in tensorflow def bn_init_as_tf(m): if isinstance(m, nn.BatchNorm2d): m.track_running_stats = True # These two lines enable using stats (moving mean and var) loaded from pretrained model m.eval() # or zero mean and variance of one if the batch norm layer has no pretrained values m.affine = True m.requires_grad = True def weights_init_xavier(m): if isinstance(m, nn.Conv2d): torch.nn.init.xavier_uniform_(m.weight) if m.bias is not None: torch.nn.init.zeros_(m.bias) #_______________________________________________________________________________________# class silog_loss(nn.Module): def __init__(self, variance_focus): super(silog_loss, self).__init__() self.variance_focus = variance_focus def forward(self, depth_est, depth_gt, mask): d = torch.log(depth_est[mask]) - torch.log(depth_gt[mask]) return torch.sqrt((d ** 2).mean() - self.variance_focus * (d.mean() ** 2)) * 10.0 class entropy_loss(nn.Module): def __init__(self): super(entropy_loss, self).__init__() def forward(self, embedding): embedding = nn.Softmax(dim=1)(embedding) minus_entropy = embedding * torch.log(embedding) minus_entropy = torch.sum(minus_entropy, dim=1) return minus_entropy.mean() class uncertainty_loss(nn.Module): def __init__(self, args): super(uncertainty_loss, self).__init__() self.max_depth = args.max_depth def forward(self, uncer, final_depth, depth_gt, mask): abs_error = abs(final_depth.detach() - depth_gt)/self.max_depth abs_error[abs_error>1] = 1 abs_error = abs_error[mask].detach() loss = nn.BCEWithLogitsLoss(pos_weight = torch.tensor([5.0]).cuda(), reduction='mean')(uncer[mask], abs_error) return loss class dissimilar_loss(nn.Module): def __init__(self): super(dissimilar_loss, self).__init__() def forward(self, protos): loss = -1 * torch.mean(torch.cdist(protos, protos)) return loss #_______________________________________________________________________________________# class atrous_conv(nn.Sequential): def __init__(self, in_channels, out_channels, dilation, apply_bn_first=True): super(atrous_conv, self).__init__() self.atrous_conv = torch.nn.Sequential() if apply_bn_first: self.atrous_conv.add_module('first_bn', nn.BatchNorm2d(in_channels, momentum=0.01, affine=True, track_running_stats=True, eps=1.1e-5)) self.atrous_conv.add_module('aconv_sequence', nn.Sequential(nn.ReLU(), nn.Conv2d(in_channels=in_channels, out_channels=out_channels*2, bias=False, kernel_size=1, stride=1, padding=0), nn.BatchNorm2d(out_channels*2, momentum=0.01, affine=True, track_running_stats=True), nn.ReLU(), nn.Conv2d(in_channels=out_channels * 2, out_channels=out_channels, bias=False, kernel_size=3, stride=1, padding=(dilation, dilation), dilation=dilation))) def forward(self, x): return self.atrous_conv.forward(x) class upconv(nn.Module): def __init__(self, in_channels, out_channels, ratio=2): super(upconv, self).__init__() self.elu = nn.ELU() self.conv = nn.Conv2d(in_channels=in_channels, out_channels=out_channels, bias=False, kernel_size=3, stride=1, padding=1) self.ratio = ratio def forward(self, x): up_x = torch_nn_func.interpolate(x, scale_factor=self.ratio, mode='nearest') out = self.conv(up_x) out = self.elu(out) return out class reduction_1x1(nn.Sequential): def __init__(self, num_in_filters, num_out_filters, max_depth, is_final=False): super(reduction_1x1, self).__init__() self.max_depth = max_depth self.is_final = is_final self.sigmoid = nn.Sigmoid() self.reduc = torch.nn.Sequential() while num_out_filters >= 4: if num_out_filters < 8: if self.is_final: self.reduc.add_module('final', torch.nn.Sequential(nn.Conv2d(num_in_filters, out_channels=1, bias=False, kernel_size=1, stride=1, padding=0), nn.Sigmoid())) else: self.reduc.add_module('plane_params', torch.nn.Conv2d(num_in_filters, out_channels=3, bias=False, kernel_size=1, stride=1, padding=0)) break else: self.reduc.add_module('inter_{}_{}'.format(num_in_filters, num_out_filters), torch.nn.Sequential(nn.Conv2d(in_channels=num_in_filters, out_channels=num_out_filters, bias=False, kernel_size=1, stride=1, padding=0), nn.ELU())) num_in_filters = num_out_filters num_out_filters = num_out_filters // 2 def forward(self, net): net = self.reduc.forward(net) if not self.is_final: theta = self.sigmoid(net[:, 0, :, :]) * math.pi / 3 phi = self.sigmoid(net[:, 1, :, :]) * math.pi * 2 dist = self.sigmoid(net[:, 2, :, :]) * self.max_depth n1 = torch.mul(torch.sin(theta), torch.cos(phi)).unsqueeze(1) n2 = torch.mul(torch.sin(theta), torch.sin(phi)).unsqueeze(1) n3 = torch.cos(theta).unsqueeze(1) n4 = dist.unsqueeze(1) net = torch.cat([n1, n2, n3, n4], dim=1) return net class local_planar_guidance(nn.Module): def __init__(self, upratio): super(local_planar_guidance, self).__init__() self.upratio = upratio self.u = torch.arange(self.upratio).reshape([1, 1, self.upratio]).float() self.v = torch.arange(int(self.upratio)).reshape([1, self.upratio, 1]).float() self.upratio = float(upratio) def forward(self, plane_eq, focal): plane_eq_expanded = torch.repeat_interleave(plane_eq, int(self.upratio), 2) plane_eq_expanded = torch.repeat_interleave(plane_eq_expanded, int(self.upratio), 3) n1 = plane_eq_expanded[:, 0, :, :] n2 = plane_eq_expanded[:, 1, :, :] n3 = plane_eq_expanded[:, 2, :, :] n4 = plane_eq_expanded[:, 3, :, :] u = self.u.repeat(plane_eq.size(0), plane_eq.size(2) * int(self.upratio), plane_eq.size(3)).cuda() u = (u - (self.upratio - 1) * 0.5) / self.upratio v = self.v.repeat(plane_eq.size(0), plane_eq.size(2), plane_eq.size(3) * int(self.upratio)).cuda() v = (v - (self.upratio - 1) * 0.5) / self.upratio return n4 / (n1 * u + n2 * v + n3) class Distanceminimi_Layer_learned(nn.Module): def __init__(self, in_features=0, out_features=0, dist='lin'): super(Distanceminimi_Layer_learned, self).__init__() self.in_features = in_features self.out_features = out_features self.dist=dist self.omega = nn.Parameter(torch.Tensor(1, out_features, in_features, 1, 1)) self.reset_parameters() def reset_parameters(self): nn.init.normal_(self.omega, mean=0, std=1)#/self.out_features) def forward(self, x): x = x.unsqueeze(1) out = torch_nn_func.cosine_similarity(x, self.omega, dim=2, eps=1e-30) return out, self.omega class bts(nn.Module): def __init__(self, params, feat_out_channels, num_features=512, nb_prototype = 80): super(bts, self).__init__() self.params = params self.upconv5 = upconv(feat_out_channels[4], num_features) self.bn5 = nn.BatchNorm2d(num_features, momentum=0.01, affine=True, eps=1.1e-5) self.conv5 = torch.nn.Sequential(nn.Conv2d(num_features + feat_out_channels[3], num_features, 3, 1, 1, bias=False), nn.ELU()) self.upconv4 = upconv(num_features, num_features // 2) self.bn4 = nn.BatchNorm2d(num_features // 2, momentum=0.01, affine=True, eps=1.1e-5) self.conv4 = torch.nn.Sequential(nn.Conv2d(num_features // 2 + feat_out_channels[2], num_features // 2, 3, 1, 1, bias=False), nn.ELU()) self.bn4_2 = nn.BatchNorm2d(num_features // 2, momentum=0.01, affine=True, eps=1.1e-5) self.daspp_3 = atrous_conv(num_features // 2, num_features // 4, 3, apply_bn_first=False) self.daspp_6 = atrous_conv(num_features // 2 + num_features // 4 + feat_out_channels[2], num_features // 4, 6) self.daspp_12 = atrous_conv(num_features + feat_out_channels[2], num_features // 4, 12) self.daspp_18 = atrous_conv(num_features + num_features // 4 + feat_out_channels[2], num_features // 4, 18) self.daspp_24 = atrous_conv(num_features + num_features // 2 + feat_out_channels[2], num_features // 4, 24) self.daspp_conv = torch.nn.Sequential(nn.Conv2d(num_features + num_features // 2 + num_features // 4, num_features // 4, 3, 1, 1, bias=False), nn.ELU()) self.reduc8x8 = reduction_1x1(num_features // 4, num_features // 4, self.params.max_depth) self.lpg8x8 = local_planar_guidance(8) self.upconv3 = upconv(num_features // 4, num_features // 4) self.bn3 = nn.BatchNorm2d(num_features // 4, momentum=0.01, affine=True, eps=1.1e-5) self.conv3 = torch.nn.Sequential(nn.Conv2d(num_features // 4 + feat_out_channels[1] + 1, num_features // 4, 3, 1, 1, bias=False), nn.ELU()) self.reduc4x4 = reduction_1x1(num_features // 4, num_features // 8, self.params.max_depth) self.lpg4x4 = local_planar_guidance(4) self.upconv2 = upconv(num_features // 4, num_features // 8) self.bn2 = nn.BatchNorm2d(num_features // 8, momentum=0.01, affine=True, eps=1.1e-5) self.conv2 = torch.nn.Sequential(nn.Conv2d(num_features // 8 + feat_out_channels[0] + 1, num_features // 8, 3, 1, 1, bias=False), nn.ELU()) self.reduc2x2 = reduction_1x1(num_features // 8, num_features // 16, self.params.max_depth) self.lpg2x2 = local_planar_guidance(2) self.upconv1 = upconv(num_features // 8, num_features // 16) self.reduc1x1 = reduction_1x1(num_features // 16, num_features // 32, self.params.max_depth, is_final=True) self.conv1 = torch.nn.Sequential(nn.Conv2d(num_features // 16 + 4, num_features // 16, 3, 1, 1, bias=False), nn.ELU()) self.DMlayer = Distanceminimi_Layer_learned(in_features=(num_features // 16), out_features = nb_prototype, dist='cos') self.DMBN = nn.BatchNorm2d(nb_prototype) self.get_uncer = nn.Conv2d(nb_prototype, 1, 1) self.get_depth = nn.Sequential(nn.Conv2d(nb_prototype, 1, 1), nn.Sigmoid()) def forward(self, features, focal): skip0, skip1, skip2, skip3 = features[0], features[1], features[2], features[3] dense_features = torch.nn.ReLU()(features[4]) upconv5 = self.upconv5(dense_features) # H/16 upconv5 = self.bn5(upconv5) concat5 = torch.cat([upconv5, skip3], dim=1) iconv5 = self.conv5(concat5) upconv4 = self.upconv4(iconv5) # H/8 upconv4 = self.bn4(upconv4) concat4 = torch.cat([upconv4, skip2], dim=1) iconv4 = self.conv4(concat4) iconv4 = self.bn4_2(iconv4) daspp_3 = self.daspp_3(iconv4) concat4_2 = torch.cat([concat4, daspp_3], dim=1) daspp_6 = self.daspp_6(concat4_2) concat4_3 = torch.cat([concat4_2, daspp_6], dim=1) daspp_12 = self.daspp_12(concat4_3) concat4_4 = torch.cat([concat4_3, daspp_12], dim=1) daspp_18 = self.daspp_18(concat4_4) concat4_5 = torch.cat([concat4_4, daspp_18], dim=1) daspp_24 = self.daspp_24(concat4_5) concat4_daspp = torch.cat([iconv4, daspp_3, daspp_6, daspp_12, daspp_18, daspp_24], dim=1) daspp_feat = self.daspp_conv(concat4_daspp) reduc8x8 = self.reduc8x8(daspp_feat) plane_normal_8x8 = reduc8x8[:, :3, :, :] plane_normal_8x8 = torch_nn_func.normalize(plane_normal_8x8, 2, 1) plane_dist_8x8 = reduc8x8[:, 3, :, :] plane_eq_8x8 = torch.cat([plane_normal_8x8, plane_dist_8x8.unsqueeze(1)], 1) depth_8x8 = self.lpg8x8(plane_eq_8x8, focal) depth_8x8_scaled = depth_8x8.unsqueeze(1) / self.params.max_depth depth_8x8_scaled_ds = torch_nn_func.interpolate(depth_8x8_scaled, scale_factor=0.25, mode='nearest') upconv3 = self.upconv3(daspp_feat) # H/4 upconv3 = self.bn3(upconv3) concat3 = torch.cat([upconv3, skip1, depth_8x8_scaled_ds], dim=1) iconv3 = self.conv3(concat3) reduc4x4 = self.reduc4x4(iconv3) plane_normal_4x4 = reduc4x4[:, :3, :, :] plane_normal_4x4 = torch_nn_func.normalize(plane_normal_4x4, 2, 1) plane_dist_4x4 = reduc4x4[:, 3, :, :] plane_eq_4x4 = torch.cat([plane_normal_4x4, plane_dist_4x4.unsqueeze(1)], 1) depth_4x4 = self.lpg4x4(plane_eq_4x4, focal) depth_4x4_scaled = depth_4x4.unsqueeze(1) / self.params.max_depth depth_4x4_scaled_ds = torch_nn_func.interpolate(depth_4x4_scaled, scale_factor=0.5, mode='nearest') upconv2 = self.upconv2(iconv3) # H/2 upconv2 = self.bn2(upconv2) concat2 = torch.cat([upconv2, skip0, depth_4x4_scaled_ds], dim=1) iconv2 = self.conv2(concat2) reduc2x2 = self.reduc2x2(iconv2) plane_normal_2x2 = reduc2x2[:, :3, :, :] plane_normal_2x2 = torch_nn_func.normalize(plane_normal_2x2, 2, 1) plane_dist_2x2 = reduc2x2[:, 3, :, :] plane_eq_2x2 = torch.cat([plane_normal_2x2, plane_dist_2x2.unsqueeze(1)], 1) depth_2x2 = self.lpg2x2(plane_eq_2x2, focal) depth_2x2_scaled = depth_2x2.unsqueeze(1) / self.params.max_depth upconv1 = self.upconv1(iconv2) reduc1x1 = self.reduc1x1(upconv1) concat1 = torch.cat([upconv1, reduc1x1, depth_2x2_scaled, depth_4x4_scaled, depth_8x8_scaled], dim=1) feature_output = self.conv1(concat1) # Before the last layer, DM layer is added embedding_, omega = self.DMlayer(feature_output) embedding = torch.exp(-embedding_) out = self.DMBN(embedding) final_uncer = self.get_uncer(out) final_depth = self.get_depth(out) * self.params.max_depth if self.training: return final_depth, final_uncer, omega.squeeze(), embedding_ else: return final_depth, torch.sigmoid(final_uncer) class encoder(nn.Module): def __init__(self, params): super(encoder, self).__init__() self.params = params import torchvision.models as models if params.encoder == 'densenet121_bts': self.base_model = models.densenet121(pretrained=True).features self.feat_names = ['relu0', 'pool0', 'transition1', 'transition2', 'norm5'] self.feat_out_channels = [64, 64, 128, 256, 1024] elif params.encoder == 'densenet161_bts': self.base_model = models.densenet161(pretrained=True).features self.feat_names = ['relu0', 'pool0', 'transition1', 'transition2', 'norm5'] self.feat_out_channels = [96, 96, 192, 384, 2208] elif params.encoder == 'resnet50_bts': self.base_model = models.resnet50(pretrained=True) self.feat_names = ['relu', 'layer1', 'layer2', 'layer3', 'layer4'] self.feat_out_channels = [64, 256, 512, 1024, 2048] elif params.encoder == 'resnet101_bts': self.base_model = models.resnet101(pretrained=True) self.feat_names = ['relu', 'layer1', 'layer2', 'layer3', 'layer4'] self.feat_out_channels = [64, 256, 512, 1024, 2048] elif params.encoder == 'resnext50_bts': self.base_model = models.resnext50_32x4d(pretrained=True) self.feat_names = ['relu', 'layer1', 'layer2', 'layer3', 'layer4'] self.feat_out_channels = [64, 256, 512, 1024, 2048] elif params.encoder == 'resnext101_bts': self.base_model = models.resnext101_32x8d(pretrained=True) self.feat_names = ['relu', 'layer1', 'layer2', 'layer3', 'layer4'] self.feat_out_channels = [64, 256, 512, 1024, 2048] elif params.encoder == 'mobilenetv2_bts': self.base_model = models.mobilenet_v2(pretrained=True).features self.feat_inds = [2, 4, 7, 11, 19] self.feat_out_channels = [16, 24, 32, 64, 1280] self.feat_names = [] else: print('Not supported encoder: {}'.format(params.encoder)) def forward(self, x): feature = x skip_feat = [] i = 1 for k, v in self.base_model._modules.items(): if 'fc' in k or 'avgpool' in k: continue feature = v(feature) if self.params.encoder == 'mobilenetv2_bts': if i == 2 or i == 4 or i == 7 or i == 11 or i == 19: skip_feat.append(feature) else: if any(x in k for x in self.feat_names): skip_feat.append(feature) i = i + 1 return skip_feat class BtsModel(nn.Module): def __init__(self, params): super(BtsModel, self).__init__() self.encoder = encoder(params) self.decoder = bts(params, self.encoder.feat_out_channels, params.bts_size, params.nb_proto) def forward(self, x, focal): skip_feat = self.encoder(x) return self.decoder(skip_feat, focal)

py

LDU

LDU-main/monocular_depth_estimation/pytorch/bts_test_kitti_ldu.py

# Copyright (C) 2019 Jin Han Lee # # This file is a part of BTS. # This program is free software: you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation, either version 3 of the License, or # (at your option) any later version. # # This program is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with this program. If not, see <http://www.gnu.org/licenses/> from __future__ import absolute_import, division, print_function import os import argparse import time import numpy as np import sys import torch from torch.autograd import Variable from tqdm import tqdm from bts_dataloader import * from sparsification import sparsification_error_gpu from bts_ldu import BtsModel def convert_arg_line_to_args(arg_line): for arg in arg_line.split(): if not arg.strip(): continue yield arg parser = argparse.ArgumentParser(description='BTS PyTorch implementation.', fromfile_prefix_chars='@') parser.convert_arg_line_to_args = convert_arg_line_to_args parser.add_argument('--model_name', type=str, help='model name', default='bts_nyu_v2') parser.add_argument('--encoder', type=str, help='type of encoder, vgg or desenet121_bts or densenet161_bts', default='densenet161_bts') parser.add_argument('--data_path_eval', type=str, help='path to the data', required=True) parser.add_argument('--gt_path_eval', type=str, help='path to the data', required=True) parser.add_argument('--filenames_file_eval', type=str, help='path to the filenames text file', required=True) parser.add_argument('--input_height', type=int, help='input height', default=480) parser.add_argument('--input_width', type=int, help='input width', default=640) parser.add_argument('--max_depth', type=float, help='maximum depth in estimation', default=80) parser.add_argument('--checkpoint_path', type=str, help='path to a specific checkpoint to load', default='') parser.add_argument('--dataset', type=str, help='dataset to train on, make3d or nyudepthv2', default='nyu') parser.add_argument('--do_kb_crop', help='if set, crop input images as kitti benchmark images', action='store_true') parser.add_argument('--garg_crop', help='if set, crops according to Garg ECCV16', action='store_true') parser.add_argument('--bts_size', type=int, help='initial num_filters in bts', default=512) parser.add_argument('--clip_gt', help='if set, clipping the ground truth to the min-max depth', action='store_true') parser.add_argument('--min_depth_eval', type=float, help='minimum depth for evaluation', default=1e-3) parser.add_argument('--max_depth_eval', type=float, help='maximum depth for evaluation', default=80) parser.add_argument('--nb_proto', type=int, help='initial num_proto in bts', default=30) if sys.argv.__len__() == 2: arg_filename_with_prefix = '@' + sys.argv[1] args = parser.parse_args([arg_filename_with_prefix]) else: args = parser.parse_args() def get_num_lines(file_path): f = open(file_path, 'r') lines = f.readlines() f.close() return len(lines) def compute_errors(gt, pred): thresh = np.maximum((gt / pred), (pred / gt)) d1 = (thresh < 1.25).mean() d2 = (thresh < 1.25 ** 2).mean() d3 = (thresh < 1.25 ** 3).mean() rmse = (gt - pred) ** 2 rmse = np.sqrt(rmse.mean()) rmse_log = (np.log(gt) - np.log(pred)) ** 2 rmse_log = np.sqrt(rmse_log.mean()) abs_rel = np.mean(np.abs(gt - pred) / gt) sq_rel = np.mean(((gt - pred)**2) / gt) err = np.log(pred) - np.log(gt) silog = np.sqrt(np.mean(err ** 2) - np.mean(err) ** 2) * 100 err = np.abs(np.log10(pred) - np.log10(gt)) log10 = np.mean(err) return silog, log10, abs_rel, sq_rel, rmse, rmse_log, d1, d2, d3 inv_normalize = transforms.Normalize( mean=[-0.485/0.229, -0.456/0.224, -0.406/0.225], std=[1/0.229, 1/0.224, 1/0.225] ) def test(): """Test function.""" args.mode = 'online_eval' args.distributed = False dataloader = BtsDataLoader(args, 'online_eval') model = BtsModel(params=args) model = torch.nn.DataParallel(model) if os.path.exists(args.checkpoint_path): checkpoint = torch.load(args.checkpoint_path) else: print('Wrong checkpoint path. Exit.') exit() model.load_state_dict(checkpoint['model']) model.eval() model.cuda() num_params = sum([np.prod(p.size()) for p in model.parameters()]) print("Total number of parameters: {}".format(num_params)) num_test_samples = get_num_lines(args.filenames_file_eval) print('now testing {} files with {}'.format(num_test_samples, args.checkpoint_path)) start_time = time.time() with torch.no_grad(): num_samples = len(dataloader.data) print(num_samples) nb_valid = 0 silog = np.zeros(num_samples, np.float32) log10 = np.zeros(num_samples, np.float32) rms = np.zeros(num_samples, np.float32) log_rms = np.zeros(num_samples, np.float32) abs_rel = np.zeros(num_samples, np.float32) sq_rel = np.zeros(num_samples, np.float32) d1 = np.zeros(num_samples, np.float32) d2 = np.zeros(num_samples, np.float32) d3 = np.zeros(num_samples, np.float32) hist_pred_rmses = 0 hist_oracle_rmses = 0 nb_remain_rmses = 0 hist_pred_rmses = 0 hist_oracle_rmses = 0 nb_remain_rmses = 0 ausc_rmse = np.zeros(num_samples, np.float32) hist_pred_absrels = 0 hist_oracle_absrels = 0 nb_remain_absrels = 0 hist_pred_absrels = 0 hist_oracle_absrels = 0 nb_remain_absrels = 0 ausc_absrel = np.zeros(num_samples, np.float32) spar_rmse = 0 spar_absr = 0 for i, sample in tqdm(enumerate(tqdm(dataloader.data))): is_valid = sample['has_valid_depth'] if not is_valid: continue else: nb_valid += 1 image = Variable(sample['image'].cuda()) focal = Variable(sample['focal'].cuda()) depth_gt = Variable(sample['depth'].cuda()) # Predict depth_gt = depth_gt.cpu().numpy().squeeze() depth_est, uncertainty = model(image, focal) depth_est = depth_est.cpu().numpy().squeeze() uncertainty = uncertainty.cpu().numpy().squeeze() if args.clip_gt: valid_mask = np.logical_and(depth_gt > args.min_depth_eval, depth_gt < args.max_depth_eval) else: valid_mask = (depth_gt > args.min_depth_eval) # We are using online-eval here, and the following operation is to imitate the operation in the test case in the original work. if args.do_kb_crop: height, width = depth_gt.shape top_margin = int(height - 352) left_margin = int((width - 1216) / 2) pred_depth_uncropped = np.zeros((height, width), dtype=np.float32) pred_depth_uncropped[top_margin:top_margin + 352, left_margin:left_margin + 1216] = depth_est depth_est = pred_depth_uncropped pred_depth_uncropped = np.zeros((height, width), dtype=np.float32) pred_depth_uncropped[top_margin:top_margin + 352, left_margin:left_margin + 1216] = uncertainty uncertainty = pred_depth_uncropped if args.clip_gt: depth_est[depth_est < args.min_depth_eval] = args.min_depth_eval depth_est[depth_est > args.max_depth_eval] = args.max_depth_eval depth_est[np.isinf(depth_est)] = args.max_depth_eval depth_gt[np.isinf(depth_gt)] = args.max_depth_eval depth_gt[np.isnan(depth_gt)] = args.min_depth_eval if args.garg_crop: gt_height, gt_width = depth_gt.shape eval_mask = np.zeros(valid_mask.shape) eval_mask[int(0.40810811 * gt_height):int(0.99189189 * gt_height), int(0.03594771 * gt_width):int(0.96405229 * gt_width)] = 1 valid_mask = np.logical_and(valid_mask, eval_mask) uncertainty = torch.tensor(uncertainty).cuda() depth_est = torch.tensor(depth_est).cuda() depth_gt = torch.tensor(depth_gt).cuda() valid_mask = torch.tensor(valid_mask).cuda() hist_pred_rmse, hist_oracle_rmse, nb_remain_rmse, ausc_rmse[i] = sparsification_error_gpu(unc_tensor = uncertainty[valid_mask], pred_tensor = depth_est[valid_mask], gt_tensor = depth_gt[valid_mask], is_rmse = True) hist_pred_rmses += hist_pred_rmse hist_oracle_rmses += hist_oracle_rmse nb_remain_rmses += nb_remain_rmse spar_rmse += np.trapz((hist_pred_rmse - hist_oracle_rmse), x = list(np.arange(start=0.0, stop=1.0, step=(1/100)))) hist_pred_absrel, hist_oracle_absrel, nb_remain_absrel, ausc_absrel[i] = sparsification_error_gpu(unc_tensor = uncertainty[valid_mask], pred_tensor = depth_est[valid_mask], gt_tensor = depth_gt[valid_mask], is_rmse = False) hist_pred_absrels += hist_pred_absrel hist_oracle_absrels += hist_oracle_absrel nb_remain_absrels += nb_remain_absrel spar_absr += np.trapz((hist_pred_absrel - hist_oracle_absrel), x = list(np.arange(start=0.0, stop=1.0, step=(1/100)))) depth_est = depth_est.cpu().numpy() depth_gt = depth_gt.cpu().numpy() valid_mask = valid_mask.cpu().numpy() silog[i], log10[i], abs_rel[i], sq_rel[i], rms[i], log_rms[i], d1[i], d2[i], d3[i] = compute_errors(depth_gt[valid_mask], depth_est[valid_mask]) print(nb_valid) print("{:>7}, {:>7}, {:>7}, {:>7}, {:>7}, {:>7}, {:>7}, {:>7}, {:>7}".format( 'd1', 'd2', 'd3', 'AbsRel', 'SqRel', 'RMSE', 'RMSElog', 'SILog', 'log10')) print("{:7.3f}, {:7.3f}, {:7.3f}, {:7.3f}, {:7.3f}, {:7.3f}, {:7.3f}, {:7.3f}, {:7.3f}".format( d1.sum()/nb_valid, d2.sum()/nb_valid, d3.sum()/nb_valid, abs_rel.sum()/nb_valid, sq_rel.sum()/nb_valid, rms.sum()/nb_valid, log_rms.sum()/nb_valid, silog.sum()/nb_valid, log10.sum()/nb_valid)) hist_pred_rmses = hist_pred_rmses/nb_valid hist_oracle_rmses = hist_oracle_rmses/nb_valid nb_remain_rmses = nb_remain_rmses/nb_valid hist_pred_absrels = hist_pred_absrels/nb_valid hist_oracle_absrels = hist_oracle_absrels/nb_valid nb_remain_absrels = nb_remain_absrels/nb_valid spar_rmse = spar_rmse/nb_valid spar_absr = spar_absr/nb_valid # to verify that the averages obtained by the two different methods are consistent. print('ausc_rmse', np.trapz((hist_pred_rmses - hist_oracle_rmses), x = list(np.arange(start=0.0, stop=1.0, step=(1/100))))) print('ausc_abrel', np.trapz((hist_pred_absrels - hist_oracle_absrels), x = list(np.arange(start=0.0, stop=1.0, step=(1/100))))) print('ausc_rmse', spar_rmse) print('ausc_abrel', spar_absr) elapsed_time = time.time() - start_time print('Elapesed time: %s' % str(elapsed_time)) print('Done.') if __name__ == '__main__': test()

py

LDU

LDU-main/monocular_depth_estimation/pytorch/sparsification.py

import numpy as np import torch """Calculate the sparsification error. Calcualte the sparsification error for a given array according to a reference array. Args: unc_tensor: Flatten estimated uncertainty tensor. pred_tensor: Flatten depth prediction tensor. gt_tensor: Flatten ground truth tensor. nb_bins: Number of bins using for uncertainty estimation. Each time, 1/nb_bins * 100% items with highest value will be removed. return_hist: if return histograms for drawing the sparsification curve, otherwise, directly return the sum of sparsification error. Returns: By default, sum of the sparsification error after removing all the items in two given vectors given nb_bins. Given return_hist = True, three arrays corresponding to the components of sparsification curve. """ def sparsification_error_gpu(unc_tensor, pred_tensor, gt_tensor, nb_bins = 100, return_hist=True, is_rmse = True): hist_pred = [] hist_oracle = [] nb_remain = [] # From small to big argsorted_U = torch.argsort(unc_tensor) err_tensor = abs(pred_tensor - gt_tensor) if not is_rmse: err_tensor = err_tensor/gt_tensor else: err_tensor = err_tensor**2 argsorted_E = torch.argsort(err_tensor) total_len = len(unc_tensor) sigma_pred_curves = [] error_curves = [] fractions = list(torch.arange(start=0.0, end=1.0, step=(1/nb_bins))) for fraction in fractions: if is_rmse: sigma_pred_curve = torch.mean(err_tensor[argsorted_U[0:int((1.0-fraction)*total_len)]]) error_curve = torch.mean(err_tensor[argsorted_E[0:int((1.0-fraction)*total_len)]]) sigma_pred_curve = torch.sqrt(sigma_pred_curve) error_curve = torch.sqrt(error_curve) else: sigma_pred_curve = torch.mean(err_tensor[argsorted_U[0:int((1.0-fraction)*total_len)]]) error_curve = torch.mean(err_tensor[argsorted_E[0:int((1.0-fraction)*total_len)]]) sigma_pred_curves.append(sigma_pred_curve) error_curves.append(error_curve) nb_remain.append(int((1.0-fraction)*total_len)) hist_oracle = torch.tensor(error_curves)/error_curves[0].cpu() hist_pred = torch.tensor(sigma_pred_curves)/sigma_pred_curves[0].cpu() nb_remain = torch.tensor(nb_remain) sparsification_errors_pred = torch.trapz((hist_pred - hist_oracle), torch.arange(start=0.0, end=1.0, step=(1/nb_bins))) # without normalization. in our paper we use the codes shown above. # hist_oracle = torch.tensor(error_curves) # hist_pred = torch.tensor(sigma_pred_curves) # nb_remain = torch.tensor(nb_remain) # sparsification_errors_pred = torch.trapz((hist_pred), torch.arange(start=0.0, end=1.0, step=(1/nb_bins))) - torch.trapz((hist_oracle), torch.arange(start=0.0, end=1.0, step=(1/nb_bins))) if return_hist: return hist_pred, hist_oracle, nb_remain, sparsification_errors_pred else: return sparsification_errors_pred

py

LDU

LDU-main/monocular_depth_estimation/pytorch/bts_dataloader.py

# Copyright (C) 2019 Jin Han Lee # # This file is a part of BTS. # This program is free software: you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation, either version 3 of the License, or # (at your option) any later version. # # This program is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with this program. If not, see <http://www.gnu.org/licenses/> import numpy as np import torch from torch.utils.data import Dataset, DataLoader import torch.utils.data.distributed from torchvision import transforms from PIL import Image import os import random from distributed_sampler_no_evenly_divisible import * def _is_pil_image(img): return isinstance(img, Image.Image) def _is_numpy_image(img): return isinstance(img, np.ndarray) and (img.ndim in {2, 3}) def preprocessing_transforms(mode): return transforms.Compose([ ToTensor(mode=mode) ]) class BtsDataLoader(object): def __init__(self, args, mode): if mode == 'train': self.training_samples = DataLoadPreprocess(args, mode, transform=preprocessing_transforms(mode)) if args.distributed: self.train_sampler = torch.utils.data.distributed.DistributedSampler(self.training_samples) else: self.train_sampler = None self.data = DataLoader(self.training_samples, args.batch_size, shuffle=(self.train_sampler is None), num_workers=args.num_threads, pin_memory=True, sampler=self.train_sampler) elif mode == 'online_eval': self.testing_samples = DataLoadPreprocess(args, mode, transform=preprocessing_transforms(mode)) if args.distributed: # self.eval_sampler = torch.utils.data.distributed.DistributedSampler(self.testing_samples, shuffle=False) self.eval_sampler = DistributedSamplerNoEvenlyDivisible(self.testing_samples, shuffle=False) else: self.eval_sampler = None self.data = DataLoader(self.testing_samples, 1, shuffle=False, num_workers=1, pin_memory=True, sampler=self.eval_sampler) elif mode == 'test': self.testing_samples = DataLoadPreprocess(args, mode, transform=preprocessing_transforms(mode)) self.data = DataLoader(self.testing_samples, 1, shuffle=False, num_workers=1) else: print('mode should be one of \'train, test, online_eval\'. Got {}'.format(mode)) class DataLoadPreprocess(Dataset): def __init__(self, args, mode, transform=None, is_for_online_eval=False): self.args = args if mode == 'online_eval': with open(args.filenames_file_eval, 'r') as f: self.filenames = f.readlines() else: with open(args.filenames_file, 'r') as f: self.filenames = f.readlines() self.mode = mode self.transform = transform self.to_tensor = ToTensor self.is_for_online_eval = is_for_online_eval def __getitem__(self, idx): sample_path = self.filenames[idx] focal = float(sample_path.split()[2]) if self.mode == 'train': if self.args.dataset == 'kitti' and self.args.use_right is True and random.random() > 0.5: image_path = os.path.join(self.args.data_path, "./" + sample_path.split()[3]) depth_path = os.path.join(self.args.gt_path, "./" + sample_path.split()[4]) else: image_path = os.path.join(self.args.data_path, "./" + sample_path.split()[0]) depth_path = os.path.join(self.args.gt_path, "./" + sample_path.split()[1]) image = Image.open(image_path) depth_gt = Image.open(depth_path) if self.args.do_kb_crop is True: height = image.height width = image.width top_margin = int(height - 352) left_margin = int((width - 1216) / 2) depth_gt = depth_gt.crop((left_margin, top_margin, left_margin + 1216, top_margin + 352)) image = image.crop((left_margin, top_margin, left_margin + 1216, top_margin + 352)) # To avoid blank boundaries due to pixel registration if self.args.dataset == 'nyu': depth_gt = depth_gt.crop((43, 45, 608, 472)) image = image.crop((43, 45, 608, 472)) if self.args.do_random_rotate is True: random_angle = (random.random() - 0.5) * 2 * self.args.degree image = self.rotate_image(image, random_angle) depth_gt = self.rotate_image(depth_gt, random_angle, flag=Image.NEAREST) image = np.asarray(image, dtype=np.float32) / 255.0 depth_gt = np.asarray(depth_gt, dtype=np.float32) depth_gt = np.expand_dims(depth_gt, axis=2) if self.args.dataset == 'nyu': depth_gt = depth_gt / 1000.0 else: depth_gt = depth_gt / 256.0 image, depth_gt = self.random_crop(image, depth_gt, self.args.input_height, self.args.input_width) image, depth_gt = self.train_preprocess(image, depth_gt) sample = {'image': image, 'depth': depth_gt, 'focal': focal} else: if self.mode == 'online_eval': data_path = self.args.data_path_eval else: data_path = self.args.data_path image_path = os.path.join(data_path, "./" + sample_path.split()[0]) image = np.asarray(Image.open(image_path), dtype=np.float32) / 255.0 if self.mode == 'online_eval': gt_path = self.args.gt_path_eval depth_path = os.path.join(gt_path, "./" + sample_path.split()[1]) has_valid_depth = False try: depth_gt = Image.open(depth_path) has_valid_depth = True except IOError: depth_gt = False # print('Missing gt for {}'.format(image_path)) if has_valid_depth: depth_gt = np.asarray(depth_gt, dtype=np.float32) depth_gt = np.expand_dims(depth_gt, axis=2) if self.args.dataset == 'nyu': depth_gt = depth_gt / 1000.0 else: depth_gt = depth_gt / 256.0 if self.args.do_kb_crop is True: height = image.shape[0] width = image.shape[1] top_margin = int(height - 352) left_margin = int((width - 1216) / 2) image = image[top_margin:top_margin + 352, left_margin:left_margin + 1216, :] if self.mode == 'online_eval' and has_valid_depth: depth_gt = depth_gt[top_margin:top_margin + 352, left_margin:left_margin + 1216, :] if self.mode == 'online_eval': sample = {'image': image, 'depth': depth_gt, 'focal': focal, 'has_valid_depth': has_valid_depth} else: sample = {'image': image, 'focal': focal} if self.transform: sample = self.transform(sample) return sample def rotate_image(self, image, angle, flag=Image.BILINEAR): result = image.rotate(angle, resample=flag) return result def random_crop(self, img, depth, height, width): assert img.shape[0] >= height assert img.shape[1] >= width assert img.shape[0] == depth.shape[0] assert img.shape[1] == depth.shape[1] x = random.randint(0, img.shape[1] - width) y = random.randint(0, img.shape[0] - height) img = img[y:y + height, x:x + width, :] depth = depth[y:y + height, x:x + width, :] return img, depth def train_preprocess(self, image, depth_gt): # Random flipping do_flip = random.random() if do_flip > 0.5: image = (image[:, ::-1, :]).copy() depth_gt = (depth_gt[:, ::-1, :]).copy() # Random gamma, brightness, color augmentation do_augment = random.random() if do_augment > 0.5: image = self.augment_image(image) return image, depth_gt def augment_image(self, image): # gamma augmentation gamma = random.uniform(0.9, 1.1) image_aug = image ** gamma # brightness augmentation if self.args.dataset == 'nyu': brightness = random.uniform(0.75, 1.25) else: brightness = random.uniform(0.9, 1.1) image_aug = image_aug * brightness # color augmentation colors = np.random.uniform(0.9, 1.1, size=3) white = np.ones((image.shape[0], image.shape[1])) color_image = np.stack([white * colors[i] for i in range(3)], axis=2) image_aug *= color_image image_aug = np.clip(image_aug, 0, 1) return image_aug def __len__(self): return len(self.filenames) class ToTensor(object): def __init__(self, mode): self.mode = mode self.normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) def __call__(self, sample): image, focal = sample['image'], sample['focal'] image = self.to_tensor(image) image = self.normalize(image) if self.mode == 'test': return {'image': image, 'focal': focal} depth = sample['depth'] if self.mode == 'train': depth = self.to_tensor(depth) return {'image': image, 'depth': depth, 'focal': focal} else: has_valid_depth = sample['has_valid_depth'] return {'image': image, 'depth': depth, 'focal': focal, 'has_valid_depth': has_valid_depth} def to_tensor(self, pic): if not (_is_pil_image(pic) or _is_numpy_image(pic)): raise TypeError( 'pic should be PIL Image or ndarray. Got {}'.format(type(pic))) if isinstance(pic, np.ndarray): img = torch.from_numpy(pic.transpose((2, 0, 1))) return img # handle PIL Image if pic.mode == 'I': img = torch.from_numpy(np.array(pic, np.int32, copy=False)) elif pic.mode == 'I;16': img = torch.from_numpy(np.array(pic, np.int16, copy=False)) else: img = torch.ByteTensor(torch.ByteStorage.from_buffer(pic.tobytes())) # PIL image mode: 1, L, P, I, F, RGB, YCbCr, RGBA, CMYK if pic.mode == 'YCbCr': nchannel = 3 elif pic.mode == 'I;16': nchannel = 1 else: nchannel = len(pic.mode) img = img.view(pic.size[1], pic.size[0], nchannel) img = img.transpose(0, 1).transpose(0, 2).contiguous() if isinstance(img, torch.ByteTensor): return img.float() else: return img

py

LDU

LDU-main/monocular_depth_estimation/pytorch/bts_eval.py

# Copyright (C) 2019 Jin Han Lee # # This file is a part of BTS. # This program is free software: you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation, either version 3 of the License, or # (at your option) any later version. # # This program is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with this program. If not, see <http://www.gnu.org/licenses/> from __future__ import absolute_import, division, print_function import os import argparse import time import numpy as np import cv2 import sys import torch import torch.nn as nn import torch.nn.utils as utils import torchvision.utils as vutils import torch.backends.cudnn as cudnn from torch.autograd import Variable from tensorboardX import SummaryWriter from bts_dataloader import * def convert_arg_line_to_args(arg_line): for arg in arg_line.split(): if not arg.strip(): continue yield arg parser = argparse.ArgumentParser(description='BTS PyTorch implementation.', fromfile_prefix_chars='@') parser.convert_arg_line_to_args = convert_arg_line_to_args parser.add_argument('--model_name', type=str, help='model name', default='bts_v0_0_1') parser.add_argument('--encoder', type=str, help='type of encoder, desenet121_bts or densenet161_bts', default='densenet161_bts') parser.add_argument('--data_path', type=str, help='path to the data', required=True) parser.add_argument('--gt_path', type=str, help='path to the groundtruth data', required=False) parser.add_argument('--filenames_file', type=str, help='path to the filenames text file', required=True) parser.add_argument('--input_height', type=int, help='input height', default=480) parser.add_argument('--input_width', type=int, help='input width', default=640) parser.add_argument('--max_depth', type=float, help='maximum depth in estimation', default=80) parser.add_argument('--output_directory', type=str, help='output directory for summary, if empty outputs to checkpoint folder', default='') parser.add_argument('--checkpoint_path', type=str, help='path to a specific checkpoint to load', default='') parser.add_argument('--dataset', type=str, help='dataset to train on, make3d or nyudepthv2', default='nyu') parser.add_argument('--eigen_crop', help='if set, crops according to Eigen NIPS14', action='store_true') parser.add_argument('--garg_crop', help='if set, crops according to Garg ECCV16', action='store_true') parser.add_argument('--min_depth_eval', type=float, help='minimum depth for evaluation', default=1e-3) parser.add_argument('--max_depth_eval', type=float, help='maximum depth for evaluation', default=80) parser.add_argument('--do_kb_crop', help='if set, crop input images as kitti benchmark images', action='store_true') parser.add_argument('--bts_size', type=int, help='initial num_filters in bts', default=512) if sys.argv.__len__() == 2: arg_filename_with_prefix = '@' + sys.argv[1] args = parser.parse_args([arg_filename_with_prefix]) else: args = parser.parse_args() model_dir = os.path.dirname(args.checkpoint_path) sys.path.append(model_dir) for key, val in vars(__import__(args.model_name)).items(): if key.startswith('__') and key.endswith('__'): continue vars()[key] = val def compute_errors(gt, pred): thresh = np.maximum((gt / pred), (pred / gt)) d1 = (thresh < 1.25).mean() d2 = (thresh < 1.25 ** 2).mean() d3 = (thresh < 1.25 ** 3).mean() rmse = (gt - pred) ** 2 rmse = np.sqrt(rmse.mean()) rmse_log = (np.log(gt) - np.log(pred)) ** 2 rmse_log = np.sqrt(rmse_log.mean()) abs_rel = np.mean(np.abs(gt - pred) / gt) sq_rel = np.mean(((gt - pred) ** 2) / gt) err = np.log(pred) - np.log(gt) silog = np.sqrt(np.mean(err ** 2) - np.mean(err) ** 2) * 100 err = np.abs(np.log10(pred) - np.log10(gt)) log10 = np.mean(err) return silog, log10, abs_rel, sq_rel, rmse, rmse_log, d1, d2, d3 def get_num_lines(file_path): f = open(file_path, 'r') lines = f.readlines() f.close() return len(lines) def test(params): global gt_depths, is_missing, missing_ids gt_depths = [] is_missing = [] missing_ids = set() write_summary = False steps = set() if os.path.isdir(args.checkpoint_path): import glob models = [f for f in glob.glob(args.checkpoint_path + "/model*")] for model in models: step = model.split('-')[-1] steps.add('{:06d}'.format(int(step))) lines = [] if os.path.exists(args.checkpoint_path + '/evaluated_checkpoints'): with open(args.checkpoint_path + '/evaluated_checkpoints') as file: lines = file.readlines() for line in lines: if line.rstrip() in steps: steps.remove(line.rstrip()) steps = sorted(steps) if args.output_directory != '': summary_path = os.path.join(args.output_directory, args.model_name) else: summary_path = os.path.join(args.checkpoint_path, 'eval') write_summary = True else: steps.add('{:06d}'.format(int(args.checkpoint_path.split('-')[-1]))) if len(steps) == 0: print('No new model to evaluate. Abort.') return args.mode = 'test' dataloader = BtsDataLoader(args, 'eval') model = BtsModel(params=params) model = torch.nn.DataParallel(model) cudnn.benchmark = True if write_summary: summary_writer = SummaryWriter(summary_path, flush_secs=30) for step in steps: if os.path.isdir(args.checkpoint_path): checkpoint = torch.load(os.path.join(args.checkpoint_path, 'model-' + str(int(step)))) model.load_state_dict(checkpoint['model']) else: checkpoint = torch.load(args.checkpoint_path) model.load_state_dict(checkpoint['model']) model.eval() model.cuda() num_test_samples = get_num_lines(args.filenames_file) with open(args.filenames_file) as f: lines = f.readlines() print('now testing {} files for step {}'.format(num_test_samples, step)) pred_depths = [] start_time = time.time() with torch.no_grad(): for _, sample in enumerate(dataloader.data): image = Variable(sample['image'].cuda()) focal = Variable(sample['focal'].cuda()) # image = Variable(sample['image']) # focal = Variable(sample['focal']) # Predict lpg8x8, lpg4x4, lpg2x2, reduc1x1, depth_est = model(image, focal) pred_depths.append(depth_est.cpu().numpy().squeeze()) elapsed_time = time.time() - start_time print('Elapesed time: %s' % str(elapsed_time)) print('Done.') if len(gt_depths) == 0: for t_id in range(num_test_samples): gt_depth_path = os.path.join(args.gt_path, lines[t_id].split()[1]) depth = cv2.imread(gt_depth_path, -1) if depth is None: print('Missing: %s ' % gt_depth_path) missing_ids.add(t_id) continue if args.dataset == 'nyu': depth = depth.astype(np.float32) / 1000.0 else: depth = depth.astype(np.float32) / 256.0 gt_depths.append(depth) print('Computing errors') silog, log10, abs_rel, sq_rel, rms, log_rms, d1, d2, d3 = eval(pred_depths, int(step)) if write_summary: summary_writer.add_scalar('silog', silog.mean(), int(step)) summary_writer.add_scalar('abs_rel', abs_rel.mean(), int(step)) summary_writer.add_scalar('log10', log10.mean(), int(step)) summary_writer.add_scalar('sq_rel', sq_rel.mean(), int(step)) summary_writer.add_scalar('rms', rms.mean(), int(step)) summary_writer.add_scalar('log_rms', log_rms.mean(), int(step)) summary_writer.add_scalar('d1', d1.mean(), int(step)) summary_writer.add_scalar('d2', d2.mean(), int(step)) summary_writer.add_scalar('d3', d3.mean(), int(step)) summary_writer.flush() with open(os.path.dirname(args.checkpoint_path) + '/evaluated_checkpoints', 'a') as file: file.write(step + '\n') print('Evaluation done') def eval(pred_depths, step): num_samples = get_num_lines(args.filenames_file) pred_depths_valid = [] for t_id in range(num_samples): if t_id in missing_ids: continue pred_depths_valid.append(pred_depths[t_id]) num_samples = num_samples - len(missing_ids) silog = np.zeros(num_samples, np.float32) log10 = np.zeros(num_samples, np.float32) rms = np.zeros(num_samples, np.float32) log_rms = np.zeros(num_samples, np.float32) abs_rel = np.zeros(num_samples, np.float32) sq_rel = np.zeros(num_samples, np.float32) d1 = np.zeros(num_samples, np.float32) d2 = np.zeros(num_samples, np.float32) d3 = np.zeros(num_samples, np.float32) for i in range(num_samples): gt_depth = gt_depths[i] pred_depth = pred_depths_valid[i] if args.do_kb_crop: height, width = gt_depth.shape top_margin = int(height - 352) left_margin = int((width - 1216) / 2) pred_depth_uncropped = np.zeros((height, width), dtype=np.float32) pred_depth_uncropped[top_margin:top_margin + 352, left_margin:left_margin + 1216] = pred_depth pred_depth = pred_depth_uncropped pred_depth[pred_depth < args.min_depth_eval] = args.min_depth_eval pred_depth[pred_depth > args.max_depth_eval] = args.max_depth_eval pred_depth[np.isinf(pred_depth)] = args.max_depth_eval pred_depth[np.isnan(pred_depth)] = args.min_depth_eval valid_mask = np.logical_and(gt_depth > args.min_depth_eval, gt_depth < args.max_depth_eval) if args.garg_crop or args.eigen_crop: gt_height, gt_width = gt_depth.shape eval_mask = np.zeros(valid_mask.shape) if args.garg_crop: eval_mask[int(0.40810811 * gt_height):int(0.99189189 * gt_height), int(0.03594771 * gt_width):int(0.96405229 * gt_width)] = 1 elif args.eigen_crop: if args.dataset == 'kitti': eval_mask[int(0.3324324 * gt_height):int(0.91351351 * gt_height), int(0.0359477 * gt_width):int(0.96405229 * gt_width)] = 1 else: eval_mask[45:471, 41:601] = 1 valid_mask = np.logical_and(valid_mask, eval_mask) silog[i], log10[i], abs_rel[i], sq_rel[i], rms[i], log_rms[i], d1[i], d2[i], d3[i] = compute_errors( gt_depth[valid_mask], pred_depth[valid_mask]) print("{:>7}, {:>7}, {:>7}, {:>7}, {:>7}, {:>7}, {:>7}, {:>7}, {:>7}".format('silog', 'abs_rel', 'log10', 'rms', 'sq_rel', 'log_rms', 'd1', 'd2', 'd3')) print("{:7.4f}, {:7.4f}, {:7.3f}, {:7.3f}, {:7.3f}, {:7.3f}, {:7.3f}, {:7.3f}, {:7.3f}".format( silog.mean(), abs_rel.mean(), log10.mean(), rms.mean(), sq_rel.mean(), log_rms.mean(), d1.mean(), d2.mean(), d3.mean())) return silog, log10, abs_rel, sq_rel, rms, log_rms, d1, d2, d3 if __name__ == '__main__': test(args)

py

LDU

LDU-main/monocular_depth_estimation/pytorch/bts_test.py

# Copyright (C) 2019 Jin Han Lee # # This file is a part of BTS. # This program is free software: you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation, either version 3 of the License, or # (at your option) any later version. # # This program is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with this program. If not, see <http://www.gnu.org/licenses/> from __future__ import absolute_import, division, print_function import os import argparse import time import numpy as np import cv2 import sys import torch import torch.nn as nn from torch.autograd import Variable from bts_dataloader import * import errno import matplotlib.pyplot as plt from tqdm import tqdm from bts_dataloader import * def convert_arg_line_to_args(arg_line): for arg in arg_line.split(): if not arg.strip(): continue yield arg parser = argparse.ArgumentParser(description='BTS PyTorch implementation.', fromfile_prefix_chars='@') parser.convert_arg_line_to_args = convert_arg_line_to_args parser.add_argument('--model_name', type=str, help='model name', default='bts_nyu_v2') parser.add_argument('--encoder', type=str, help='type of encoder, vgg or desenet121_bts or densenet161_bts', default='densenet161_bts') parser.add_argument('--data_path', type=str, help='path to the data', required=True) parser.add_argument('--filenames_file', type=str, help='path to the filenames text file', required=True) parser.add_argument('--input_height', type=int, help='input height', default=480) parser.add_argument('--input_width', type=int, help='input width', default=640) parser.add_argument('--max_depth', type=float, help='maximum depth in estimation', default=80) parser.add_argument('--checkpoint_path', type=str, help='path to a specific checkpoint to load', default='') parser.add_argument('--dataset', type=str, help='dataset to train on, make3d or nyudepthv2', default='nyu') parser.add_argument('--do_kb_crop', help='if set, crop input images as kitti benchmark images', action='store_true') parser.add_argument('--save_lpg', help='if set, save outputs from lpg layers', action='store_true') parser.add_argument('--bts_size', type=int, help='initial num_filters in bts', default=512) if sys.argv.__len__() == 2: arg_filename_with_prefix = '@' + sys.argv[1] args = parser.parse_args([arg_filename_with_prefix]) else: args = parser.parse_args() model_dir = os.path.dirname(args.checkpoint_path) sys.path.append(model_dir) for key, val in vars(__import__(args.model_name)).items(): if key.startswith('__') and key.endswith('__'): continue vars()[key] = val def get_num_lines(file_path): f = open(file_path, 'r') lines = f.readlines() f.close() return len(lines) def test(params): """Test function.""" args.mode = 'test' dataloader = BtsDataLoader(args, 'test') model = BtsModel(params=args) model = torch.nn.DataParallel(model) checkpoint = torch.load(args.checkpoint_path) model.load_state_dict(checkpoint['model']) model.eval() model.cuda() num_params = sum([np.prod(p.size()) for p in model.parameters()]) print("Total number of parameters: {}".format(num_params)) num_test_samples = get_num_lines(args.filenames_file) with open(args.filenames_file) as f: lines = f.readlines() print('now testing {} files with {}'.format(num_test_samples, args.checkpoint_path)) pred_depths = [] pred_8x8s = [] pred_4x4s = [] pred_2x2s = [] pred_1x1s = [] start_time = time.time() with torch.no_grad(): for _, sample in enumerate(tqdm(dataloader.data)): image = Variable(sample['image'].cuda()) focal = Variable(sample['focal'].cuda()) # Predict lpg8x8, lpg4x4, lpg2x2, reduc1x1, depth_est = model(image, focal) pred_depths.append(depth_est.cpu().numpy().squeeze()) pred_8x8s.append(lpg8x8[0].cpu().numpy().squeeze()) pred_4x4s.append(lpg4x4[0].cpu().numpy().squeeze()) pred_2x2s.append(lpg2x2[0].cpu().numpy().squeeze()) pred_1x1s.append(reduc1x1[0].cpu().numpy().squeeze()) elapsed_time = time.time() - start_time print('Elapesed time: %s' % str(elapsed_time)) print('Done.') save_name = 'result_' + args.model_name print('Saving result pngs..') if not os.path.exists(os.path.dirname(save_name)): try: os.mkdir(save_name) os.mkdir(save_name + '/raw') os.mkdir(save_name + '/cmap') os.mkdir(save_name + '/rgb') os.mkdir(save_name + '/gt') except OSError as e: if e.errno != errno.EEXIST: raise for s in tqdm(range(num_test_samples)): if args.dataset == 'kitti': date_drive = lines[s].split('/')[1] filename_pred_png = save_name + '/raw/' + date_drive + '_' + lines[s].split()[0].split('/')[-1].replace( '.jpg', '.png') filename_cmap_png = save_name + '/cmap/' + date_drive + '_' + lines[s].split()[0].split('/')[ -1].replace('.jpg', '.png') filename_image_png = save_name + '/rgb/' + date_drive + '_' + lines[s].split()[0].split('/')[-1] elif args.dataset == 'kitti_benchmark': filename_pred_png = save_name + '/raw/' + lines[s].split()[0].split('/')[-1].replace('.jpg', '.png') filename_cmap_png = save_name + '/cmap/' + lines[s].split()[0].split('/')[-1].replace('.jpg', '.png') filename_image_png = save_name + '/rgb/' + lines[s].split()[0].split('/')[-1] else: scene_name = lines[s].split()[0].split('/')[0] filename_pred_png = save_name + '/raw/' + scene_name + '_' + lines[s].split()[0].split('/')[1].replace( '.jpg', '.png') filename_cmap_png = save_name + '/cmap/' + scene_name + '_' + lines[s].split()[0].split('/')[1].replace( '.jpg', '.png') filename_gt_png = save_name + '/gt/' + scene_name + '_' + lines[s].split()[0].split('/')[1].replace( '.jpg', '.png') filename_image_png = save_name + '/rgb/' + scene_name + '_' + lines[s].split()[0].split('/')[1] rgb_path = os.path.join(args.data_path, './' + lines[s].split()[0]) image = cv2.imread(rgb_path) if args.dataset == 'nyu': gt_path = os.path.join(args.data_path, './' + lines[s].split()[1]) gt = cv2.imread(gt_path, -1).astype(np.float32) / 1000.0 # Visualization purpose only gt[gt == 0] = np.amax(gt) pred_depth = pred_depths[s] pred_8x8 = pred_8x8s[s] pred_4x4 = pred_4x4s[s] pred_2x2 = pred_2x2s[s] pred_1x1 = pred_1x1s[s] if args.dataset == 'kitti' or args.dataset == 'kitti_benchmark': pred_depth_scaled = pred_depth * 256.0 else: pred_depth_scaled = pred_depth * 1000.0 pred_depth_scaled = pred_depth_scaled.astype(np.uint16) cv2.imwrite(filename_pred_png, pred_depth_scaled, [cv2.IMWRITE_PNG_COMPRESSION, 0]) if args.save_lpg: cv2.imwrite(filename_image_png, image[10:-1 - 9, 10:-1 - 9, :]) if args.dataset == 'nyu': plt.imsave(filename_gt_png, np.log10(gt[10:-1 - 9, 10:-1 - 9]), cmap='Greys') pred_depth_cropped = pred_depth[10:-1 - 9, 10:-1 - 9] plt.imsave(filename_cmap_png, np.log10(pred_depth_cropped), cmap='Greys') pred_8x8_cropped = pred_8x8[10:-1 - 9, 10:-1 - 9] filename_lpg_cmap_png = filename_cmap_png.replace('.png', '_8x8.png') plt.imsave(filename_lpg_cmap_png, np.log10(pred_8x8_cropped), cmap='Greys') pred_4x4_cropped = pred_4x4[10:-1 - 9, 10:-1 - 9] filename_lpg_cmap_png = filename_cmap_png.replace('.png', '_4x4.png') plt.imsave(filename_lpg_cmap_png, np.log10(pred_4x4_cropped), cmap='Greys') pred_2x2_cropped = pred_2x2[10:-1 - 9, 10:-1 - 9] filename_lpg_cmap_png = filename_cmap_png.replace('.png', '_2x2.png') plt.imsave(filename_lpg_cmap_png, np.log10(pred_2x2_cropped), cmap='Greys') pred_1x1_cropped = pred_1x1[10:-1 - 9, 10:-1 - 9] filename_lpg_cmap_png = filename_cmap_png.replace('.png', '_1x1.png') plt.imsave(filename_lpg_cmap_png, np.log10(pred_1x1_cropped), cmap='Greys') else: plt.imsave(filename_cmap_png, np.log10(pred_depth), cmap='Greys') filename_lpg_cmap_png = filename_cmap_png.replace('.png', '_8x8.png') plt.imsave(filename_lpg_cmap_png, np.log10(pred_8x8), cmap='Greys') filename_lpg_cmap_png = filename_cmap_png.replace('.png', '_4x4.png') plt.imsave(filename_lpg_cmap_png, np.log10(pred_4x4), cmap='Greys') filename_lpg_cmap_png = filename_cmap_png.replace('.png', '_2x2.png') plt.imsave(filename_lpg_cmap_png, np.log10(pred_2x2), cmap='Greys') filename_lpg_cmap_png = filename_cmap_png.replace('.png', '_1x1.png') plt.imsave(filename_lpg_cmap_png, np.log10(pred_1x1), cmap='Greys') return if __name__ == '__main__': test(args)

py

LDU

LDU-main/monocular_depth_estimation/pytorch/bts.py

# Copyright (C) 2019 Jin Han Lee # # This file is a part of BTS. # This program is free software: you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation, either version 3 of the License, or # (at your option) any later version. # # This program is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with this program. If not, see <http://www.gnu.org/licenses/> import torch import torch.nn as nn import torch.nn.functional as torch_nn_func import math from collections import namedtuple # This sets the batch norm layers in pytorch as if {'is_training': False, 'scale': True} in tensorflow def bn_init_as_tf(m): if isinstance(m, nn.BatchNorm2d): m.track_running_stats = True # These two lines enable using stats (moving mean and var) loaded from pretrained model m.eval() # or zero mean and variance of one if the batch norm layer has no pretrained values m.affine = True m.requires_grad = True def weights_init_xavier(m): if isinstance(m, nn.Conv2d): torch.nn.init.xavier_uniform_(m.weight) if m.bias is not None: torch.nn.init.zeros_(m.bias) class silog_loss(nn.Module): def __init__(self, variance_focus): super(silog_loss, self).__init__() self.variance_focus = variance_focus def forward(self, depth_est, depth_gt, mask): d = torch.log(depth_est[mask]) - torch.log(depth_gt[mask]) return torch.sqrt((d ** 2).mean() - self.variance_focus * (d.mean() ** 2)) * 10.0 class atrous_conv(nn.Sequential): def __init__(self, in_channels, out_channels, dilation, apply_bn_first=True): super(atrous_conv, self).__init__() self.atrous_conv = torch.nn.Sequential() if apply_bn_first: self.atrous_conv.add_module('first_bn', nn.BatchNorm2d(in_channels, momentum=0.01, affine=True, track_running_stats=True, eps=1.1e-5)) self.atrous_conv.add_module('aconv_sequence', nn.Sequential(nn.ReLU(), nn.Conv2d(in_channels=in_channels, out_channels=out_channels*2, bias=False, kernel_size=1, stride=1, padding=0), nn.BatchNorm2d(out_channels*2, momentum=0.01, affine=True, track_running_stats=True), nn.ReLU(), nn.Conv2d(in_channels=out_channels * 2, out_channels=out_channels, bias=False, kernel_size=3, stride=1, padding=(dilation, dilation), dilation=dilation))) def forward(self, x): return self.atrous_conv.forward(x) class upconv(nn.Module): def __init__(self, in_channels, out_channels, ratio=2): super(upconv, self).__init__() self.elu = nn.ELU() self.conv = nn.Conv2d(in_channels=in_channels, out_channels=out_channels, bias=False, kernel_size=3, stride=1, padding=1) self.ratio = ratio def forward(self, x): up_x = torch_nn_func.interpolate(x, scale_factor=self.ratio, mode='nearest') out = self.conv(up_x) out = self.elu(out) return out class reduction_1x1(nn.Sequential): def __init__(self, num_in_filters, num_out_filters, max_depth, is_final=False): super(reduction_1x1, self).__init__() self.max_depth = max_depth self.is_final = is_final self.sigmoid = nn.Sigmoid() self.reduc = torch.nn.Sequential() while num_out_filters >= 4: if num_out_filters < 8: if self.is_final: self.reduc.add_module('final', torch.nn.Sequential(nn.Conv2d(num_in_filters, out_channels=1, bias=False, kernel_size=1, stride=1, padding=0), nn.Sigmoid())) else: self.reduc.add_module('plane_params', torch.nn.Conv2d(num_in_filters, out_channels=3, bias=False, kernel_size=1, stride=1, padding=0)) break else: self.reduc.add_module('inter_{}_{}'.format(num_in_filters, num_out_filters), torch.nn.Sequential(nn.Conv2d(in_channels=num_in_filters, out_channels=num_out_filters, bias=False, kernel_size=1, stride=1, padding=0), nn.ELU())) num_in_filters = num_out_filters num_out_filters = num_out_filters // 2 def forward(self, net): net = self.reduc.forward(net) if not self.is_final: theta = self.sigmoid(net[:, 0, :, :]) * math.pi / 3 phi = self.sigmoid(net[:, 1, :, :]) * math.pi * 2 dist = self.sigmoid(net[:, 2, :, :]) * self.max_depth n1 = torch.mul(torch.sin(theta), torch.cos(phi)).unsqueeze(1) n2 = torch.mul(torch.sin(theta), torch.sin(phi)).unsqueeze(1) n3 = torch.cos(theta).unsqueeze(1) n4 = dist.unsqueeze(1) net = torch.cat([n1, n2, n3, n4], dim=1) return net class local_planar_guidance(nn.Module): def __init__(self, upratio): super(local_planar_guidance, self).__init__() self.upratio = upratio self.u = torch.arange(self.upratio).reshape([1, 1, self.upratio]).float() self.v = torch.arange(int(self.upratio)).reshape([1, self.upratio, 1]).float() self.upratio = float(upratio) def forward(self, plane_eq, focal): plane_eq_expanded = torch.repeat_interleave(plane_eq, int(self.upratio), 2) plane_eq_expanded = torch.repeat_interleave(plane_eq_expanded, int(self.upratio), 3) n1 = plane_eq_expanded[:, 0, :, :] n2 = plane_eq_expanded[:, 1, :, :] n3 = plane_eq_expanded[:, 2, :, :] n4 = plane_eq_expanded[:, 3, :, :] u = self.u.repeat(plane_eq.size(0), plane_eq.size(2) * int(self.upratio), plane_eq.size(3)).cuda() u = (u - (self.upratio - 1) * 0.5) / self.upratio v = self.v.repeat(plane_eq.size(0), plane_eq.size(2), plane_eq.size(3) * int(self.upratio)).cuda() v = (v - (self.upratio - 1) * 0.5) / self.upratio return n4 / (n1 * u + n2 * v + n3) class bts(nn.Module): def __init__(self, params, feat_out_channels, num_features=512): super(bts, self).__init__() self.params = params self.upconv5 = upconv(feat_out_channels[4], num_features) self.bn5 = nn.BatchNorm2d(num_features, momentum=0.01, affine=True, eps=1.1e-5) self.conv5 = torch.nn.Sequential(nn.Conv2d(num_features + feat_out_channels[3], num_features, 3, 1, 1, bias=False), nn.ELU()) self.upconv4 = upconv(num_features, num_features // 2) self.bn4 = nn.BatchNorm2d(num_features // 2, momentum=0.01, affine=True, eps=1.1e-5) self.conv4 = torch.nn.Sequential(nn.Conv2d(num_features // 2 + feat_out_channels[2], num_features // 2, 3, 1, 1, bias=False), nn.ELU()) self.bn4_2 = nn.BatchNorm2d(num_features // 2, momentum=0.01, affine=True, eps=1.1e-5) self.daspp_3 = atrous_conv(num_features // 2, num_features // 4, 3, apply_bn_first=False) self.daspp_6 = atrous_conv(num_features // 2 + num_features // 4 + feat_out_channels[2], num_features // 4, 6) self.daspp_12 = atrous_conv(num_features + feat_out_channels[2], num_features // 4, 12) self.daspp_18 = atrous_conv(num_features + num_features // 4 + feat_out_channels[2], num_features // 4, 18) self.daspp_24 = atrous_conv(num_features + num_features // 2 + feat_out_channels[2], num_features // 4, 24) self.daspp_conv = torch.nn.Sequential(nn.Conv2d(num_features + num_features // 2 + num_features // 4, num_features // 4, 3, 1, 1, bias=False), nn.ELU()) self.reduc8x8 = reduction_1x1(num_features // 4, num_features // 4, self.params.max_depth) self.lpg8x8 = local_planar_guidance(8) self.upconv3 = upconv(num_features // 4, num_features // 4) self.bn3 = nn.BatchNorm2d(num_features // 4, momentum=0.01, affine=True, eps=1.1e-5) self.conv3 = torch.nn.Sequential(nn.Conv2d(num_features // 4 + feat_out_channels[1] + 1, num_features // 4, 3, 1, 1, bias=False), nn.ELU()) self.reduc4x4 = reduction_1x1(num_features // 4, num_features // 8, self.params.max_depth) self.lpg4x4 = local_planar_guidance(4) self.upconv2 = upconv(num_features // 4, num_features // 8) self.bn2 = nn.BatchNorm2d(num_features // 8, momentum=0.01, affine=True, eps=1.1e-5) self.conv2 = torch.nn.Sequential(nn.Conv2d(num_features // 8 + feat_out_channels[0] + 1, num_features // 8, 3, 1, 1, bias=False), nn.ELU()) self.reduc2x2 = reduction_1x1(num_features // 8, num_features // 16, self.params.max_depth) self.lpg2x2 = local_planar_guidance(2) self.upconv1 = upconv(num_features // 8, num_features // 16) self.reduc1x1 = reduction_1x1(num_features // 16, num_features // 32, self.params.max_depth, is_final=True) self.conv1 = torch.nn.Sequential(nn.Conv2d(num_features // 16 + 4, num_features // 16, 3, 1, 1, bias=False), nn.ELU()) self.get_depth = torch.nn.Sequential(nn.Conv2d(num_features // 16, 1, 3, 1, 1, bias=False), nn.Sigmoid()) def forward(self, features, focal): skip0, skip1, skip2, skip3 = features[0], features[1], features[2], features[3] dense_features = torch.nn.ReLU()(features[4]) upconv5 = self.upconv5(dense_features) # H/16 upconv5 = self.bn5(upconv5) concat5 = torch.cat([upconv5, skip3], dim=1) iconv5 = self.conv5(concat5) upconv4 = self.upconv4(iconv5) # H/8 upconv4 = self.bn4(upconv4) concat4 = torch.cat([upconv4, skip2], dim=1) iconv4 = self.conv4(concat4) iconv4 = self.bn4_2(iconv4) daspp_3 = self.daspp_3(iconv4) concat4_2 = torch.cat([concat4, daspp_3], dim=1) daspp_6 = self.daspp_6(concat4_2) concat4_3 = torch.cat([concat4_2, daspp_6], dim=1) daspp_12 = self.daspp_12(concat4_3) concat4_4 = torch.cat([concat4_3, daspp_12], dim=1) daspp_18 = self.daspp_18(concat4_4) concat4_5 = torch.cat([concat4_4, daspp_18], dim=1) daspp_24 = self.daspp_24(concat4_5) concat4_daspp = torch.cat([iconv4, daspp_3, daspp_6, daspp_12, daspp_18, daspp_24], dim=1) daspp_feat = self.daspp_conv(concat4_daspp) reduc8x8 = self.reduc8x8(daspp_feat) plane_normal_8x8 = reduc8x8[:, :3, :, :] plane_normal_8x8 = torch_nn_func.normalize(plane_normal_8x8, 2, 1) plane_dist_8x8 = reduc8x8[:, 3, :, :] plane_eq_8x8 = torch.cat([plane_normal_8x8, plane_dist_8x8.unsqueeze(1)], 1) depth_8x8 = self.lpg8x8(plane_eq_8x8, focal) depth_8x8_scaled = depth_8x8.unsqueeze(1) / self.params.max_depth depth_8x8_scaled_ds = torch_nn_func.interpolate(depth_8x8_scaled, scale_factor=0.25, mode='nearest') upconv3 = self.upconv3(daspp_feat) # H/4 upconv3 = self.bn3(upconv3) concat3 = torch.cat([upconv3, skip1, depth_8x8_scaled_ds], dim=1) iconv3 = self.conv3(concat3) reduc4x4 = self.reduc4x4(iconv3) plane_normal_4x4 = reduc4x4[:, :3, :, :] plane_normal_4x4 = torch_nn_func.normalize(plane_normal_4x4, 2, 1) plane_dist_4x4 = reduc4x4[:, 3, :, :] plane_eq_4x4 = torch.cat([plane_normal_4x4, plane_dist_4x4.unsqueeze(1)], 1) depth_4x4 = self.lpg4x4(plane_eq_4x4, focal) depth_4x4_scaled = depth_4x4.unsqueeze(1) / self.params.max_depth depth_4x4_scaled_ds = torch_nn_func.interpolate(depth_4x4_scaled, scale_factor=0.5, mode='nearest') upconv2 = self.upconv2(iconv3) # H/2 upconv2 = self.bn2(upconv2) concat2 = torch.cat([upconv2, skip0, depth_4x4_scaled_ds], dim=1) iconv2 = self.conv2(concat2) reduc2x2 = self.reduc2x2(iconv2) plane_normal_2x2 = reduc2x2[:, :3, :, :] plane_normal_2x2 = torch_nn_func.normalize(plane_normal_2x2, 2, 1) plane_dist_2x2 = reduc2x2[:, 3, :, :] plane_eq_2x2 = torch.cat([plane_normal_2x2, plane_dist_2x2.unsqueeze(1)], 1) depth_2x2 = self.lpg2x2(plane_eq_2x2, focal) depth_2x2_scaled = depth_2x2.unsqueeze(1) / self.params.max_depth upconv1 = self.upconv1(iconv2) reduc1x1 = self.reduc1x1(upconv1) concat1 = torch.cat([upconv1, reduc1x1, depth_2x2_scaled, depth_4x4_scaled, depth_8x8_scaled], dim=1) iconv1 = self.conv1(concat1) final_depth = self.params.max_depth * self.get_depth(iconv1) if self.params.dataset == 'kitti': final_depth = final_depth * focal.view(-1, 1, 1, 1).float() / 715.0873 return depth_8x8_scaled, depth_4x4_scaled, depth_2x2_scaled, reduc1x1, final_depth class encoder(nn.Module): def __init__(self, params): super(encoder, self).__init__() self.params = params import torchvision.models as models if params.encoder == 'densenet121_bts': self.base_model = models.densenet121(pretrained=True).features self.feat_names = ['relu0', 'pool0', 'transition1', 'transition2', 'norm5'] self.feat_out_channels = [64, 64, 128, 256, 1024] elif params.encoder == 'densenet161_bts': self.base_model = models.densenet161(pretrained=True).features self.feat_names = ['relu0', 'pool0', 'transition1', 'transition2', 'norm5'] self.feat_out_channels = [96, 96, 192, 384, 2208] elif params.encoder == 'resnet50_bts': self.base_model = models.resnet50(pretrained=True) self.feat_names = ['relu', 'layer1', 'layer2', 'layer3', 'layer4'] self.feat_out_channels = [64, 256, 512, 1024, 2048] elif params.encoder == 'resnet101_bts': self.base_model = models.resnet101(pretrained=True) self.feat_names = ['relu', 'layer1', 'layer2', 'layer3', 'layer4'] self.feat_out_channels = [64, 256, 512, 1024, 2048] elif params.encoder == 'resnext50_bts': self.base_model = models.resnext50_32x4d(pretrained=True) self.feat_names = ['relu', 'layer1', 'layer2', 'layer3', 'layer4'] self.feat_out_channels = [64, 256, 512, 1024, 2048] elif params.encoder == 'resnext101_bts': self.base_model = models.resnext101_32x8d(pretrained=True) self.feat_names = ['relu', 'layer1', 'layer2', 'layer3', 'layer4'] self.feat_out_channels = [64, 256, 512, 1024, 2048] elif params.encoder == 'mobilenetv2_bts': self.base_model = models.mobilenet_v2(pretrained=True).features self.feat_inds = [2, 4, 7, 11, 19] self.feat_out_channels = [16, 24, 32, 64, 1280] self.feat_names = [] else: print('Not supported encoder: {}'.format(params.encoder)) def forward(self, x): feature = x skip_feat = [] i = 1 for k, v in self.base_model._modules.items(): if 'fc' in k or 'avgpool' in k: continue feature = v(feature) if self.params.encoder == 'mobilenetv2_bts': if i == 2 or i == 4 or i == 7 or i == 11 or i == 19: skip_feat.append(feature) else: if any(x in k for x in self.feat_names): skip_feat.append(feature) i = i + 1 return skip_feat class BtsModel(nn.Module): def __init__(self, params): super(BtsModel, self).__init__() self.encoder = encoder(params) self.decoder = bts(params, self.encoder.feat_out_channels, params.bts_size) def forward(self, x, focal): skip_feat = self.encoder(x) return self.decoder(skip_feat, focal)

py

DelayResolvedRL

DelayResolvedRL-main/Gym(Stochastic)/agent.py

import tensorflow as tf import numpy as np import random import copy from statistics import mean from collections import deque GPUs = tf.config.experimental.list_physical_devices('GPU') if GPUs: try: for gpu in GPUs: tf.config.experimental.set_memory_growth(gpu, True) except RuntimeError as e: print(e) def to_onehot(size, value): """1 hot encoding for observed state""" return np.eye(size)[value] class Model(tf.keras.Model): """DQN Model""" def __init__(self, num_states, hidden_units, num_actions, alg, use_stochastic_delay, max_dimension): super(Model, self).__init__() if alg == 'IS': if use_stochastic_delay: self.input_layer = tf.keras.layers.InputLayer(input_shape=(num_states + 1 + max_dimension,)) else: self.input_layer = tf.keras.layers.InputLayer(input_shape=(num_states + max_dimension,)) else: self.input_layer = tf.keras.layers.InputLayer(input_shape=(num_states,)) self.hidden_layers = [] for i in hidden_units: self.hidden_layers.append(tf.keras.layers.Dense( i, activation='tanh', kernel_initializer='RandomNormal')) self.output_layer = tf.keras.layers.Dense( num_actions, activation='linear', kernel_initializer='RandomNormal') @tf.function def call(self, inputs): z = self.input_layer(inputs) for layer in self.hidden_layers: z = layer(z) output = self.output_layer(z) return output class DQN: def __init__(self, num_states, num_actions, model_params, alg_params): np.random.seed(alg_params['seed']) tf.random.set_seed(alg_params['seed']) random.seed(alg_params['seed']) self.num_actions = num_actions self.alg = alg_params['algorithm'] self.batch_size = alg_params['batch_size'] self.optimizer = tf.optimizers.Adam(alg_params['learning_rate']) self.use_stochastic_delay = alg_params['use_stochastic_delay'] self.max_dimension = model_params['max_dimension'] hidden_units = model_params['hidden_units'] self.delay = alg_params['delay'] self.gamma = alg_params['gamma'] self.model = Model(num_states, hidden_units, num_actions, self.use_stochastic_delay, self.max_dimension, self.alg) self.experience = {'s': [], 'a': [], 'r': [], 's2': [], 'done': []} self.max_experiences = model_params['max_buffer_size'] self.min_experiences = model_params['min_buffer_size'] if self.alg != 'normal': self.action_buffer = deque(maxlen=self.max_dimension + 1) self.action_buffer_padded = deque(maxlen=self.max_dimension + 1) def predict(self, inputs): return self.model(np.atleast_2d(inputs.astype('float32'))) def fill_up_buffer(self): self.action_buffer_padded.clear() for _ in range(self.max_dimension): self.action_buffer_padded.append(0) def buffer_padding(self): current_length = len(self.action_buffer) self.action_buffer_padded = copy.deepcopy(self.action_buffer) for _ in range(0, self.max_dimension - current_length): self.action_buffer_padded.append(0) def train(self, TargetNet): if len(self.experience['s']) < self.min_experiences: return 0 ids = np.random.randint(low=0, high=len(self.experience['s']), size=self.batch_size) states = np.asarray([self.experience['s'][i] for i in ids]) actions = np.asarray([self.experience['a'][i] for i in ids]) rewards = np.asarray([self.experience['r'][i] for i in ids]) states_next = np.asarray([self.experience['s2'][i] for i in ids]) dones = np.asarray([self.experience['done'][i] for i in ids]) value_next = np.max(TargetNet.predict(states_next), axis=1) actual_values = np.where(dones, rewards, rewards + self.gamma * value_next) with tf.GradientTape() as tape: selected_action_values = tf.math.reduce_sum( self.predict(states) * tf.one_hot(actions, self.num_actions), axis=1) loss = tf.math.reduce_mean(tf.square(actual_values - selected_action_values)) variables = self.model.trainable_variables gradients = tape.gradient(loss, variables) self.optimizer.apply_gradients(zip(gradients, variables)) return loss def get_action(self, states, epsilon): if np.random.random() < epsilon: return np.random.choice(self.num_actions) else: return np.argmax(self.predict(np.atleast_2d(states))[0]) def add_experience(self, exp): if len(self.experience['s']) >= self.max_experiences: for key in self.experience.keys(): self.experience[key].pop(0) for key, value in exp.items(): self.experience[key].append(value) def copy_weights(self, TrainNet): variables1 = self.model.trainable_variables variables2 = TrainNet.model.trainable_variables for v1, v2 in zip(variables1, variables2): v1.assign(v2.numpy()) def play_game(global_step, env, TrainNet, TargetNet, epsilon, copy_step): rewards = 0 episode_step = 0 last_state_observed = 0 done = False observations = env.reset() observations_original = observations if env.game_name.startswith('Frozen'): observations = to_onehot(env.state_space.n, observations) if TrainNet.alg != 'normal': TrainNet.fill_up_buffer() losses = list() clear = False while not done: delay = env.delay len_buffer = len(env.state_buffer) if TrainNet.alg == 'normal': action = TrainNet.get_action(observations, epsilon) prev_observations = observations observations, reward, done = env.step(observations_original, action) observations_original = observations if env.game_name.startswith('Frozen'): observations = to_onehot(env.state_space.n, observations) else: if episode_step == 0: if env.use_stochastic_delay: last_state_observed = (episode_step - env.turn_limit / 2) / env.turn_limit action_state = np.append(last_state_observed, TrainNet.action_buffer_padded) information_state = np.append(observations, action_state) # information_state = np.append(observations, TrainNet.action_buffer_padded) else: information_state = np.append(observations, TrainNet.action_buffer_padded) if TrainNet.alg == 'IS': action = TrainNet.get_action(information_state, epsilon) else: action = TrainNet.get_action(observations, epsilon) prev_observations = observations prev_information_state = information_state observations, reward, done = env.step(observations_original, action) observations_original = observations if env.game_name.startswith('Frozen'): observations = to_onehot(env.state_space.n, observations) episode_step += 1 if env.train: last_state_observed = (episode_step - 1 - env.turn_limit / 2) / env.turn_limit TrainNet.action_buffer.append(action + 1) for i in range(len_buffer + 1 - delay): TrainNet.action_buffer.popleft() - 1 TrainNet.buffer_padding() else: # delayed_action = random.randint(0, TrainNet.num_actions) TrainNet.action_buffer.append(action + 1) TrainNet.buffer_padding() if env.delay == 0: delayed_action = action else: if not TrainNet.action_buffer: delayed_action = random.randint(0, TrainNet.num_actions) else: delayed_action = TrainNet.action_buffer[0] if delay == 0: delayed_action = action if len(TrainNet.action_buffer) == TrainNet.max_dimension + 1: TrainNet.action_buffer.clear() TrainNet.buffer_padding() observations = env.state_buffer.pop() env.state_buffer.clear() reward = np.sum(env.reward_buffer) done = env.done_buffer.pop() env.done_buffer.clear() env.reward_buffer.clear() clear = True if env.use_stochastic_delay: action_state = np.append(last_state_observed, TrainNet.action_buffer_padded) information_state = np.append(observations, action_state) # information_state = np.append(observations, TrainNet.action_buffer_padded) else: information_state = np.append(observations, TrainNet.action_buffer_padded) rewards += reward if done: episode_step = 0 env.reset() if TrainNet.alg != 'normal': TrainNet.action_buffer.clear() TrainNet.buffer_padding() global_step += 1 if TrainNet.alg == 'normal': exp = {'s': prev_observations, 'a': action, 'r': reward, 's2': observations, 'done': done} if TrainNet.alg == 'delay': exp = {'s': prev_observations, 'a': delayed_action, 'r': reward, 's2': observations, 'done': done} if TrainNet.alg == 'IS': exp = {'s': prev_information_state, 'a': action, 'r': reward, 's2': information_state, 'done': done} TrainNet.add_experience(exp) loss = TrainNet.train(TargetNet) if isinstance(loss, int): losses.append(loss) else: losses.append(loss.numpy()) if global_step % copy_step == 0: TargetNet.copy_weights(TrainNet) return global_step, rewards, mean(losses) def test(env, TrainNet, logs, num_episodes): for _ in range(num_episodes): observation = env.reset() rewards = 0 steps = 0 done = False while not done: action = TrainNet.get_action(observation, 0) observation, reward, done, _ = env.step(action) steps += 1 rewards += reward with open(logs['log_file_name'], "a") as f: print("Testing steps: {} rewards :{} ".format(steps, rewards), file=f) print("Testing steps: {} rewards :{} ".format(steps, rewards)) def train_agent(env, num_frames, model_params, algorithm_params, logs, verbose): num_actions = env.number_of_actions try: state_space = len(env.state_space.sample()) except TypeError: state_space = env.state_space.n copy_step = model_params['copy_step'] TrainNet = DQN(state_space, num_actions, model_params, algorithm_params) TargetNet = DQN(state_space, num_actions, model_params, algorithm_params) # N = num_episodes total_rewards_list = [] total_losses_list = [] epsilon_start = algorithm_params['start_epsilon'] decay = algorithm_params['epsilon_decay'] min_epsilon = algorithm_params['stop_epsilon'] global_step = 1 n = 0 while True: epsilon = min_epsilon + (epsilon_start - min_epsilon) * np.exp(-decay * global_step) global_step, total_reward, losses = play_game(global_step, env, TrainNet, TargetNet, epsilon, copy_step) total_rewards_list.append(total_reward) total_losses_list.append(losses) total_rewards = np.array(total_rewards_list) total_losses = np.array(total_losses_list) avg_rewards = total_rewards[max(0, n - 100):(n + 1)].mean() avg_losses = total_losses[max(0, n - 100):(n + 1)].mean() if n % logs['log_interval'] == 0: if verbose: with open(logs['log_file_name'], "a") as f: print("episode:{}, eps:{:.3f}, avg reward (last 100):{:.2f}, avg loss:{:.2f}" .format(n, epsilon, avg_rewards, avg_losses), file=f) if not verbose: print("episode:{}, eps:{:.3f}, avg reward (last 100):{:.2f}" .format(n, epsilon, avg_rewards)) # test(env, TrainNet, logs, 100) n += 1 if global_step > num_frames: break env.close() return total_rewards, total_losses

py

DelayResolvedRL

DelayResolvedRL-main/W-Maze/DQN/agent.py

import tensorflow as tf import numpy as np import random import copy from statistics import mean from collections import deque GPUs = tf.config.experimental.list_physical_devices('GPU') if GPUs: try: for gpu in GPUs: tf.config.experimental.set_memory_growth(gpu, True) except RuntimeError as e: print(e) '''DQN Model''' class Model(tf.keras.Model): def __init__(self, state_space_shape, hidden_units, num_actions, use_stochastic_delay, max_dimension, alg): super(Model, self).__init__() input_shape = state_space_shape.ndim if alg == 'IS': if use_stochastic_delay: self.input_layer = tf.keras.layers.InputLayer(input_shape=(input_shape + 1 + max_dimension,)) else: self.input_layer = tf.keras.layers.InputLayer(input_shape=(input_shape + max_dimension,)) else: self.input_layer = tf.keras.layers.InputLayer(input_shape=(input_shape,)) self.hidden_layers = [] for i in hidden_units: self.hidden_layers.append(tf.keras.layers.Dense( i, activation='tanh', kernel_initializer='RandomNormal')) self.output_layer = tf.keras.layers.Dense( num_actions, activation='linear', kernel_initializer='RandomNormal') @tf.function def call(self, inputs): z = self.input_layer(inputs) for layer in self.hidden_layers: z = layer(z) output = self.output_layer(z) return output class DQN: def __init__(self, state_space_shape, num_actions, model_params, alg_params): self.num_actions = num_actions self.actions = np.linspace(1, self.num_actions, num=self.num_actions, dtype=np.int32) self.alg = alg_params['algorithm'] self.batch_size = alg_params['batch_size'] self.optimizer = tf.optimizers.Adam(alg_params['learning_rate']) self.delay = alg_params['delay'] self.gamma = alg_params['gamma'] self.use_stochastic_delay = alg_params['use_stochastic_delay'] self.max_dimension = model_params['max_dimension'] hidden_units = model_params['hidden_units'] self.model = Model(state_space_shape, hidden_units, num_actions, self.use_stochastic_delay, self.max_dimension, self.alg) self.experience = {'s': [], 'a': [], 'r': [], 's2': [], 'done': []} self.max_experiences = model_params['max_buffer_size'] self.min_experiences = model_params['min_buffer_size'] if self.alg != 'normal': self.action_buffer = deque(maxlen=self.max_dimension+1) self.action_buffer_padded = deque(maxlen=self.max_dimension+1) def predict(self, inputs): return self.model(np.atleast_2d(inputs.astype('float32'))) def fill_up_buffer(self): self.action_buffer_padded.clear() for _ in range(self.max_dimension): self.action_buffer_padded.append(0) def buffer_padding(self): current_length = len(self.action_buffer) self.action_buffer_padded = copy.deepcopy(self.action_buffer) for _ in range(0, self.max_dimension - current_length): self.action_buffer_padded.append(0) def train(self, TargetNet): if len(self.experience['s']) < self.min_experiences: return 0 ids = np.random.randint(low=0, high=len(self.experience['s']), size=self.batch_size) states = np.asarray([self.experience['s'][i] for i in ids]) actions = np.asarray([self.experience['a'][i] for i in ids]) rewards = np.asarray([self.experience['r'][i] for i in ids]) states_next = np.asarray([self.experience['s2'][i] for i in ids]) dones = np.asarray([self.experience['done'][i] for i in ids]) value_next = np.max(TargetNet.predict(states_next), axis=1) actual_values = np.where(dones, rewards, rewards + self.gamma * value_next) with tf.GradientTape() as tape: selected_action_values = tf.math.reduce_sum( self.predict(states) * tf.one_hot(actions, self.num_actions), axis=1) loss = tf.math.reduce_mean(tf.square(actual_values - selected_action_values)) variables = self.model.trainable_variables gradients = tape.gradient(loss, variables) self.optimizer.apply_gradients(zip(gradients, variables)) return loss def get_action(self, states, epsilon): if np.random.random() < epsilon: return np.random.choice(self.num_actions) else: return np.argmax(self.predict(np.atleast_2d(states))[0]) def add_experience(self, exp): if len(self.experience['s']) >= self.max_experiences: for key in self.experience.keys(): self.experience[key].pop(0) for key, value in exp.items(): self.experience[key].append(value) def copy_weights(self, TrainNet): variables1 = self.model.trainable_variables variables2 = TrainNet.model.trainable_variables for v1, v2 in zip(variables1, variables2): v1.assign(v2.numpy()) def play_game(global_step, env, TrainNet, TargetNet, epsilon, copy_step): rewards = 0 episode_step = 0 last_state_observed = 0 done = False observations = env.reset() if TrainNet.alg != 'normal': TrainNet.fill_up_buffer() losses = list() clear = False while not done: delay = env.delay len_buffer = len(env.state_buffer) if TrainNet.alg == 'normal': action = TrainNet.get_action(observations, epsilon) prev_observations = observations observations, reward, done = env.step(observations, action) else: if episode_step == 0: if env.use_stochastic_delay: # append the last time this state was observed normalized by the max step of the episode last_state_observed = (episode_step-env.turn_limit/2)/env.turn_limit action_state = np.append(last_state_observed, TrainNet.action_buffer_padded) information_state = np.append(observations, action_state) else: information_state = np.append(observations, TrainNet.action_buffer_padded) if TrainNet.alg == 'IS': action = TrainNet.get_action(information_state, epsilon) else: action = TrainNet.get_action(observations, epsilon) prev_observations = observations prev_information_state = information_state observations, reward, done = env.step(observations, action) episode_step += 1 if env.delay == 0: delayed_action = action else: if not TrainNet.action_buffer: # buffer empty delayed_action = random.randint(0, TrainNet.num_actions) else: delayed_action = TrainNet.action_buffer[0] if env.train: last_state_observed = (episode_step-env.turn_limit/2)/env.turn_limit TrainNet.action_buffer.append(action + 1) for i in range(len_buffer + 1 - delay): TrainNet.action_buffer.popleft() - 1 TrainNet.buffer_padding() else: TrainNet.action_buffer.append(action + 1) TrainNet.buffer_padding() if len(TrainNet.action_buffer) == TrainNet.max_dimension+1: TrainNet.action_buffer.clear() TrainNet.buffer_padding() observations = env.state_buffer.pop() env.state_buffer.clear() reward = np.sum(env.reward_buffer) done = env.done_buffer.pop() env.done_buffer.clear() env.reward_buffer.clear() clear = True if env.use_stochastic_delay: # append the last time this state was observed normalized by the max step of the episode action_state = np.append(last_state_observed, TrainNet.action_buffer_padded) information_state = np.append(observations, action_state) else: information_state = np.append(observations, TrainNet.action_buffer_padded) rewards += reward if done: episode_step = 0 env.reset() if TrainNet.alg != 'normal': TrainNet.action_buffer.clear() TrainNet.buffer_padding() global_step += 1 if TrainNet.alg == 'normal': exp = {'s': prev_observations, 'a': action, 'r': reward, 's2': observations, 'done': done} if TrainNet.alg == 'delay': exp = {'s': prev_observations, 'a': delayed_action, 'r': reward, 's2': observations, 'done': done} if TrainNet.alg == 'IS': exp = {'s': prev_information_state, 'a': action, 'r': reward, 's2': information_state, 'done': done} TrainNet.add_experience(exp) loss = TrainNet.train(TargetNet) if isinstance(loss, int): losses.append(loss) else: losses.append(loss.numpy()) if global_step % copy_step == 0: TargetNet.copy_weights(TrainNet) return global_step, rewards, mean(losses) def test(env, TrainNet, logs, num_episodes): for _ in range(num_episodes): observation = env.reset() rewards = 0 steps = 0 done = False while not done: action = TrainNet.get_action(observation, 0) observation, reward, done, _ = env.step(action) steps += 1 rewards += reward with open(logs['log_file_name'], "a") as f: print("Testing steps: {} rewards :{} ".format(steps, rewards), file=f) print("Testing steps: {} rewards :{} ".format(steps, rewards)) def train_agent(env, num_frames, model_params, algorithm_params, logs, verbose): np.random.seed(algorithm_params['seed']) tf.random.set_seed(algorithm_params['seed']) random.seed(algorithm_params['seed']) num_actions = env.number_of_actions state_space = env.state_space.shape copy_step = model_params['copy_step'] TrainNet = DQN(state_space, num_actions, model_params, algorithm_params) TargetNet = DQN(state_space, num_actions, model_params, algorithm_params) # N = num_episodes total_rewards_list = [] total_losses_list = [] epsilon_start = algorithm_params['start_epsilon'] decay = algorithm_params['epsilon_decay'] min_epsilon = algorithm_params['stop_epsilon'] global_step = 1 n = 0 while True: epsilon = min_epsilon + (epsilon_start - min_epsilon) * np.exp(-decay * global_step) global_step, total_reward, losses = play_game(global_step, env, TrainNet, TargetNet, epsilon, copy_step) total_rewards_list.append(total_reward) total_losses_list.append(losses) total_rewards = np.array(total_rewards_list) total_losses = np.array(total_losses_list) avg_rewards = total_rewards[max(0, n - 100):(n + 1)].mean() avg_losses = total_losses[max(0, n - 100):(n + 1)].mean() if n % logs['log_interval'] == 0: if verbose: with open(logs['log_file_name'], "a") as f: print("episode:{}, eps:{:.4f}, avg reward (last 100):{:.2f}, avg loss:{:.2f}" .format(n, epsilon, avg_rewards, avg_losses), file=f) if not verbose: print("episode:{}, eps:{:.3f}, avg reward (last 100):{:.2f}" .format(n, epsilon, avg_rewards)) # test(env, TrainNet, logs, 100) n += 1 if global_step > num_frames: break # env.close() return total_rewards, total_losses

py

DelayResolvedRL

DelayResolvedRL-main/Gym(Constant)/dqn_agents.py

from collections import deque from keras.models import Sequential from keras.layers import Dense, Conv2D, MaxPool2D, Flatten from copy import deepcopy import random from keras.optimizers import Adam from keras import backend as K import tensorflow as tf import numpy as np def reshape_state(state, is_atari_env, state_size): reshaped = state if not is_atari_env: reshaped = np.reshape(state, [1, state_size]) else: if len(state.shape) < 4: reshaped = np.expand_dims(state, axis=0) return reshaped def update_loss(loss, sample_loss): if loss is not None and sample_loss is not None: for key, val in sample_loss.items(): if key in loss: loss[key] += val else: loss[key] = val def concatenate_state_action(state, action): out = np.concatenate((state[0], [action])) out = np.reshape(out, [1, len(out)]) return out class DQNAgent: def __init__(self, seed, state_size, action_size, is_atari_env, is_delayed_agent=False, delay_value=0, epsilon_min=0.001, epsilon_decay=0.999, learning_rate=0.001, epsilon=1.0, use_m_step_reward=False, use_latest_reward=True, loss='mse', **kwargs): np.random.seed(seed) tf.random.set_seed(seed) random.seed(seed) self.state_size = state_size self.action_size = action_size self.is_atari_env = is_atari_env mem_len = 50000 if self.is_atari_env else 1000 self.memory = deque(maxlen=mem_len) self.gamma = 0.99 # discount rate self.epsilon = epsilon # exploration rate self.epsilon_min = epsilon_min self.epsilon_decay = epsilon_decay self.learning_rate = learning_rate self.sample_buffer = deque() self.is_delayed_agent = is_delayed_agent self.delay_value = delay_value self.model = self._build_model(loss=loss) self.use_m_step_reward = use_m_step_reward self.use_latest_reward = use_latest_reward def _huber_loss(self, y_true, y_pred, clip_delta=1.0): """Huber loss for Q Learning References: https://en.wikipedia.org/wiki/Huber_loss https://www.tensorflow.org/api_docs/python/tf/losses/huber_loss """ error = y_true - y_pred cond = K.abs(error) <= clip_delta squared_loss = 0.5 * K.square(error) quadratic_loss = 0.5 * K.square(clip_delta) + clip_delta * (K.abs(error) - clip_delta) return K.mean(tf.where(cond, squared_loss, quadratic_loss)) def _build_forward_model(self, loss='mse', input_size=None, output_size=None): input_size = self.state_size if input_size is None else input_size output_size = self.action_size if output_size is None else output_size model = Sequential() model.add(Dense(200, input_dim=input_size, activation='relu')) model.add(Dense(200, activation='relu')) model.add(Dense(output_size, activation='linear')) model.compile(loss=loss, optimizer=Adam(lr=self.learning_rate)) return model def _build_model(self, loss=None, input_size=None, output_size=None): loss = self._huber_loss if loss is 'huber' else loss input_size = self.state_size if input_size is None else input_size output_size = self.action_size if output_size is None else output_size # Neural Net for Deep-Q learning Model model = Sequential() if self.is_atari_env: model.add(Conv2D(32, 8, strides=(4,4), input_shape=input_size, activation='relu')) model.add(MaxPool2D()) model.add(Conv2D(64, 4, strides=(2,2), activation='relu')) model.add(MaxPool2D()) model.add(Conv2D(64, 3, strides=(1,1), activation='relu')) model.add(MaxPool2D()) model.add(Flatten()) model.add(Dense(64, activation='relu')) model.add(Dense(64, activation='relu')) model.add(Dense(output_size, activation='linear')) else: # model.add(Dense(24, input_dim=input_size, activation='relu')) # model.add(Dense(24, activation='relu')) # model.add(Dense(output_size, activation='linear')) model.add(Dense(200, input_dim=input_size, activation='tanh', kernel_initializer='RandomNormal')) # model.add(Dense(200, activation='tanh')) model.add(Dense(output_size, activation='linear', kernel_initializer='RandomNormal')) model.compile(loss=loss, optimizer=Adam(lr=self.learning_rate)) return model def memorize(self, state, action, reward, next_state, done): if self.is_delayed_agent: # for earlier time than delay_value, the data is problematic (non-delayed response) # Construct modified tuple by keeping old s_t with new a_{t+m}, r_{t+m} s_{t+m+1} new_tuple = (state, action, reward, next_state, done) self.sample_buffer.append(new_tuple) if len(self.sample_buffer) - 1 >= self.delay_value: old_tuple = self.sample_buffer.popleft() modified_tuple = list(deepcopy(old_tuple)) modified_tuple[1] = action modified_tuple[2] = self.m_step_reward(first_reward=old_tuple[2]) # trying to use s_{t+1} instead of s_{t+m} as in the original ICML2020 submission # modified_tuple[3] = next_state modified_tuple = tuple(modified_tuple) self.memory.append(modified_tuple) else: self.memory.append((state, action, reward, next_state, done)) def act(self, state, eval=False): if not eval and np.random.rand() <= self.epsilon: return random.randrange(self.action_size) act_values = self.model.predict(state) return np.argmax(act_values[0]) # returns action def m_step_reward(self, first_reward): if not self.use_m_step_reward: if self.use_latest_reward: return self.sample_buffer[-1][2] else: return first_reward else: discounted_rew = first_reward for i in range(self.delay_value): discounted_rew += self.gamma ** (i + 1) * self.sample_buffer[i][2] return discounted_rew def effective_gamma(self): return self.gamma if not self.use_m_step_reward else (self.gamma ** (self.delay_value + 1)) def replay(self, batch_size, global_step): minibatch = random.sample(self.memory, batch_size) for state, action, reward, next_state, done in minibatch: target = reward if not done: target = (reward + self.effective_gamma() * np.amax(self.model.predict(next_state)[0])) target_f = self.model.predict(state) target_f[0][action] = target # self.model.fit(state, target_f, epochs=1, verbose=0, # callbacks=[WandbCallback()]) self.model.fit(state, target_f, epochs=1, verbose=0) self.epsilon = self.epsilon_min + (1.0 - self.epsilon_min) * np.exp(-self.epsilon_decay * global_step) def load(self, name): self.model.load_weights(name) def save(self, name): self.model.save_weights(name) def clear_action_buffer(self): self.sample_buffer.clear() class DDQNAgent(DQNAgent): def __init__(self, seed, state_size, action_size, is_atari_env, is_delayed_agent=False, delay_value=0, epsilon_min=0.001, epsilon_decay=0.999, learning_rate=0.001, epsilon=1.0, use_m_step_reward=False, use_latest_reward=True): super().__init__(seed, state_size, action_size, is_atari_env=is_atari_env, is_delayed_agent=is_delayed_agent, delay_value=delay_value, epsilon_min=epsilon_min, epsilon_decay=epsilon_decay, learning_rate=learning_rate, epsilon=epsilon, use_m_step_reward=use_m_step_reward, use_latest_reward=use_latest_reward, loss='huber') # self.model = self._build_model() self.target_model = self._build_model(loss='mse') self.update_target_model() def update_target_model(self): # copy weights from model to target_model self.target_model.set_weights(self.model.get_weights()) def train_model(self, batch): state_vec, action_vec, reward_vec, next_state_vec, done_vec = batch target = self.model.predict(state_vec) t = self.target_model.predict(next_state_vec) not_done_arr = np.invert(np.asarray(done_vec)) new_targets = reward_vec + not_done_arr * self.effective_gamma() * np.amax(t, axis=1) for i in range(len(batch[0])): target[i][action_vec[i]] = new_targets[i] train_history = self.model.fit(state_vec, target, epochs=1, verbose=0) q_loss = train_history.history['loss'][0] loss_dict = {'q_loss': q_loss} return loss_dict def _create_batch(self, indices): state_vec, action_vec, reward_vec, next_state_vec, done_vec = [], [], [], [], [] for i in indices: data = self.memory[i] state, action, reward, next_state, done = data state_vec.append(np.array(state, copy=False)) action_vec.append(action) reward_vec.append(reward) next_state_vec.append(np.array(next_state, copy=False)) done_vec.append(done) return np.concatenate(state_vec, axis=0), action_vec, reward_vec, np.concatenate(next_state_vec, axis=0), done_vec def replay(self, batch_size, global_step): loss = {} indices = np.random.choice(len(self.memory), batch_size) batch = self._create_batch(indices) sample_loss = self.train_model(batch) update_loss(loss, sample_loss) self.epsilon = self.epsilon_min + (1.0 - self.epsilon_min) * np.exp(-self.epsilon_decay * global_step) return loss class DDQNPlanningAgent(DDQNAgent): def __init__(self, seed, state_size, action_size, is_atari_env, is_delayed_agent=False, delay_value=0, epsilon_min=0.001, epsilon_decay=0.999, learning_rate=0.001, epsilon=1.0, use_m_step_reward=False, use_latest_reward=True, env=None, use_learned_forward_model=True): super().__init__(seed, state_size, action_size, is_atari_env=is_atari_env, is_delayed_agent=is_delayed_agent, delay_value=delay_value, epsilon_min=epsilon_min, epsilon_decay=epsilon_decay, learning_rate=learning_rate, epsilon=epsilon, use_m_step_reward=use_m_step_reward, use_latest_reward=use_latest_reward) self.use_learned_forward_model = use_learned_forward_model if self.use_learned_forward_model: keras_forward_model = self._build_forward_model(loss='mse', input_size=self.state_size + 1, output_size=self.state_size) self.forward_model = ForwardModel(keras_forward_model) else: self.forward_model = env def train_model(self, batch): loss_dict = super().train_model(batch) if self.use_learned_forward_model and self.delay_value > 0: state_vec, action_vec, _, next_state_vec, _ = batch act_t = np.asarray([action_vec]).transpose() concat_vec = np.concatenate((state_vec, act_t), axis=1) train_history = self.forward_model.keras_model.fit(concat_vec, next_state_vec, epochs=1, verbose=0) f_model_loss = train_history.history['loss'][0] loss_dict['f_model_loss'] = f_model_loss return loss_dict def act(self, state, pending_actions, eval): if not eval and np.random.rand() <= self.epsilon: return random.randrange(self.action_size) last_state = state if self.delay_value > 0: if not self.use_learned_forward_model: self.forward_model.store_initial_state() # initial_state = deepcopy(state) for curr_action in pending_actions: last_state = self.forward_model.get_next_state(state=last_state, action=curr_action) if not self.use_learned_forward_model: self.forward_model.restore_initial_state() last_state_r = reshape_state(last_state, self.is_atari_env, self.state_size) act_values = self.model.predict(last_state_r) return np.argmax(act_values[0]) # returns best action for last state def memorize(self, state, action, reward, next_state, done): # for earlier time than delay_value, the data is problematic (non-delayed response) # Construct modified tuple by keeping old s_t with new a_{t+m}, r_{t+m} s_{t+m+1} new_tuple = (state, action, reward, next_state, done) self.sample_buffer.append(new_tuple) if len(self.sample_buffer) - 1 >= self.delay_value: old_tuple = self.sample_buffer.popleft() modified_tuple = list(deepcopy(old_tuple)) # build time-coherent tuple from new tuple and old action modified_tuple[0] = state # modified_tuple[1] = action modified_tuple[2] = reward # self.m_step_reward(first_reward=old_tuple[2]) modified_tuple[3] = next_state modified_tuple = tuple(modified_tuple) self.memory.append(modified_tuple) class ForwardModel: def __init__(self, keras_model): self.keras_model = keras_model def get_next_state(self, state, action): input = concatenate_state_action(state, action) return self.keras_model.predict(input) def reset_to_state(self, state): # not necessary here. Only used if the forward_model is the actual env instance pass

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/OpenNMT/translate.py

#!/usr/bin/env python from __future__ import division from builtins import bytes import os import argparse import math import codecs import torch import onmt import onmt.IO import opts from itertools import takewhile, count try: from itertools import zip_longest except ImportError: from itertools import izip_longest as zip_longest parser = argparse.ArgumentParser( description='translate.py', formatter_class=argparse.ArgumentDefaultsHelpFormatter) opts.add_md_help_argument(parser) opts.translate_opts(parser) opt = parser.parse_args() if opt.batch_size != 1: print("WARNING: -batch_size isn't supported currently, " "we set it to 1 for now!") opt.batch_size = 1 def report_score(name, score_total, words_total): print("%s AVG SCORE: %.4f, %s PPL: %.4f" % ( name, score_total / words_total, name, math.exp(-score_total/words_total))) def get_src_words(src_indices, index2str): words = [] raw_words = (index2str[i] for i in src_indices) words = takewhile(lambda w: w != onmt.IO.PAD_WORD, raw_words) return " ".join(words) def main(): dummy_parser = argparse.ArgumentParser(description='train.py') opts.model_opts(dummy_parser) dummy_opt = dummy_parser.parse_known_args([])[0] opt.cuda = opt.gpu > -1 if opt.cuda: torch.cuda.set_device(opt.gpu) translator = onmt.Translator(opt, dummy_opt.__dict__) out_file = codecs.open(opt.output, 'w', 'utf-8') pred_score_total, pred_words_total = 0, 0 gold_score_total, gold_words_total = 0, 0 if opt.dump_beam != "": import json translator.initBeamAccum() data = onmt.IO.ONMTDataset( opt.src, opt.tgt, translator.fields, use_filter_pred=False) test_data = onmt.IO.OrderedIterator( dataset=data, device=opt.gpu, batch_size=opt.batch_size, train=False, sort=False, shuffle=False) counter = count(1) for batch in test_data: pred_batch, gold_batch, pred_scores, gold_scores, attn, src \ = translator.translate(batch, data) pred_score_total += sum(score[0] for score in pred_scores) pred_words_total += sum(len(x[0]) for x in pred_batch) if opt.tgt: gold_score_total += sum(gold_scores) gold_words_total += sum(len(x) for x in batch.tgt[1:]) # z_batch: an iterator over the predictions, their scores, # the gold sentence, its score, and the source sentence for each # sentence in the batch. It has to be zip_longest instead of # plain-old zip because the gold_batch has length 0 if the target # is not included. z_batch = zip_longest( pred_batch, gold_batch, pred_scores, gold_scores, (sent.squeeze(1) for sent in src.split(1, dim=1))) for pred_sents, gold_sent, pred_score, gold_score, src_sent in z_batch: n_best_preds = [" ".join(pred) for pred in pred_sents[:opt.n_best]] out_file.write('\n'.join(n_best_preds)) out_file.write('\n') out_file.flush() if opt.verbose: sent_number = next(counter) words = get_src_words( src_sent, translator.fields["src"].vocab.itos) os.write(1, bytes('\nSENT %d: %s\n' % (sent_number, words), 'UTF-8')) best_pred = n_best_preds[0] best_score = pred_score[0] os.write(1, bytes('PRED %d: %s\n' % (sent_number, best_pred), 'UTF-8')) print("PRED SCORE: %.4f" % best_score) if opt.tgt: tgt_sent = ' '.join(gold_sent) os.write(1, bytes('GOLD %d: %s\n' % (sent_number, tgt_sent), 'UTF-8')) print("GOLD SCORE: %.4f" % gold_score) if len(n_best_preds) > 1: print('\nBEST HYP:') for score, sent in zip(pred_score, n_best_preds): os.write(1, bytes("[%.4f] %s\n" % (score, sent), 'UTF-8')) report_score('PRED', pred_score_total, pred_words_total) if opt.tgt: report_score('GOLD', gold_score_total, gold_words_total) if opt.dump_beam: json.dump(translator.beam_accum, codecs.open(opt.dump_beam, 'w', 'utf-8')) if __name__ == "__main__": main()

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/OpenNMT/train.py

#!/usr/bin/env python from __future__ import division import os import sys import argparse import torch import torch.nn as nn from torch import cuda import onmt import onmt.Models import onmt.ModelConstructor import onmt.modules from onmt.Utils import aeq, use_gpu import opts parser = argparse.ArgumentParser( description='train.py', formatter_class=argparse.ArgumentDefaultsHelpFormatter) # opts.py opts.add_md_help_argument(parser) opts.model_opts(parser) opts.train_opts(parser) opt = parser.parse_args() if opt.word_vec_size != -1: opt.src_word_vec_size = opt.word_vec_size opt.tgt_word_vec_size = opt.word_vec_size if opt.layers != -1: opt.enc_layers = opt.layers opt.dec_layers = opt.layers opt.brnn = (opt.encoder_type == "brnn") if opt.seed > 0: torch.manual_seed(opt.seed) if opt.rnn_type == "SRU" and not opt.gpuid: raise AssertionError("Using SRU requires -gpuid set.") if torch.cuda.is_available() and not opt.gpuid: print("WARNING: You have a CUDA device, should run with -gpuid 0") if opt.gpuid: cuda.set_device(opt.gpuid[0]) if opt.seed > 0: torch.cuda.manual_seed(opt.seed) if len(opt.gpuid) > 1: sys.stderr.write("Sorry, multigpu isn't supported yet, coming soon!\n") sys.exit(1) # Set up the Crayon logging server. if opt.exp_host != "": from pycrayon import CrayonClient cc = CrayonClient(hostname=opt.exp_host) experiments = cc.get_experiment_names() print(experiments) if opt.exp in experiments: cc.remove_experiment(opt.exp) experiment = cc.create_experiment(opt.exp) def report_func(epoch, batch, num_batches, start_time, lr, report_stats): """ This is the user-defined batch-level traing progress report function. Args: epoch(int): current epoch count. batch(int): current batch count. num_batches(int): total number of batches. start_time(float): last report time. lr(float): current learning rate. report_stats(Statistics): old Statistics instance. Returns: report_stats(Statistics): updated Statistics instance. """ if batch % opt.report_every == -1 % opt.report_every: report_stats.output(epoch, batch+1, num_batches, start_time) if opt.exp_host: report_stats.log("progress", experiment, lr) report_stats = onmt.Statistics() return report_stats def make_train_data_iter(train_data, opt): """ This returns user-defined train data iterator for the trainer to iterate over during each train epoch. We implement simple ordered iterator strategy here, but more sophisticated strategy like curriculum learning is ok too. """ return onmt.IO.OrderedIterator( dataset=train_data, batch_size=opt.batch_size, device=opt.gpuid[0] if opt.gpuid else -1, repeat=False) def make_valid_data_iter(valid_data, opt): """ This returns user-defined validate data iterator for the trainer to iterate over during each validate epoch. We implement simple ordered iterator strategy here, but more sophisticated strategy is ok too. """ return onmt.IO.OrderedIterator( dataset=valid_data, batch_size=opt.batch_size, device=opt.gpuid[0] if opt.gpuid else -1, train=False, sort=True) def make_loss_compute(model, tgt_vocab, dataset, opt): """ This returns user-defined LossCompute object, which is used to compute loss in train/validate process. You can implement your own *LossCompute class, by subclassing LossComputeBase. """ if opt.copy_attn: compute = onmt.modules.CopyGeneratorLossCompute( model.generator, tgt_vocab, dataset, opt.copy_attn_force) else: compute = onmt.Loss.NMTLossCompute(model.generator, tgt_vocab) if use_gpu(opt): compute.cuda() return compute def train_model(model, train_data, valid_data, fields, optim): min_ppl, max_accuracy = float('inf'), -1 train_iter = make_train_data_iter(train_data, opt) valid_iter = make_valid_data_iter(valid_data, opt) train_loss = make_loss_compute(model, fields["tgt"].vocab, train_data, opt) valid_loss = make_loss_compute(model, fields["tgt"].vocab, valid_data, opt) trunc_size = opt.truncated_decoder # Badly named... shard_size = opt.max_generator_batches trainer = onmt.Trainer(model, train_iter, valid_iter, train_loss, valid_loss, optim, trunc_size, shard_size) for epoch in range(opt.start_epoch, opt.epochs + 1): print('') # 1. Train for one epoch on the training set. train_stats = trainer.train(epoch, report_func) print('Train perplexity: %g' % train_stats.ppl()) print('Train accuracy: %g' % train_stats.accuracy()) # 2. Validate on the validation set. valid_stats = trainer.validate() print('Validation perplexity: %g' % valid_stats.ppl()) print('Validation accuracy: %g' % valid_stats.accuracy()) # 3. Log to remote server. if opt.exp_host: train_stats.log("train", experiment, optim.lr) valid_stats.log("valid", experiment, optim.lr) # 4. Update the learning rate trainer.epoch_step(valid_stats.ppl(), epoch) # 5. Drop a checkpoint if needed. if epoch >= opt.start_checkpoint_at: if valid_stats.accuracy() > max_accuracy: # 5.1 drop checkpoint when bigger accuracy is achieved. min_ppl = min(valid_stats.ppl(), min_ppl) max_accuracy = max(valid_stats.accuracy(), max_accuracy) trainer.drop_checkpoint(opt, epoch, fields, valid_stats) print('Save model according to biggest-ever accuracy: acc: {0}, ppl: {1}'.format(max_accuracy, min_ppl)) elif valid_stats.ppl() < min_ppl: # 5.2 drop checkpoint when smaller ppl is achieved. min_ppl = min(valid_stats.ppl(), min_ppl) max_accuracy = max(valid_stats.accuracy(), max_accuracy) trainer.drop_checkpoint(opt, epoch, fields, valid_stats) print('Save model according to lowest-ever ppl: acc: {0}, ppl: {1}'.format(max_accuracy, min_ppl)) def check_save_model_path(): save_model_path = os.path.abspath(opt.save_model) model_dirname = os.path.dirname(save_model_path) if not os.path.exists(model_dirname): os.makedirs(model_dirname) def tally_parameters(model): n_params = sum([p.nelement() for p in model.parameters()]) print('* number of parameters: %d' % n_params) enc = 0 dec = 0 for name, param in model.named_parameters(): if 'encoder' in name: enc += param.nelement() elif 'decoder' or 'generator' in name: dec += param.nelement() print('encoder: ', enc) print('decoder: ', dec) def load_fields(train, valid, checkpoint): fields = onmt.IO.load_fields( torch.load(opt.data + '.vocab.pt')) fields = dict([(k, f) for (k, f) in fields.items() if k in train.examples[0].__dict__]) train.fields = fields valid.fields = fields if opt.train_from: print('Loading vocab from checkpoint at %s.' % opt.train_from) fields = onmt.IO.load_fields(checkpoint['vocab']) print(' * vocabulary size. source = %d; target = %d' % (len(fields['src'].vocab), len(fields['tgt'].vocab))) return fields def collect_features(train, fields): # TODO: account for target features. # Also, why does fields need to have the structure it does? src_features = onmt.IO.collect_features(fields) aeq(len(src_features), train.n_src_feats) return src_features def build_model(model_opt, opt, fields, checkpoint): print('Building model...') model = onmt.ModelConstructor.make_base_model(model_opt, fields, use_gpu(opt), checkpoint) if len(opt.gpuid) > 1: print('Multi gpu training: ', opt.gpuid) model = nn.DataParallel(model, device_ids=opt.gpuid, dim=1) print(model) return model def build_optim(model, checkpoint): if opt.train_from: print('Loading optimizer from checkpoint.') optim = checkpoint['optim'] optim.optimizer.load_state_dict( checkpoint['optim'].optimizer.state_dict()) else: # what members of opt does Optim need? optim = onmt.Optim( opt.optim, opt.learning_rate, opt.max_grad_norm, lr_decay=opt.learning_rate_decay, start_decay_at=opt.start_decay_at, opt=opt ) optim.set_parameters(model.parameters()) return optim def main(): # Load train and validate data. print("Loading train and validate data from '%s'" % opt.data) train = torch.load(opt.data + '.train.pt') valid = torch.load(opt.data + '.valid.pt') print(' * number of training sentences: %d' % len(train)) print(' * maximum batch size: %d' % opt.batch_size) # Load checkpoint if we resume from a previous training. if opt.train_from: print('Loading checkpoint from %s' % opt.train_from) checkpoint = torch.load(opt.train_from, map_location=lambda storage, loc: storage) model_opt = checkpoint['opt'] # I don't like reassigning attributes of opt: it's not clear opt.start_epoch = checkpoint['epoch'] + 1 else: checkpoint = None model_opt = opt # Load fields generated from preprocess phase. fields = load_fields(train, valid, checkpoint) # Collect features. src_features = collect_features(train, fields) for j, feat in enumerate(src_features): print(' * src feature %d size = %d' % (j, len(fields[feat].vocab))) # Build model. model = build_model(model_opt, opt, fields, checkpoint) tally_parameters(model) check_save_model_path() # Build optimizer. optim = build_optim(model, checkpoint) # Do training. train_model(model, train, valid, fields, optim) if __name__ == "__main__": main()

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/OpenNMT/preprocess.py

#!/usr/bin/env python # -*- coding: utf-8 -*- import argparse import codecs import torch import onmt import onmt.IO import opts parser = argparse.ArgumentParser( description='preprocess.py', formatter_class=argparse.ArgumentDefaultsHelpFormatter) opts.add_md_help_argument(parser) # **Preprocess Options** parser.add_argument('-config', help="Read options from this file") parser.add_argument('-data_type', default="text", help="Type of the source input. Options are [text|img].") parser.add_argument('-data_img_dir', default=".", help="Location of source images") parser.add_argument('-train_src', required=True, help="Path to the training source data") parser.add_argument('-train_tgt', required=True, help="Path to the training target data") parser.add_argument('-valid_src', required=True, help="Path to the validation source data") parser.add_argument('-valid_tgt', required=True, help="Path to the validation target data") parser.add_argument('-save_data', required=True, help="Output file for the prepared data") parser.add_argument('-src_vocab', help="Path to an existing source vocabulary") parser.add_argument('-tgt_vocab', help="Path to an existing target vocabulary") parser.add_argument('-features_vocabs_prefix', type=str, default='', help="Path prefix to existing features vocabularies") parser.add_argument('-seed', type=int, default=3435, help="Random seed") parser.add_argument('-report_every', type=int, default=100000, help="Report status every this many sentences") opts.preprocess_opts(parser) opt = parser.parse_args() torch.manual_seed(opt.seed) def main(): print('Preparing training ...') with codecs.open(opt.train_src, "r", "utf-8") as src_file: src_line = src_file.readline().strip().split() _, _, n_src_features = onmt.IO.extract_features(src_line) with codecs.open(opt.train_tgt, "r", "utf-8") as tgt_file: tgt_line = tgt_file.readline().strip().split() _, _, n_tgt_features = onmt.IO.extract_features(tgt_line) fields = onmt.IO.get_fields(n_src_features, n_tgt_features) print("Building Training...") train = onmt.IO.ONMTDataset( opt.train_src, opt.train_tgt, fields, opt.src_seq_length, opt.tgt_seq_length, src_seq_length_trunc=opt.src_seq_length_trunc, tgt_seq_length_trunc=opt.tgt_seq_length_trunc, dynamic_dict=opt.dynamic_dict) print("Building Vocab...") onmt.IO.build_vocab(train, opt) print("Building Valid...") valid = onmt.IO.ONMTDataset( opt.valid_src, opt.valid_tgt, fields, opt.src_seq_length, opt.tgt_seq_length, src_seq_length_trunc=opt.src_seq_length_trunc, tgt_seq_length_trunc=opt.tgt_seq_length_trunc, dynamic_dict=opt.dynamic_dict) print("Saving train/valid/fields") # Can't save fields, so remove/reconstruct at training time. torch.save(onmt.IO.save_vocab(fields), open(opt.save_data + '.vocab.pt', 'wb')) train.fields = [] valid.fields = [] torch.save(train, open(opt.save_data + '.train.pt', 'wb')) torch.save(valid, open(opt.save_data + '.valid.pt', 'wb')) if __name__ == "__main__": main()

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/OpenNMT/tools/extract_embeddings.py

from __future__ import division import torch import argparse from onmt.ModelConstructor import make_embeddings, \ make_encoder, make_decoder parser = argparse.ArgumentParser(description='translate.py') parser.add_argument('-model', required=True, help='Path to model .pt file') parser.add_argument('-output_dir', default='.', help="""Path to output the embeddings""") parser.add_argument('-gpu', type=int, default=-1, help="Device to run on") def write_embeddings(filename, dict, embeddings): with open(filename, 'w') as file: for i in range(len(embeddings)): str = dict.idxToLabel[i].encode("utf-8") for j in range(len(embeddings[0])): str = str + " %5f" % (embeddings[i][j]) file.write(str + "\n") def main(): opt = parser.parse_args() checkpoint = torch.load(opt.model) opt.cuda = opt.gpu > -1 if opt.cuda: torch.cuda.set_device(opt.gpu) model_opt = checkpoint['opt'] src_dict = checkpoint['dicts']['src'] tgt_dict = checkpoint['dicts']['tgt'] feature_dicts = [] embeddings = make_embeddings(model_opt, src_dict, feature_dicts) encoder = make_encoder(model_opt, embeddings) embeddings = make_embeddings(model_opt, tgt_dict, feature_dicts, for_encoder=False) decoder = make_decoder(model_opt, embeddings) encoder_embeddings = encoder.word_lut.weight.data.tolist() decoder_embeddings = decoder.word_lut.weight.data.tolist() print("Writing source embeddings") write_embeddings(opt.output_dir + "/src_embeddings.txt", src_dict, encoder_embeddings) print("Writing target embeddings") write_embeddings(opt.output_dir + "/tgt_embeddings.txt", tgt_dict, decoder_embeddings) print('... done.') print('Converting model...') if __name__ == "__main__": main()

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/OpenNMT/test/test_models.py

import argparse import copy import unittest import torch from torch.autograd import Variable import onmt import opts from onmt.ModelConstructor import make_embeddings, \ make_encoder, make_decoder parser = argparse.ArgumentParser(description='train.py') opts.model_opts(parser) opts.train_opts(parser) # -data option is required, but not used in this test, so dummy. opt = parser.parse_known_args(['-data', 'dummy'])[0] print(opt) class TestModel(unittest.TestCase): def __init__(self, *args, **kwargs): super(TestModel, self).__init__(*args, **kwargs) self.opt = opt # Helper to generate a vocabulary def get_vocab(self): src = onmt.IO.get_fields(0, 0)["src"] src.build_vocab([]) return src.vocab def get_batch(self, sourceL=3, bsize=1): # len x batch x nfeat test_src = Variable(torch.ones(sourceL, bsize, 1)).long() test_tgt = Variable(torch.ones(sourceL, bsize, 1)).long() test_length = torch.ones(bsize).fill_(sourceL) return test_src, test_tgt, test_length def embeddings_forward(self, opt, sourceL=3, bsize=1): ''' Tests if the embeddings works as expected args: opt: set of options sourceL: Length of generated input sentence bsize: Batchsize of generated input ''' word_dict = self.get_vocab() feature_dicts = [] emb = make_embeddings(opt, word_dict, feature_dicts) test_src, _, __ = self.get_batch(sourceL=sourceL, bsize=bsize) if opt.decoder_type == 'transformer': input = torch.cat([test_src, test_src], 0) res = emb(input) compare_to = torch.zeros(sourceL * 2, bsize, opt.src_word_vec_size) else: res = emb(test_src) compare_to = torch.zeros(sourceL, bsize, opt.src_word_vec_size) self.assertEqual(res.size(), compare_to.size()) def encoder_forward(self, opt, sourceL=3, bsize=1): ''' Tests if the encoder works as expected args: opt: set of options sourceL: Length of generated input sentence bsize: Batchsize of generated input ''' word_dict = self.get_vocab() feature_dicts = [] embeddings = make_embeddings(opt, word_dict, feature_dicts) enc = make_encoder(opt, embeddings) test_src, test_tgt, test_length = self.get_batch(sourceL=sourceL, bsize=bsize) hidden_t, outputs = enc(test_src, test_length) # Initialize vectors to compare size with test_hid = torch.zeros(self.opt.enc_layers, bsize, opt.rnn_size) test_out = torch.zeros(sourceL, bsize, opt.rnn_size) # Ensure correct sizes and types self.assertEqual(test_hid.size(), hidden_t[0].size(), hidden_t[1].size()) self.assertEqual(test_out.size(), outputs.size()) self.assertEqual(type(outputs), torch.autograd.Variable) self.assertEqual(type(outputs.data), torch.FloatTensor) def ntmmodel_forward(self, opt, sourceL=3, bsize=1): """ Creates a ntmmodel with a custom opt function. Forwards a testbatch anc checks output size. Args: opt: Namespace with options sourceL: length of input sequence bsize: batchsize """ word_dict = self.get_vocab() feature_dicts = [] embeddings = make_embeddings(opt, word_dict, feature_dicts) enc = make_encoder(opt, embeddings) embeddings = make_embeddings(opt, word_dict, feature_dicts, for_encoder=False) dec = make_decoder(opt, embeddings) model = onmt.Models.NMTModel(enc, dec) test_src, test_tgt, test_length = self.get_batch(sourceL=sourceL, bsize=bsize) outputs, attn, _ = model(test_src, test_tgt, test_length) outputsize = torch.zeros(sourceL - 1, bsize, opt.rnn_size) # Make sure that output has the correct size and type self.assertEqual(outputs.size(), outputsize.size()) self.assertEqual(type(outputs), torch.autograd.Variable) self.assertEqual(type(outputs.data), torch.FloatTensor) def _add_test(paramSetting, methodname): """ Adds a Test to TestModel according to settings Args: paramSetting: list of tuples of (param, setting) methodname: name of the method that gets called """ def test_method(self): if paramSetting: opt = copy.deepcopy(self.opt) for param, setting in paramSetting: setattr(opt, param, setting) else: opt = self.opt getattr(self, methodname)(opt) if paramSetting: name = 'test_' + methodname + "_" + "_".join(str(paramSetting).split()) else: name = 'test_' + methodname + '_standard' setattr(TestModel, name, test_method) test_method.__name__ = name ''' TEST PARAMETERS ''' test_embeddings = [[], [('decoder_type', 'transformer')] ] for p in test_embeddings: _add_test(p, 'embeddings_forward') tests_encoder = [[], [('encoder_type', 'mean')], # [('encoder_type', 'transformer'), # ('word_vec_size', 16), ('rnn_size', 16)], [] ] for p in tests_encoder: _add_test(p, 'encoder_forward') tests_ntmodel = [[('rnn_type', 'GRU')], [('layers', 10)], [('input_feed', 0)], [('decoder_type', 'transformer'), ('encoder_type', 'transformer'), ('src_word_vec_size', 16), ('tgt_word_vec_size', 16), ('rnn_size', 16)], # [('encoder_type', 'transformer'), # ('word_vec_size', 16), # ('rnn_size', 16)], [('decoder_type', 'transformer'), ('encoder_type', 'transformer'), ('src_word_vec_size', 16), ('tgt_word_vec_size', 16), ('rnn_size', 16), ('position_encoding', True)], [('coverage_attn', True)], [('copy_attn', True)], [('global_attention', 'mlp')], [('context_gate', 'both')], [('context_gate', 'target')], [('context_gate', 'source')], [('encoder_type', "brnn"), ('brnn_merge', 'sum')], [('encoder_type', "brnn")], [('decoder_type', 'cnn'), ('encoder_type', 'cnn')], [] ] if onmt.modules.check_sru_requirement(): """ Only do SRU test if requirment is safisfied. """ # SRU doesn't support input_feed. tests_ntmodel.append([('rnn_type', 'SRU'), ('input_feed', 0)]) for p in tests_ntmodel: _add_test(p, 'ntmmodel_forward')

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/OpenNMT/test/test_preprocess.py

import argparse import copy import unittest import onmt import opts import torchtext from collections import Counter parser = argparse.ArgumentParser(description='preprocess.py') opts.preprocess_opts(parser) opt = parser.parse_known_args()[0] opt.train_src = 'data/src-train.txt' opt.train_tgt = 'data/tgt-train.txt' opt.valid_src = 'data/src-val.txt' opt.valid_tgt = 'data/tgt-val.txt' print(opt) class TestData(unittest.TestCase): def __init__(self, *args, **kwargs): super(TestData, self).__init__(*args, **kwargs) self.opt = opt def dataset_build(self, opt): fields = onmt.IO.get_fields(0, 0) train = onmt.IO.ONMTDataset( opt.train_src, opt.train_tgt, fields, opt.src_seq_length, opt.tgt_seq_length, src_seq_length_trunc=opt.src_seq_length_trunc, tgt_seq_length_trunc=opt.tgt_seq_length_trunc, dynamic_dict=opt.dynamic_dict) onmt.IO.build_vocab(train, opt) onmt.IO.ONMTDataset( opt.valid_src, opt.valid_tgt, fields, opt.src_seq_length, opt.tgt_seq_length, src_seq_length_trunc=opt.src_seq_length_trunc, tgt_seq_length_trunc=opt.tgt_seq_length_trunc, dynamic_dict=opt.dynamic_dict) def test_merge_vocab(self): va = torchtext.vocab.Vocab(Counter('abbccc')) vb = torchtext.vocab.Vocab(Counter('eeabbcccf')) merged = onmt.IO.merge_vocabs([va, vb], 2) self.assertEqual(Counter({'c': 6, 'b': 4, 'a': 2, 'e': 2, 'f': 1}), merged.freqs) self.assertEqual(6, len(merged.itos)) self.assertTrue('b' in merged.itos) def _add_test(paramSetting, methodname): """ Adds a Test to TestData according to settings Args: paramSetting: list of tuples of (param, setting) methodname: name of the method that gets called """ def test_method(self): if paramSetting: opt = copy.deepcopy(self.opt) for param, setting in paramSetting: setattr(opt, param, setting) else: opt = self.opt getattr(self, methodname)(opt) if paramSetting: name = 'test_' + methodname + "_" + "_".join(str(paramSetting).split()) else: name = 'test_' + methodname + '_standard' setattr(TestData, name, test_method) test_method.__name__ = name test_databuild = [[], [('src_vocab_size', 1), ('tgt_vocab_size', 1)], [('src_vocab_size', 10000), ('tgt_vocab_size', 10000)], [('src_seq_length', 1)], [('src_seq_length', 5000)], [('src_seq_length_trunc', 1)], [('src_seq_length_trunc', 5000)], [('tgt_seq_length', 1)], [('tgt_seq_length', 5000)], [('tgt_seq_length_trunc', 1)], [('tgt_seq_length_trunc', 5000)], [('shuffle', 0)], [('lower', True)], [('dynamic_dict', True)], [('share_vocab', True)], [('dynamic_dict', True), ('share_vocab', True)], ] for p in test_databuild: _add_test(p, 'dataset_build')

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/OpenNMT/onmt/Loss.py

""" This file handles the details of the loss function during training. This includes: LossComputeBase and the standard NMTLossCompute, and sharded loss compute stuff. """ from __future__ import division import torch import torch.nn as nn from torch.autograd import Variable import onmt class LossComputeBase(nn.Module): """ This is the loss criterion base class. Users can implement their own loss computation strategy by making subclass of this one. Users need to implement the compute_loss() and make_shard_state() methods. We inherits from nn.Module to leverage the cuda behavior. """ def __init__(self, generator, tgt_vocab): super(LossComputeBase, self).__init__() self.generator = generator self.tgt_vocab = tgt_vocab self.padding_idx = tgt_vocab.stoi[onmt.IO.PAD_WORD] def make_shard_state(self, batch, output, range_, attns=None): """ Make shard state dictionary for shards() to return iterable shards for efficient loss computation. Subclass must define this method to match its own compute_loss() interface. Args: batch: the current batch. output: the predict output from the model. range_: the range of examples for computing, the whole batch or a trunc of it? attns: the attns dictionary returned from the model. """ return NotImplementedError def compute_loss(self, batch, output, target, **kwargs): """ Compute the loss. Subclass must define this method. Args: batch: the current batch. output: the predict output from the model. target: the validate target to compare output with. **kwargs(optional): additional info for computing loss. """ return NotImplementedError def monolithic_compute_loss(self, batch, output, attns): """ Compute the loss monolithically, not dividing into shards. """ range_ = (0, batch.tgt.size(0)) shard_state = self.make_shard_state(batch, output, range_, attns) _, batch_stats = self.compute_loss(batch, **shard_state) return batch_stats def sharded_compute_loss(self, batch, output, attns, cur_trunc, trunc_size, shard_size): """ Compute the loss in shards for efficiency. """ batch_stats = onmt.Statistics() range_ = (cur_trunc, cur_trunc + trunc_size) shard_state = self.make_shard_state(batch, output, range_, attns) for shard in shards(shard_state, shard_size): loss, stats = self.compute_loss(batch, **shard) loss.div(batch.batch_size).backward() batch_stats.update(stats) return batch_stats def stats(self, loss, scores, target): """ Compute and return a Statistics object. Args: loss(Tensor): the loss computed by the loss criterion. scores(Tensor): a sequence of predict output with scores. """ pred = scores.max(1)[1] non_padding = target.ne(self.padding_idx) num_correct = pred.eq(target) \ .masked_select(non_padding) \ .sum() return onmt.Statistics(loss[0], non_padding.sum(), num_correct) def bottle(self, v): return v.view(-1, v.size(2)) def unbottle(self, v, batch_size): return v.view(-1, batch_size, v.size(1)) class NMTLossCompute(LossComputeBase): """ Standard NMT Loss Computation. """ def __init__(self, generator, tgt_vocab): super(NMTLossCompute, self).__init__(generator, tgt_vocab) weight = torch.ones(len(tgt_vocab)) weight[self.padding_idx] = 0 self.criterion = nn.NLLLoss(weight, size_average=False) def make_shard_state(self, batch, output, range_, attns=None): """ See base class for args description. """ return { "output": output, "target": batch.tgt[range_[0] + 1: range_[1]], } def compute_loss(self, batch, output, target): """ See base class for args description. """ scores = self.generator(self.bottle(output)) target = target.view(-1) loss = self.criterion(scores, target) loss_data = loss.data.clone() stats = self.stats(loss_data, scores.data, target.data) return loss, stats def filter_shard_state(state): for k, v in state.items(): if v is not None: if isinstance(v, Variable) and v.requires_grad: v = Variable(v.data, requires_grad=True, volatile=False) yield k, v def shards(state, shard_size, eval=False): """ Args: state: A dictionary which corresponds to the output of *LossCompute.make_shard_state(). The values for those keys are Tensor-like or None. shard_size: The maximum size of the shards yielded by the model. eval: If True, only yield the state, nothing else. Otherwise, yield shards. Yields: Each yielded shard is a dict. Side effect: After the last shard, this function does back-propagation. """ if eval: yield state else: # non_none: the subdict of the state dictionary where the values # are not None. non_none = dict(filter_shard_state(state)) # Now, the iteration: # state is a dictionary of sequences of tensor-like but we # want a sequence of dictionaries of tensors. # First, unzip the dictionary into a sequence of keys and a # sequence of tensor-like sequences. keys, values = zip(*((k, torch.split(v, shard_size)) for k, v in non_none.items())) # Now, yield a dictionary for each shard. The keys are always # the same. values is a sequence of length #keys where each # element is a sequence of length #shards. We want to iterate # over the shards, not over the keys: therefore, the values need # to be re-zipped by shard and then each shard can be paired # with the keys. for shard_tensors in zip(*values): yield dict(zip(keys, shard_tensors)) # Assumed backprop'd variables = ((state[k], v.grad.data) for k, v in non_none.items() if isinstance(v, Variable) and v.grad is not None) inputs, grads = zip(*variables) torch.autograd.backward(inputs, grads)

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/OpenNMT/onmt/Beam.py

from __future__ import division import torch import onmt """ Class for managing the internals of the beam search process. Takes care of beams, back pointers, and scores. """ class Beam(object): def __init__(self, size, n_best=1, cuda=False, vocab=None, global_scorer=None): self.size = size self.tt = torch.cuda if cuda else torch # The score for each translation on the beam. self.scores = self.tt.FloatTensor(size).zero_() self.allScores = [] # The backpointers at each time-step. self.prevKs = [] # The outputs at each time-step. self.nextYs = [self.tt.LongTensor(size) .fill_(vocab.stoi[onmt.IO.PAD_WORD])] self.nextYs[0][0] = vocab.stoi[onmt.IO.BOS_WORD] self.vocab = vocab # Has EOS topped the beam yet. self._eos = self.vocab.stoi[onmt.IO.EOS_WORD] self.eosTop = False # The attentions (matrix) for each time. self.attn = [] # Time and k pair for finished. self.finished = [] self.n_best = n_best # Information for global scoring. self.globalScorer = global_scorer self.globalState = {} def getCurrentState(self): "Get the outputs for the current timestep." return self.nextYs[-1] def getCurrentOrigin(self): "Get the backpointers for the current timestep." return self.prevKs[-1] def advance(self, wordLk, attnOut): """ Given prob over words for every last beam `wordLk` and attention `attnOut`: Compute and update the beam search. Parameters: * `wordLk`- probs of advancing from the last step (K x words) * `attnOut`- attention at the last step Returns: True if beam search is complete. """ numWords = wordLk.size(1) # Sum the previous scores. if len(self.prevKs) > 0: beamLk = wordLk + self.scores.unsqueeze(1).expand_as(wordLk) # Don't let EOS have children. for i in range(self.nextYs[-1].size(0)): if self.nextYs[-1][i] == self._eos: beamLk[i] = -1e20 else: beamLk = wordLk[0] flatBeamLk = beamLk.view(-1) bestScores, bestScoresId = flatBeamLk.topk(self.size, 0, True, True) self.allScores.append(self.scores) self.scores = bestScores # bestScoresId is flattened beam x word array, so calculate which # word and beam each score came from prevK = bestScoresId / numWords self.prevKs.append(prevK) self.nextYs.append((bestScoresId - prevK * numWords)) self.attn.append(attnOut.index_select(0, prevK)) if self.globalScorer is not None: self.globalScorer.updateGlobalState(self) for i in range(self.nextYs[-1].size(0)): if self.nextYs[-1][i] == self._eos: s = self.scores[i] if self.globalScorer is not None: globalScores = self.globalScorer.score(self, self.scores) s = globalScores[i] self.finished.append((s, len(self.nextYs) - 1, i)) # End condition is when top-of-beam is EOS and no global score. if self.nextYs[-1][0] == self.vocab.stoi[onmt.IO.EOS_WORD]: # self.allScores.append(self.scores) self.eosTop = True def done(self): return self.eosTop and len(self.finished) >= self.n_best def sortFinished(self, minimum=None): if minimum is not None: i = 0 # Add from beam until we have minimum outputs. while len(self.finished) < minimum: s = self.scores[i] if self.globalScorer is not None: globalScores = self.globalScorer.score(self, self.scores) s = globalScores[i] self.finished.append((s, len(self.nextYs) - 1, i)) self.finished.sort(key=lambda a: -a[0]) scores = [sc for sc, _, _ in self.finished] ks = [(t, k) for _, t, k in self.finished] return scores, ks def getHyp(self, timestep, k): """ Walk back to construct the full hypothesis. """ hyp, attn = [], [] for j in range(len(self.prevKs[:timestep]) - 1, -1, -1): hyp.append(self.nextYs[j+1][k]) attn.append(self.attn[j][k]) k = self.prevKs[j][k] return hyp[::-1], torch.stack(attn[::-1]) class GNMTGlobalScorer(object): """ Google NMT ranking score from Wu et al. """ def __init__(self, alpha, beta): self.alpha = alpha self.beta = beta def score(self, beam, logprobs): "Additional term add to log probability" cov = beam.globalState["coverage"] pen = self.beta * torch.min(cov, cov.clone().fill_(1.0)).log().sum(1) l_term = (((5 + len(beam.nextYs)) ** self.alpha) / ((5 + 1) ** self.alpha)) return (logprobs / l_term) + pen def updateGlobalState(self, beam): "Keeps the coverage vector as sum of attens" if len(beam.prevKs) == 1: beam.globalState["coverage"] = beam.attn[-1] else: beam.globalState["coverage"] = beam.globalState["coverage"] \ .index_select(0, beam.prevKs[-1]).add(beam.attn[-1])

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/OpenNMT/onmt/Translator.py

import torch from torch.autograd import Variable import onmt import onmt.Models import onmt.ModelConstructor import onmt.modules import onmt.IO from onmt.Utils import use_gpu NOISE_TRANSELATE = False class Translator(object): def __init__(self, opt, dummy_opt={}): # Add in default model arguments, possibly added since training. self.opt = opt checkpoint = torch.load(opt.model, map_location=lambda storage, loc: storage) self.fields = onmt.IO.load_fields(checkpoint['vocab']) model_opt = checkpoint['opt'] for arg in dummy_opt: if arg not in model_opt: model_opt.__dict__[arg] = dummy_opt[arg] self._type = model_opt.encoder_type self.copy_attn = model_opt.copy_attn self.model = onmt.ModelConstructor.make_base_model( model_opt, self.fields, use_gpu(opt), checkpoint) self.model.eval() self.model.generator.eval() # for debugging self.beam_accum = None def initBeamAccum(self): self.beam_accum = { "predicted_ids": [], "beam_parent_ids": [], "scores": [], "log_probs": []} def buildTargetTokens(self, pred, src, attn, copy_vocab): vocab = self.fields["tgt"].vocab tokens = [] for tok in pred: if tok < len(vocab): tokens.append(vocab.itos[tok]) else: tokens.append(copy_vocab.itos[tok - len(vocab)]) if tokens[-1] == onmt.IO.EOS_WORD: tokens = tokens[:-1] break if self.opt.replace_unk and attn is not None: for i in range(len(tokens)): if tokens[i] == vocab.itos[onmt.IO.UNK]: _, maxIndex = attn[i].max(0) tokens[i] = self.fields["src"].vocab.itos[src[maxIndex[0]]] return tokens def _runTarget(self, batch, data): _, src_lengths = batch.src src = onmt.IO.make_features(batch, 'src') tgt_in = onmt.IO.make_features(batch, 'tgt')[:-1] # (1) run the encoder on the src encStates, context = self.model.encoder(src, src_lengths) decStates = self.model.decoder.init_decoder_state( src, context, encStates) # (2) if a target is specified, compute the 'goldScore' # (i.e. log likelihood) of the target under the model tt = torch.cuda if self.opt.cuda else torch goldScores = tt.FloatTensor(batch.batch_size).fill_(0) decOut, decStates, attn = self.model.decoder( tgt_in, context, decStates) tgt_pad = self.fields["tgt"].vocab.stoi[onmt.IO.PAD_WORD] for dec, tgt in zip(decOut, batch.tgt[1:].data): # Log prob of each word. out = self.model.generator.forward(dec) tgt = tgt.unsqueeze(1) scores = out.data.gather(1, tgt) scores.masked_fill_(tgt.eq(tgt_pad), 0) goldScores += scores return goldScores def translateBatch(self, batch, dataset): beam_size = self.opt.beam_size batch_size = batch.batch_size # (1) Run the encoder on the src. _, src_lengths = batch.src src = onmt.IO.make_features(batch, 'src') encStates, context = self.model.encoder(src, src_lengths) # return hidden_t, outputs print(type(torch.autograd.Variable(torch.FloatTensor(encStates[0].data.shape).uniform_(-0.2, 0.2)))) print(type(encStates[0])) newEncStates = ( encStates[0] + torch.autograd.Variable(torch.FloatTensor(encStates[0].data.shape).uniform_(-0.2, 0.2)).cuda(), encStates[1] + torch.autograd.Variable(torch.FloatTensor(encStates[1].data.shape).uniform_(-0.2, 0.2).cuda()) ) if NOISE_TRANSELATE: decStates = self.model.decoder.init_decoder_state(src, context, newEncStates) else: decStates = self.model.decoder.init_decoder_state(src, context, encStates) # (1b) Initialize for the decoder. def var(a): return Variable(a, volatile=True) def rvar(a): return var(a.repeat(1, beam_size, 1)) # Repeat everything beam_size times. context = rvar(context.data) src = rvar(src.data) srcMap = rvar(batch.src_map.data) decStates.repeat_beam_size_times(beam_size) scorer = None # scorer=onmt.GNMTGlobalScorer(0.3, 0.4) beam = [onmt.Beam(beam_size, n_best=self.opt.n_best, cuda=self.opt.cuda, vocab=self.fields["tgt"].vocab, global_scorer=scorer) for __ in range(batch_size)] # (2) run the decoder to generate sentences, using beam search. def bottle(m): return m.view(batch_size * beam_size, -1) def unbottle(m): return m.view(beam_size, batch_size, -1) for i in range(self.opt.max_sent_length): if all((b.done() for b in beam)): break # Construct batch x beam_size nxt words. # Get all the pending current beam words and arrange for forward. inp = var(torch.stack([b.getCurrentState() for b in beam]) .t().contiguous().view(1, -1)) # Turn any copied words to UNKs # 0 is unk if self.copy_attn: inp = inp.masked_fill( inp.gt(len(self.fields["tgt"].vocab) - 1), 0) # Temporary kludge solution to handle changed dim expectation # in the decoder inp = inp.unsqueeze(2) # Run one step. decOut, decStates, attn = \ self.model.decoder(inp, context, decStates) decOut = decOut.squeeze(0) # decOut: beam x rnn_size # (b) Compute a vector of batch*beam word scores. if not self.copy_attn: out = self.model.generator.forward(decOut).data out = unbottle(out) # beam x tgt_vocab else: out = self.model.generator.forward(decOut, attn["copy"].squeeze(0), srcMap) # beam x (tgt_vocab + extra_vocab) out = dataset.collapse_copy_scores( unbottle(out.data), batch, self.fields["tgt"].vocab) # beam x tgt_vocab out = out.log() # (c) Advance each beam. for j, b in enumerate(beam): b.advance(out[:, j], unbottle(attn["std"]).data[:, j]) decStates.beam_update(j, b.getCurrentOrigin(), beam_size) if "tgt" in batch.__dict__: allGold = self._runTarget(batch, dataset) else: allGold = [0] * batch_size # (3) Package everything up. allHyps, allScores, allAttn = [], [], [] for b in beam: n_best = self.opt.n_best scores, ks = b.sortFinished(minimum=n_best) hyps, attn = [], [] for i, (times, k) in enumerate(ks[:n_best]): hyp, att = b.getHyp(times, k) hyps.append(hyp) attn.append(att) allHyps.append(hyps) allScores.append(scores) allAttn.append(attn) return allHyps, allScores, allAttn, allGold def translate(self, batch, data): # (1) convert words to indexes batch_size = batch.batch_size # (2) translate pred, predScore, attn, goldScore = self.translateBatch(batch, data) assert(len(goldScore) == len(pred)) pred, predScore, attn, goldScore, i = list(zip( *sorted(zip(pred, predScore, attn, goldScore, batch.indices.data), key=lambda x: x[-1]))) inds, perm = torch.sort(batch.indices.data) # (3) convert indexes to words predBatch, goldBatch = [], [] src = batch.src[0].data.index_select(1, perm) if self.opt.tgt: tgt = batch.tgt.data.index_select(1, perm) for b in range(batch_size): src_vocab = data.src_vocabs[inds[b]] predBatch.append( [self.buildTargetTokens(pred[b][n], src[:, b], attn[b][n], src_vocab) for n in range(self.opt.n_best)]) if self.opt.tgt: goldBatch.append( self.buildTargetTokens(tgt[1:, b], src[:, b], None, None)) return predBatch, goldBatch, predScore, goldScore, attn, src

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/OpenNMT/onmt/IO.py

# -*- coding: utf-8 -*- import codecs from collections import Counter, defaultdict from itertools import chain, count import torch import torchtext.data import torchtext.vocab PAD_WORD = '<blank>' UNK = 0 BOS_WORD = '<s>' EOS_WORD = '</s>' def __getstate__(self): return dict(self.__dict__, stoi=dict(self.stoi)) def __setstate__(self, state): self.__dict__.update(state) self.stoi = defaultdict(lambda: 0, self.stoi) torchtext.vocab.Vocab.__getstate__ = __getstate__ torchtext.vocab.Vocab.__setstate__ = __setstate__ def load_fields(vocab): vocab = dict(vocab) n_src_features = len(collect_features(vocab, 'src')) n_tgt_features = len(collect_features(vocab, 'tgt')) fields = get_fields(n_src_features, n_tgt_features) for k, v in vocab.items(): # Hack. Can't pickle defaultdict :( v.stoi = defaultdict(lambda: 0, v.stoi) fields[k].vocab = v return fields def collect_features(fields, side="src"): assert side in ["src", "tgt"] feats = [] for j in count(): key = side + "_feat_" + str(j) if key not in fields: break feats.append(key) return feats def extract_features(tokens): """ tokens: A list of tokens, where each token consists of a word, optionally followed by u"￨"-delimited features. returns: A sequence of words, A sequence of features, and """ if not tokens: return [], [], -1 split_tokens = [token.split(u"￨") for token in tokens] split_tokens = [token for token in split_tokens if token[0]] token_size = len(split_tokens[0]) assert all(len(token) == token_size for token in split_tokens), \ "all words must have the same number of features" words_and_features = list(zip(*split_tokens)) words = words_and_features[0] features = words_and_features[1:] return words, features, token_size - 1 def read_corpus_file(path, truncate, side): """ path: location of a src or tgt file truncate: maximum sequence length (0 for unlimited) yields: (word, features, nfeat) triples for each line """ with codecs.open(path, "r", "utf-8") as corpus_file: for i, line in enumerate(corpus_file): line = line.split() if truncate: line = line[:truncate] words, feats, n_feats = extract_features(line) example_dict = {side: words, "indices": i} if feats: prefix = side + "_feat_" example_dict.update((prefix + str(j), f) for j, f in enumerate(feats)) yield example_dict, n_feats def merge_vocabs(vocabs, vocab_size=None): """ Merge individual vocabularies (assumed to be generated from disjoint documents) into a larger vocabulary. Args: vocabs: `torchtext.vocab.Vocab` vocabularies to be merged vocab_size: `int` the final vocabulary size. `None` for no limit. Return: `torchtext.vocab.Vocab` """ merged = sum([vocab.freqs for vocab in vocabs], Counter()) return torchtext.vocab.Vocab(merged, specials=[PAD_WORD, BOS_WORD, EOS_WORD], max_size=vocab_size) def make_features(batch, side): """ Args: batch (Variable): a batch of source or target data. side (str): for source or for target. Returns: A sequence of src/tgt tensors with optional feature tensors of size (len x batch). """ assert side in ['src', 'tgt'] if isinstance(batch.__dict__[side], tuple): data = batch.__dict__[side][0] else: data = batch.__dict__[side] feat_start = side + "_feat_" features = sorted(batch.__dict__[k] for k in batch.__dict__ if feat_start in k) levels = [data] + features return torch.cat([level.unsqueeze(2) for level in levels], 2) def save_vocab(fields): vocab = [] for k, f in fields.items(): if 'vocab' in f.__dict__: f.vocab.stoi = dict(f.vocab.stoi) vocab.append((k, f.vocab)) return vocab def collect_feature_dicts(fields, side): assert side in ['src', 'tgt'] feature_dicts = [] for j in count(): key = side + "_feat_" + str(j) if key not in fields: break feature_dicts.append(fields[key].vocab) return feature_dicts def get_fields(n_src_features, n_tgt_features): """ n_src_features: the number of source features to create Field objects for. n_tgt_features: the number of target features to create Field objects for. returns: A dictionary whose keys are strings and whose values are the corresponding Field objects. """ fields = {} fields["src"] = torchtext.data.Field( pad_token=PAD_WORD, include_lengths=True) # fields = [("src_img", torchtext.data.Field( # include_lengths=True))] for j in range(n_src_features): fields["src_feat_"+str(j)] = \ torchtext.data.Field(pad_token=PAD_WORD) fields["tgt"] = torchtext.data.Field( init_token=BOS_WORD, eos_token=EOS_WORD, pad_token=PAD_WORD) for j in range(n_tgt_features): fields["tgt_feat_"+str(j)] = \ torchtext.data.Field(init_token=BOS_WORD, eos_token=EOS_WORD, pad_token=PAD_WORD) def make_src(data, _): src_size = max([t.size(0) for t in data]) src_vocab_size = max([t.max() for t in data]) + 1 alignment = torch.zeros(src_size, len(data), src_vocab_size) for i, sent in enumerate(data): for j, t in enumerate(sent): alignment[j, i, t] = 1 return alignment fields["src_map"] = torchtext.data.Field( use_vocab=False, tensor_type=torch.FloatTensor, postprocessing=make_src, sequential=False) def make_tgt(data, _): tgt_size = max([t.size(0) for t in data]) alignment = torch.zeros(tgt_size, len(data)).long() for i, sent in enumerate(data): alignment[:sent.size(0), i] = sent return alignment fields["alignment"] = torchtext.data.Field( use_vocab=False, tensor_type=torch.LongTensor, postprocessing=make_tgt, sequential=False) fields["indices"] = torchtext.data.Field( use_vocab=False, tensor_type=torch.LongTensor, sequential=False) return fields def build_vocab(train, opt): """ train: an ONMTDataset """ fields = train.fields fields["src"].build_vocab(train, max_size=opt.src_vocab_size, min_freq=opt.src_words_min_frequency) for j in range(train.n_src_feats): fields["src_feat_" + str(j)].build_vocab(train) fields["tgt"].build_vocab(train, max_size=opt.tgt_vocab_size, min_freq=opt.tgt_words_min_frequency) for j in range(train.n_tgt_feats): fields["tgt_feat_" + str(j)].build_vocab(train) # Merge the input and output vocabularies. if opt.share_vocab: # `tgt_vocab_size` is ignored when sharing vocabularies merged_vocab = merge_vocabs( [fields["src"].vocab, fields["tgt"].vocab], vocab_size=opt.src_vocab_size) fields["src"].vocab = merged_vocab fields["tgt"].vocab = merged_vocab def join_dicts(*args): """ args: dictionaries with disjoint keys returns: a single dictionary that has the union of these keys """ return dict(chain(*[d.items() for d in args])) def peek(seq): """ sequence: an iterator returns: the first thing returned by calling next() on the iterator and an iterator created by re-chaining that value to the beginning of the iterator. """ first = next(seq) return first, chain([first], seq) class OrderedIterator(torchtext.data.Iterator): def create_batches(self): if self.train: self.batches = torchtext.data.pool( self.data(), self.batch_size, self.sort_key, self.batch_size_fn, random_shuffler=self.random_shuffler) else: self.batches = [] for b in torchtext.data.batch(self.data(), self.batch_size, self.batch_size_fn): self.batches.append(sorted(b, key=self.sort_key)) class ONMTDataset(torchtext.data.Dataset): """ Defines a dataset for machine translation. An ONMTDataset is a collection that supports iteration over its examples. The parent class supports indexing as well, but future developments here may make that difficult (lazy iteration over examples because of large datasets, for example). """ @staticmethod def sort_key(ex): "Sort in reverse size order" return -len(ex.src) def __init__(self, src_path, tgt_path, fields, src_seq_length=0, tgt_seq_length=0, src_seq_length_trunc=0, tgt_seq_length_trunc=0, use_filter_pred=True, dynamic_dict=True, src_img_dir=None, **kwargs): """ Create a translation dataset given paths and fields. src_path: location of source-side data tgt_path: location of target-side data or None. If should be the same length as the source-side data if it exists, but at present this is not checked. fields: a dictionary. keys are things like 'src', 'tgt', 'src_map', and 'alignment' src_img_dir: raises an error if not None because images are not supported yet. Initializes an ONMTDataset object with the following attributes: self.examples (might be a generator, might be a list, hard to say): A sequence of torchtext Example objects. self.fields (dict): A dictionary associating str keys with Field objects. Does not necessarily have the same keys as the input fields. A dataset basically supports iteration over all the examples it contains. """ assert src_img_dir is None, "img data is not finished" # self.src_vocabs: mutated in dynamic_dict, used in # collapse_copy_scores and in Translator.py self.src_vocabs = [] src_truncate = src_seq_length_trunc src_examples = read_corpus_file(src_path, src_truncate, "src") (_, src_feats), src_examples = peek(src_examples) src_examples = (ex for ex, nfeats in src_examples) self.n_src_feats = src_feats # if tgt_path exists, then we need to do the same thing as we did # for the source data if tgt_path is not None: tgt_truncate = tgt_seq_length_trunc tgt_examples = read_corpus_file(tgt_path, tgt_truncate, "tgt") (_, tgt_feats), tgt_examples = peek(tgt_examples) tgt_examples = (ex for ex, nfeats in tgt_examples) self.n_tgt_feats = tgt_feats else: self.n_tgt_feats = 0 tgt_examples = None # examples: one for each src line or (src, tgt) line pair. # Each element is a dictionary whose keys represent at minimum # the src tokens and their indices and potentially also the # src and tgt features and alignment information. if tgt_examples is not None: examples = (join_dicts(src, tgt) for src, tgt in zip(src_examples, tgt_examples)) else: examples = src_examples if dynamic_dict: examples = self.dynamic_dict(examples) # Peek at the first to see which fields are used. ex, examples = peek(examples) keys = ex.keys() fields = [(k, fields[k]) for k in keys] example_values = ([ex[k] for k in keys] for ex in examples) out_examples = (torchtext.data.Example.fromlist(ex_values, fields) for ex_values in example_values) def filter_pred(example): return 0 < len(example.src) <= src_seq_length \ and 0 < len(example.tgt) <= tgt_seq_length super(ONMTDataset, self).__init__( out_examples, fields, filter_pred if use_filter_pred else None ) def dynamic_dict(self, examples): for example in examples: src = example["src"] src_vocab = torchtext.vocab.Vocab(Counter(src)) self.src_vocabs.append(src_vocab) # mapping source tokens to indices in the dynamic dict src_map = torch.LongTensor([src_vocab.stoi[w] for w in src]) example["src_map"] = src_map if "tgt" in example: tgt = example["tgt"] mask = torch.LongTensor( [0] + [src_vocab.stoi[w] for w in tgt] + [0]) example["alignment"] = mask yield example def __getstate__(self): return self.__dict__ def __setstate__(self, d): self.__dict__.update(d) def __reduce_ex__(self, proto): "This is a hack. Something is broken with torch pickle." return super(ONMTDataset, self).__reduce_ex__() def collapse_copy_scores(self, scores, batch, tgt_vocab): """ Given scores from an expanded dictionary corresponeding to a batch, sums together copies, with a dictionary word when it is ambigious. """ offset = len(tgt_vocab) for b in range(batch.batch_size): index = batch.indices.data[b] src_vocab = self.src_vocabs[index] for i in range(1, len(src_vocab)): sw = src_vocab.itos[i] ti = tgt_vocab.stoi[sw] if ti != 0: scores[:, b, ti] += scores[:, b, offset + i] scores[:, b, offset + i].fill_(1e-20) return scores def load_image_libs(): "Conditional import of torch image libs." global Image, transforms from PIL import Image from torchvision import transforms

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/OpenNMT/onmt/ModelConstructor.py

""" This file is for models creation, which consults options and creates each encoder and decoder accordingly. """ import torch.nn as nn import onmt import onmt.Models import onmt.modules from onmt.Models import NMTModel, MeanEncoder, RNNEncoder, \ StdRNNDecoder, InputFeedRNNDecoder from onmt.modules import Embeddings, ImageEncoder, CopyGenerator, \ TransformerEncoder, TransformerDecoder, \ CNNEncoder, CNNDecoder def make_embeddings(opt, word_dict, feature_dicts, for_encoder=True): """ Make an Embeddings instance. Args: opt: the option in current environment. word_dict(Vocab): words dictionary. feature_dicts([Vocab], optional): a list of feature dictionary. for_encoder(bool): make Embeddings for encoder or decoder? """ if for_encoder: embedding_dim = opt.src_word_vec_size else: embedding_dim = opt.tgt_word_vec_size word_padding_idx = word_dict.stoi[onmt.IO.PAD_WORD] num_word_embeddings = len(word_dict) feats_padding_idx = [feat_dict.stoi[onmt.IO.PAD_WORD] for feat_dict in feature_dicts] num_feat_embeddings = [len(feat_dict) for feat_dict in feature_dicts] return Embeddings(embedding_dim, opt.position_encoding, opt.feat_merge, opt.feat_vec_exponent, opt.feat_vec_size, opt.dropout, word_padding_idx, feats_padding_idx, num_word_embeddings, num_feat_embeddings) def make_encoder(opt, embeddings): """ Various encoder dispatcher function. Args: opt: the option in current environment. embeddings (Embeddings): vocab embeddings for this encoder. """ if opt.encoder_type == "transformer": return TransformerEncoder(opt.enc_layers, opt.rnn_size, opt.dropout, embeddings) elif opt.encoder_type == "cnn": return CNNEncoder(opt.enc_layers, opt.rnn_size, opt.cnn_kernel_width, opt.dropout, embeddings) elif opt.encoder_type == "mean": return MeanEncoder(opt.enc_layers, embeddings) else: # "rnn" or "brnn" return RNNEncoder(opt.rnn_type, opt.brnn, opt.enc_layers, opt.rnn_size, opt.dropout, embeddings) def make_decoder(opt, embeddings): """ Various decoder dispatcher function. Args: opt: the option in current environment. embeddings (Embeddings): vocab embeddings for this decoder. """ if opt.decoder_type == "transformer": return TransformerDecoder(opt.dec_layers, opt.rnn_size, opt.global_attention, opt.copy_attn, opt.dropout, embeddings) elif opt.decoder_type == "cnn": return CNNDecoder(opt.dec_layers, opt.rnn_size, opt.global_attention, opt.copy_attn, opt.cnn_kernel_width, opt.dropout, embeddings) elif opt.input_feed: return InputFeedRNNDecoder(opt.rnn_type, opt.brnn, opt.dec_layers, opt.rnn_size, opt.global_attention, opt.coverage_attn, opt.context_gate, opt.copy_attn, opt.dropout, embeddings) else: return StdRNNDecoder(opt.rnn_type, opt.brnn, opt.dec_layers, opt.rnn_size, opt.global_attention, opt.coverage_attn, opt.context_gate, opt.copy_attn, opt.dropout, embeddings) def make_base_model(model_opt, fields, gpu, checkpoint=None): """ Args: model_opt: the option loaded from checkpoint. fields: `Field` objects for the model. gpu(bool): whether to use gpu. checkpoint: the model gnerated by train phase, or a resumed snapshot model from a stopped training. Returns: the NMTModel. """ assert model_opt.model_type in ["text", "img"], \ ("Unsupported model type %s" % (model_opt.model_type)) # Make encoder. if model_opt.model_type == "text": src_dict = fields["src"].vocab feature_dicts = onmt.IO.collect_feature_dicts(fields, 'src') src_embeddings = make_embeddings(model_opt, src_dict, feature_dicts) encoder = make_encoder(model_opt, src_embeddings) else: encoder = ImageEncoder(model_opt.layers, model_opt.brnn, model_opt.rnn_size, model_opt.dropout) # Make decoder. tgt_dict = fields["tgt"].vocab # TODO: prepare for a future where tgt features are possible. feature_dicts = onmt.IO.collect_feature_dicts(fields, 'tgt') tgt_embeddings = make_embeddings(model_opt, tgt_dict, feature_dicts, for_encoder=False) # Share the embedding matrix - preprocess with share_vocab required if model_opt.share_embeddings: tgt_embeddings.word_lut.weight = src_embeddings.word_lut.weight decoder = make_decoder(model_opt, tgt_embeddings) # Make NMTModel(= encoder + decoder). model = NMTModel(encoder, decoder) # Make Generator. if not model_opt.copy_attn: generator = nn.Sequential( nn.Linear(model_opt.rnn_size, len(fields["tgt"].vocab)), nn.LogSoftmax()) if model_opt.share_decoder_embeddings: generator[0].weight = decoder.embeddings.word_lut.weight else: generator = CopyGenerator(model_opt, fields["src"].vocab, fields["tgt"].vocab) # Load the model states from checkpoint or initialize them. if checkpoint is not None: print('Loading model parameters.') model.load_state_dict(checkpoint['model']) generator.load_state_dict(checkpoint['generator']) else: if model_opt.param_init != 0.0: print('Intializing model parameters.') for p in model.parameters(): p.data.uniform_(-model_opt.param_init, model_opt.param_init) for p in generator.parameters(): p.data.uniform_(-model_opt.param_init, model_opt.param_init) model.encoder.embeddings.load_pretrained_vectors( model_opt.pre_word_vecs_enc, model_opt.fix_word_vecs_enc) model.decoder.embeddings.load_pretrained_vectors( model_opt.pre_word_vecs_dec, model_opt.fix_word_vecs_dec) # Add generator to model (this registers it as parameter of model). model.generator = generator # Make the whole model leverage GPU if indicated to do so. if gpu: model.cuda() else: model.cpu() return model

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/OpenNMT/onmt/Trainer.py

from __future__ import division """ This is the loadable seq2seq trainer library that is in charge of training details, loss compute, and statistics. See train.py for a use case of this library. Note!!! To make this a general library, we implement *only* mechanism things here(i.e. what to do), and leave the strategy things to users(i.e. how to do it). Also see train.py(one of the users of this library) for the strategy things we do. """ import time import sys import math import torch import torch.nn as nn import onmt import onmt.modules class Statistics(object): """ Train/validate loss statistics. """ def __init__(self, loss=0, n_words=0, n_correct=0): self.loss = loss self.n_words = n_words self.n_correct = n_correct self.n_src_words = 0 self.start_time = time.time() def update(self, stat): self.loss += stat.loss self.n_words += stat.n_words self.n_correct += stat.n_correct def accuracy(self): return 100 * (self.n_correct / self.n_words) def ppl(self): return math.exp(min(self.loss / self.n_words, 100)) def elapsed_time(self): return time.time() - self.start_time def output(self, epoch, batch, n_batches, start): t = self.elapsed_time() print(("Epoch %2d, %5d/%5d; acc: %6.2f; ppl: %6.2f; " + "%3.0f src tok/s; %3.0f tgt tok/s; %6.0f s elapsed") % (epoch, batch, n_batches, self.accuracy(), self.ppl(), self.n_src_words / (t + 1e-5), self.n_words / (t + 1e-5), time.time() - start)) sys.stdout.flush() def log(self, prefix, experiment, lr): t = self.elapsed_time() experiment.add_scalar_value(prefix + "_ppl", self.ppl()) experiment.add_scalar_value(prefix + "_accuracy", self.accuracy()) experiment.add_scalar_value(prefix + "_tgtper", self.n_words / t) experiment.add_scalar_value(prefix + "_lr", lr) class Trainer(object): def __init__(self, model, train_iter, valid_iter, train_loss, valid_loss, optim, trunc_size, shard_size): """ Args: model: the seq2seq model. train_iter: the train data iterator. valid_iter: the validate data iterator. train_loss: the train side LossCompute object for computing loss. valid_loss: the valid side LossCompute object for computing loss. optim: the optimizer responsible for lr update. trunc_size: a batch is divided by several truncs of this size. shard_size: compute loss in shards of this size for efficiency. """ # Basic attributes. self.model = model self.train_iter = train_iter self.valid_iter = valid_iter self.train_loss = train_loss self.valid_loss = valid_loss self.optim = optim self.trunc_size = trunc_size self.shard_size = shard_size # Set model in training mode. self.model.train() def train(self, epoch, report_func=None): """ Called for each epoch to train. """ total_stats = Statistics() report_stats = Statistics() for i, batch in enumerate(self.train_iter): target_size = batch.tgt.size(0) # Truncated BPTT trunc_size = self.trunc_size if self.trunc_size else target_size dec_state = None _, src_lengths = batch.src src = onmt.IO.make_features(batch, 'src') tgt_outer = onmt.IO.make_features(batch, 'tgt') report_stats.n_src_words += src_lengths.sum() for j in range(0, target_size-1, trunc_size): # 1. Create truncated target. tgt = tgt_outer[j: j + trunc_size] # 2. F-prop all but generator. self.model.zero_grad() outputs, attns, dec_state = \ self.model(src, tgt, src_lengths, dec_state) # 3. Compute loss in shards for memory efficiency. batch_stats = self.train_loss.sharded_compute_loss( batch, outputs, attns, j, trunc_size, self.shard_size) # 4. Update the parameters and statistics. self.optim.step() total_stats.update(batch_stats) report_stats.update(batch_stats) # If truncated, don't backprop fully. if dec_state is not None: dec_state.detach() if report_func is not None: report_stats = report_func( epoch, i, len(self.train_iter), total_stats.start_time, self.optim.lr, report_stats) return total_stats def validate(self): """ Called for each epoch to validate. """ # Set model in validating mode. self.model.eval() stats = Statistics() for batch in self.valid_iter: _, src_lengths = batch.src src = onmt.IO.make_features(batch, 'src') tgt = onmt.IO.make_features(batch, 'tgt') # F-prop through the model. outputs, attns, _ = self.model(src, tgt, src_lengths) # Compute loss. batch_stats = self.valid_loss.monolithic_compute_loss( batch, outputs, attns) # Update statistics. stats.update(batch_stats) # Set model back to training mode. self.model.train() return stats def epoch_step(self, ppl, epoch): """ Called for each epoch to update learning rate. """ return self.optim.updateLearningRate(ppl, epoch) def drop_checkpoint(self, opt, epoch, fields, valid_stats): """ Called conditionally each epoch to save a snapshot. """ real_model = (self.model.module if isinstance(self.model, nn.DataParallel) else self.model) real_generator = (real_model.generator.module if isinstance(real_model.generator, nn.DataParallel) else real_model.generator) model_state_dict = real_model.state_dict() model_state_dict = {k: v for k, v in model_state_dict.items() if 'generator' not in k} generator_state_dict = real_generator.state_dict() checkpoint = { 'model': model_state_dict, 'generator': generator_state_dict, 'vocab': onmt.IO.save_vocab(fields), 'opt': opt, 'epoch': epoch, 'optim': self.optim } torch.save(checkpoint, '%s.pt' % opt.save_model)

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/OpenNMT/onmt/Optim.py

import torch.optim as optim from torch.nn.utils import clip_grad_norm class Optim(object): def set_parameters(self, params): self.params = [p for p in params if p.requires_grad] if self.method == 'sgd': self.optimizer = optim.SGD(self.params, lr=self.lr) elif self.method == 'adagrad': self.optimizer = optim.Adagrad(self.params, lr=self.lr) elif self.method == 'adadelta': self.optimizer = optim.Adadelta(self.params, lr=self.lr) elif self.method == 'adam': self.optimizer = optim.Adam(self.params, lr=self.lr, betas=self.betas, eps=1e-9) else: raise RuntimeError("Invalid optim method: " + self.method) def __init__(self, method, lr, max_grad_norm, lr_decay=1, start_decay_at=None, beta1=0.9, beta2=0.98, opt=None): self.last_ppl = None self.lr = lr self.max_grad_norm = max_grad_norm self.method = method self.lr_decay = lr_decay self.start_decay_at = start_decay_at self.start_decay = False self._step = 0 self.betas = [beta1, beta2] self.opt = opt def _setRate(self, lr): self.lr = lr self.optimizer.param_groups[0]['lr'] = self.lr def step(self): "Compute gradients norm." self._step += 1 # Decay method used in tensor2tensor. if self.opt.__dict__.get("decay_method", "") == "noam": self._setRate( self.opt.learning_rate * (self.opt.rnn_size ** (-0.5) * min(self._step ** (-0.5), self._step * self.opt.warmup_steps**(-1.5)))) if self.max_grad_norm: clip_grad_norm(self.params, self.max_grad_norm) self.optimizer.step() def updateLearningRate(self, ppl, epoch): """ Decay learning rate if val perf does not improve or we hit the start_decay_at limit. """ if self.start_decay_at is not None and epoch >= self.start_decay_at: self.start_decay = True if self.last_ppl is not None and ppl > self.last_ppl: self.start_decay = True if self.start_decay: self.lr = self.lr * self.lr_decay print("Decaying learning rate to %g" % self.lr) self.last_ppl = ppl self.optimizer.param_groups[0]['lr'] = self.lr

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/OpenNMT/onmt/Models.py

from __future__ import division import torch import torch.nn as nn from torch.autograd import Variable from torch.nn.utils.rnn import pack_padded_sequence as pack from torch.nn.utils.rnn import pad_packed_sequence as unpack import onmt from onmt.Utils import aeq class EncoderBase(nn.Module): """ EncoderBase class for sharing code among various encoder. """ def _check_args(self, input, lengths=None, hidden=None): s_len, n_batch, n_feats = input.size() if lengths is not None: n_batch_, = lengths.size() aeq(n_batch, n_batch_) def forward(self, input, lengths=None, hidden=None): """ Args: input (LongTensor): len x batch x nfeat. lengths (LongTensor): batch hidden: Initial hidden state. Returns: hidden_t (Variable): Pair of layers x batch x rnn_size - final encoder state outputs (FloatTensor): len x batch x rnn_size - Memory bank """ raise NotImplementedError class MeanEncoder(EncoderBase): """ A trivial encoder without RNN, just takes mean as final state. """ def __init__(self, num_layers, embeddings): super(MeanEncoder, self).__init__() self.num_layers = num_layers self.embeddings = embeddings def forward(self, input, lengths=None, hidden=None): """ See EncoderBase.forward() for description of args and returns. """ self._check_args(input, lengths, hidden) emb = self.embeddings(input) s_len, batch, emb_dim = emb.size() mean = emb.mean(0).expand(self.num_layers, batch, emb_dim) return (mean, mean), emb class RNNEncoder(EncoderBase): """ The standard RNN encoder. """ def __init__(self, rnn_type, bidirectional, num_layers, hidden_size, dropout, embeddings): super(RNNEncoder, self).__init__() num_directions = 2 if bidirectional else 1 assert hidden_size % num_directions == 0 hidden_size = hidden_size // num_directions self.embeddings = embeddings self.no_pack_padded_seq = False # Use pytorch version when available. if rnn_type == "SRU": # SRU doesn't support PackedSequence. self.no_pack_padded_seq = True self.rnn = onmt.modules.SRU( input_size=embeddings.embedding_size, hidden_size=hidden_size, num_layers=num_layers, dropout=dropout, bidirectional=bidirectional) else: self.rnn = getattr(nn, rnn_type)( input_size=embeddings.embedding_size, hidden_size=hidden_size, num_layers=num_layers, dropout=dropout, bidirectional=bidirectional) def forward(self, input, lengths=None, hidden=None): """ See EncoderBase.forward() for description of args and returns.""" self._check_args(input, lengths, hidden) emb = self.embeddings(input) s_len, batch, emb_dim = emb.size() packed_emb = emb if lengths is not None and not self.no_pack_padded_seq: # Lengths data is wrapped inside a Variable. lengths = lengths.view(-1).tolist() packed_emb = pack(emb, lengths) outputs, hidden_t = self.rnn(packed_emb, hidden) if lengths is not None and not self.no_pack_padded_seq: outputs = unpack(outputs)[0] return hidden_t, outputs class RNNDecoderBase(nn.Module): """ RNN decoder base class. """ def __init__(self, rnn_type, bidirectional_encoder, num_layers, hidden_size, attn_type, coverage_attn, context_gate, copy_attn, dropout, embeddings): super(RNNDecoderBase, self).__init__() # Basic attributes. self.decoder_type = 'rnn' self.bidirectional_encoder = bidirectional_encoder self.num_layers = num_layers self.hidden_size = hidden_size self.embeddings = embeddings self.dropout = nn.Dropout(dropout) # Build the RNN. self.rnn = self._build_rnn(rnn_type, self._input_size, hidden_size, num_layers, dropout) # Set up the context gate. self.context_gate = None if context_gate is not None: self.context_gate = onmt.modules.ContextGateFactory( context_gate, self._input_size, hidden_size, hidden_size, hidden_size ) # Set up the standard attention. self._coverage = coverage_attn self.attn = onmt.modules.GlobalAttention( hidden_size, coverage=coverage_attn, attn_type=attn_type ) # Set up a separated copy attention layer, if needed. self._copy = False if copy_attn: self.copy_attn = onmt.modules.GlobalAttention( hidden_size, attn_type=attn_type ) self._copy = True def forward(self, input, context, state): """ Forward through the decoder. Args: input (LongTensor): a sequence of input tokens tensors of size (len x batch x nfeats). context (FloatTensor): output(tensor sequence) from the encoder RNN of size (src_len x batch x hidden_size). state (FloatTensor): hidden state from the encoder RNN for initializing the decoder. Returns: outputs (FloatTensor): a Tensor sequence of output from the decoder of shape (len x batch x hidden_size). state (FloatTensor): final hidden state from the decoder. attns (dict of (str, FloatTensor)): a dictionary of different type of attention Tensor from the decoder of shape (src_len x batch). """ # Args Check assert isinstance(state, RNNDecoderState) input_len, input_batch, _ = input.size() contxt_len, contxt_batch, _ = context.size() aeq(input_batch, contxt_batch) # END Args Check # Run the forward pass of the RNN. hidden, outputs, attns, coverage = \ self._run_forward_pass(input, context, state) # Update the state with the result. final_output = outputs[-1] state.update_state(hidden, final_output.unsqueeze(0), coverage.unsqueeze(0) if coverage is not None else None) # Concatenates sequence of tensors along a new dimension. outputs = torch.stack(outputs) for k in attns: attns[k] = torch.stack(attns[k]) return outputs, state, attns def _fix_enc_hidden(self, h): """ The encoder hidden is (layers*directions) x batch x dim. We need to convert it to layers x batch x (directions*dim). """ if self.bidirectional_encoder: h = torch.cat([h[0:h.size(0):2], h[1:h.size(0):2]], 2) return h def init_decoder_state(self, src, context, enc_hidden): if isinstance(enc_hidden, tuple): # LSTM return RNNDecoderState(context, self.hidden_size, tuple([self._fix_enc_hidden(enc_hidden[i]) for i in range(len(enc_hidden))])) else: # GRU return RNNDecoderState(context, self.hidden_size, self._fix_enc_hidden(enc_hidden)) class StdRNNDecoder(RNNDecoderBase): """ Stardard RNN decoder, with Attention. Currently no 'coverage_attn' and 'copy_attn' support. """ def _run_forward_pass(self, input, context, state): """ Private helper for running the specific RNN forward pass. Must be overriden by all subclasses. Args: input (LongTensor): a sequence of input tokens tensors of size (len x batch x nfeats). context (FloatTensor): output(tensor sequence) from the encoder RNN of size (src_len x batch x hidden_size). state (FloatTensor): hidden state from the encoder RNN for initializing the decoder. Returns: hidden (Variable): final hidden state from the decoder. outputs ([FloatTensor]): an array of output of every time step from the decoder. attns (dict of (str, [FloatTensor]): a dictionary of different type of attention Tensor array of every time step from the decoder. coverage (FloatTensor, optional): coverage from the decoder. """ assert not self._copy # TODO, no support yet. assert not self._coverage # TODO, no support yet. # Initialize local and return variables. outputs = [] attns = {"std": []} coverage = None emb = self.embeddings(input) # Run the forward pass of the RNN. if isinstance(self.rnn, nn.GRU): rnn_output, hidden = self.rnn(emb, state.hidden[0]) else: rnn_output, hidden = self.rnn(emb, state.hidden) # Result Check input_len, input_batch, _ = input.size() output_len, output_batch, _ = rnn_output.size() aeq(input_len, output_len) aeq(input_batch, output_batch) # END Result Check # Calculate the attention. attn_outputs, attn_scores = self.attn( rnn_output.transpose(0, 1).contiguous(), # (output_len, batch, d) context.transpose(0, 1) # (contxt_len, batch, d) ) attns["std"] = attn_scores # Calculate the context gate. if self.context_gate is not None: outputs = self.context_gate( emb.view(-1, emb.size(2)), rnn_output.view(-1, rnn_output.size(2)), attn_outputs.view(-1, attn_outputs.size(2)) ) outputs = outputs.view(input_len, input_batch, self.hidden_size) outputs = self.dropout(outputs) else: outputs = self.dropout(attn_outputs) # (input_len, batch, d) # Return result. return hidden, outputs, attns, coverage def _build_rnn(self, rnn_type, input_size, hidden_size, num_layers, dropout): """ Private helper for building standard decoder RNN. """ # Use pytorch version when available. if rnn_type == "SRU": return onmt.modules.SRU( input_size, hidden_size, num_layers=num_layers, dropout=dropout) return getattr(nn, rnn_type)( input_size, hidden_size, num_layers=num_layers, dropout=dropout) @property def _input_size(self): """ Private helper returning the number of expected features. """ return self.embeddings.embedding_size class InputFeedRNNDecoder(RNNDecoderBase): """ Stardard RNN decoder, with Input Feed and Attention. """ def _run_forward_pass(self, input, context, state): """ See StdRNNDecoder._run_forward_pass() for description of arguments and return values. """ # Additional args check. output = state.input_feed.squeeze(0) output_batch, _ = output.size() input_len, input_batch, _ = input.size() aeq(input_batch, output_batch) # END Additional args check. # Initialize local and return variables. outputs = [] attns = {"std": []} if self._copy: attns["copy"] = [] if self._coverage: attns["coverage"] = [] emb = self.embeddings(input) assert emb.dim() == 3 # len x batch x embedding_dim hidden = state.hidden coverage = state.coverage.squeeze(0) \ if state.coverage is not None else None # Input feed concatenates hidden state with # input at every time step. for i, emb_t in enumerate(emb.split(1)): emb_t = emb_t.squeeze(0) emb_t = torch.cat([emb_t, output], 1) rnn_output, hidden = self.rnn(emb_t, hidden) attn_output, attn = self.attn(rnn_output, context.transpose(0, 1)) if self.context_gate is not None: output = self.context_gate( emb_t, rnn_output, attn_output ) output = self.dropout(output) else: output = self.dropout(attn_output) outputs += [output] attns["std"] += [attn] # Update the coverage attention. if self._coverage: coverage = coverage + attn \ if coverage is not None else attn attns["coverage"] += [coverage] # Run the forward pass of the copy attention layer. if self._copy: _, copy_attn = self.copy_attn(output, context.transpose(0, 1)) attns["copy"] += [copy_attn] # Return result. return hidden, outputs, attns, coverage def _build_rnn(self, rnn_type, input_size, hidden_size, num_layers, dropout): assert not rnn_type == "SRU", "SRU doesn't support input feed! " \ "Please set -input_feed 0!" if rnn_type == "LSTM": stacked_cell = onmt.modules.StackedLSTM else: stacked_cell = onmt.modules.StackedGRU return stacked_cell(num_layers, input_size, hidden_size, dropout) @property def _input_size(self): """ Using input feed by concatenating input with attention vectors. """ return self.embeddings.embedding_size + self.hidden_size class NMTModel(nn.Module): """ The encoder + decoder Neural Machine Translation Model. """ def __init__(self, encoder, decoder, multigpu=False): """ Args: encoder(*Encoder): the various encoder. decoder(*Decoder): the various decoder. multigpu(bool): run parellel on multi-GPU? """ self.multigpu = multigpu super(NMTModel, self).__init__() self.encoder = encoder self.decoder = decoder def forward(self, src, tgt, lengths, dec_state=None): """ Args: src(FloatTensor): a sequence of source tensors with optional feature tensors of size (len x batch). tgt(FloatTensor): a sequence of target tensors with optional feature tensors of size (len x batch). lengths([int]): an array of the src length. dec_state: A decoder state object Returns: outputs (FloatTensor): (len x batch x hidden_size): decoder outputs attns (FloatTensor): Dictionary of (src_len x batch) dec_hidden (FloatTensor): tuple (1 x batch x hidden_size) Init hidden state """ src = src tgt = tgt[:-1] # exclude last target from inputs enc_hidden, context = self.encoder(src, lengths) enc_state = self.decoder.init_decoder_state(src, context, enc_hidden) out, dec_state, attns = self.decoder(tgt, context, enc_state if dec_state is None else dec_state) if self.multigpu: # Not yet supported on multi-gpu dec_state = None attns = None return out, attns, dec_state class DecoderState(object): """ DecoderState is a base class for models, used during translation for storing translation states. """ def detach(self): """ Detaches all Variables from the graph that created it, making it a leaf. """ for h in self._all: if h is not None: h.detach_() def beam_update(self, idx, positions, beam_size): """ Update when beam advances. """ for e in self._all: a, br, d = e.size() sentStates = e.view(a, beam_size, br // beam_size, d)[:, :, idx] sentStates.data.copy_( sentStates.data.index_select(1, positions)) class RNNDecoderState(DecoderState): def __init__(self, context, hidden_size, rnnstate): """ Args: context (FloatTensor): output from the encoder of size len x batch x rnn_size. hidden_size (int): the size of hidden layer of the decoder. rnnstate (Variable): final hidden state from the encoder. transformed to shape: layers x batch x (directions*dim). input_feed (FloatTensor): output from last layer of the decoder. coverage (FloatTensor): coverage output from the decoder. """ if not isinstance(rnnstate, tuple): self.hidden = (rnnstate,) else: self.hidden = rnnstate self.coverage = None # Init the input feed. batch_size = context.size(1) h_size = (batch_size, hidden_size) self.input_feed = Variable(context.data.new(*h_size).zero_(), requires_grad=False).unsqueeze(0) @property def _all(self): return self.hidden + (self.input_feed,) def update_state(self, rnnstate, input_feed, coverage): if not isinstance(rnnstate, tuple): self.hidden = (rnnstate,) else: self.hidden = rnnstate self.input_feed = input_feed self.coverage = coverage def repeat_beam_size_times(self, beam_size): """ Repeat beam_size times along batch dimension. """ vars = [Variable(e.data.repeat(1, beam_size, 1), volatile=True) for e in self._all] self.hidden = tuple(vars[:-1]) self.input_feed = vars[-1]

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/OpenNMT/onmt/modules/ConvMultiStepAttention.py

import torch import torch.nn as nn import torch.nn.functional as F from onmt.Utils import aeq SCALE_WEIGHT = 0.5 ** 0.5 def seq_linear(linear, x): # linear transform for 3-d tensor batch, hidden_size, length, _ = x.size() h = linear(torch.transpose(x, 1, 2).contiguous().view( batch * length, hidden_size)) return torch.transpose(h.view(batch, length, hidden_size, 1), 1, 2) class ConvMultiStepAttention(nn.Module): def __init__(self, input_size): super(ConvMultiStepAttention, self).__init__() self.linear_in = nn.Linear(input_size, input_size) self.mask = None def applyMask(self, mask): self.mask = mask def forward(self, base_target_emb, input, encoder_out_top, encoder_out_combine): """ It's like Luong Attetion. Conv attention takes a key matrix, a value matrix and a query vector. Attention weight is calculated by key matrix with the query vector and sum on the value matrix. And the same operation is applied in each decode conv layer. Args: base_target_emb: target emb tensor input: output of decode conv encoder_out_t: the key matrix for calculation of attetion weight, which is the top output of encode conv encoder_out_c: the value matrix for the attention-weighted sum, which is the combination of base emb and top output of encode """ # checks batch, channel, height, width = base_target_emb.size() batch_, channel_, height_, width_ = input.size() aeq(batch, batch_) aeq(height, height_) enc_batch, enc_channel, enc_height = encoder_out_top.size() enc_batch_, enc_channel_, enc_height_ = encoder_out_combine.size() aeq(enc_batch, enc_batch_) aeq(enc_height, enc_height_) preatt = seq_linear(self.linear_in, input) target = (base_target_emb + preatt) * SCALE_WEIGHT target = torch.squeeze(target, 3) target = torch.transpose(target, 1, 2) pre_attn = torch.bmm(target, encoder_out_top) if self.mask is not None: pre_attn.data.masked_fill_(self.mask, -float('inf')) pre_attn = pre_attn.transpose(0, 2) attn = F.softmax(pre_attn) attn = attn.transpose(0, 2).contiguous() context_output = torch.bmm( attn, torch.transpose(encoder_out_combine, 1, 2)) context_output = torch.transpose( torch.unsqueeze(context_output, 3), 1, 2) return context_output, attn

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/OpenNMT/onmt/modules/Transformer.py

""" Implementation of "Attention is All You Need" """ import torch import torch.nn as nn from torch.autograd import Variable import numpy as np import onmt from onmt.Models import EncoderBase from onmt.Models import DecoderState from onmt.Utils import aeq MAX_SIZE = 5000 class PositionwiseFeedForward(nn.Module): """ A two-layer Feed-Forward-Network.""" def __init__(self, size, hidden_size, dropout=0.1): """ Args: size(int): the size of input for the first-layer of the FFN. hidden_size(int): the hidden layer size of the second-layer of the FNN. droput(float): dropout probability(0-1.0). """ super(PositionwiseFeedForward, self).__init__() self.w_1 = onmt.modules.BottleLinear(size, hidden_size) self.w_2 = onmt.modules.BottleLinear(hidden_size, size) self.layer_norm = onmt.modules.BottleLayerNorm(size) self.dropout = nn.Dropout(dropout) self.relu = nn.ReLU() def forward(self, x): residual = x output = self.dropout(self.w_2(self.relu(self.w_1(x)))) return self.layer_norm(output + residual) class TransformerEncoderLayer(nn.Module): def __init__(self, size, dropout, head_count=8, hidden_size=2048): """ Args: size(int): the dimension of keys/values/queries in MultiHeadedAttention, also the input size of the first-layer of the PositionwiseFeedForward. droput(float): dropout probability(0-1.0). head_count(int): the number of head for MultiHeadedAttention. hidden_size(int): the second-layer of the PositionwiseFeedForward. """ super(TransformerEncoderLayer, self).__init__() self.self_attn = onmt.modules.MultiHeadedAttention( head_count, size, p=dropout) self.feed_forward = PositionwiseFeedForward(size, hidden_size, dropout) def forward(self, input, mask): mid, _ = self.self_attn(input, input, input, mask=mask) out = self.feed_forward(mid) return out class TransformerEncoder(EncoderBase): """ The Transformer encoder from "Attention is All You Need". """ def __init__(self, num_layers, hidden_size, dropout, embeddings): super(TransformerEncoder, self).__init__() self.num_layers = num_layers self.embeddings = embeddings self.transformer = nn.ModuleList( [TransformerEncoderLayer(hidden_size, dropout) for i in range(num_layers)]) def forward(self, input, lengths=None, hidden=None): """ See EncoderBase.forward() for description of args and returns.""" self._check_args(input, lengths, hidden) emb = self.embeddings(input) s_len, n_batch, emb_dim = emb.size() out = emb.transpose(0, 1).contiguous() words = input[:, :, 0].transpose(0, 1) # CHECKS out_batch, out_len, _ = out.size() w_batch, w_len = words.size() aeq(out_batch, w_batch) aeq(out_len, w_len) # END CHECKS # Make mask. padding_idx = self.embeddings.word_padding_idx mask = words.data.eq(padding_idx).unsqueeze(1) \ .expand(w_batch, w_len, w_len) # Run the forward pass of every layer of the tranformer. for i in range(self.num_layers): out = self.transformer[i](out, mask) return Variable(emb.data), out.transpose(0, 1).contiguous() class TransformerDecoderLayer(nn.Module): def __init__(self, size, dropout, head_count=8, hidden_size=2048): """ Args: size(int): the dimension of keys/values/queries in MultiHeadedAttention, also the input size of the first-layer of the PositionwiseFeedForward. droput(float): dropout probability(0-1.0). head_count(int): the number of head for MultiHeadedAttention. hidden_size(int): the second-layer of the PositionwiseFeedForward. """ super(TransformerDecoderLayer, self).__init__() self.self_attn = onmt.modules.MultiHeadedAttention( head_count, size, p=dropout) self.context_attn = onmt.modules.MultiHeadedAttention( head_count, size, p=dropout) self.feed_forward = PositionwiseFeedForward(size, hidden_size, dropout) self.dropout = dropout mask = self._get_attn_subsequent_mask(MAX_SIZE) # Register self.mask as a buffer in TransformerDecoderLayer, so # it gets TransformerDecoderLayer's cuda behavior automatically. self.register_buffer('mask', mask) def forward(self, input, context, src_pad_mask, tgt_pad_mask): # Args Checks input_batch, input_len, _ = input.size() contxt_batch, contxt_len, _ = context.size() aeq(input_batch, contxt_batch) src_batch, t_len, s_len = src_pad_mask.size() tgt_batch, t_len_, t_len__ = tgt_pad_mask.size() aeq(input_batch, contxt_batch, src_batch, tgt_batch) aeq(t_len, t_len_, t_len__, input_len) aeq(s_len, contxt_len) # END Args Checks dec_mask = torch.gt(tgt_pad_mask + self.mask[:, :tgt_pad_mask.size(1), :tgt_pad_mask.size(1)] .expand_as(tgt_pad_mask), 0) query, attn = self.self_attn(input, input, input, mask=dec_mask) mid, attn = self.context_attn(context, context, query, mask=src_pad_mask) output = self.feed_forward(mid) # CHECKS output_batch, output_len, _ = output.size() aeq(input_len, output_len) aeq(contxt_batch, output_batch) n_batch_, t_len_, s_len_ = attn.size() aeq(input_batch, n_batch_) aeq(contxt_len, s_len_) aeq(input_len, t_len_) # END CHECKS return output, attn def _get_attn_subsequent_mask(self, size): ''' Get an attention mask to avoid using the subsequent info.''' attn_shape = (1, size, size) subsequent_mask = np.triu(np.ones(attn_shape), k=1).astype('uint8') subsequent_mask = torch.from_numpy(subsequent_mask) return subsequent_mask class TransformerDecoder(nn.Module): """ The Transformer decoder from "Attention is All You Need". """ def __init__(self, num_layers, hidden_size, attn_type, copy_attn, dropout, embeddings): super(TransformerDecoder, self).__init__() # Basic attributes. self.decoder_type = 'transformer' self.num_layers = num_layers self.embeddings = embeddings # Build TransformerDecoder. self.transformer_layers = nn.ModuleList( [TransformerDecoderLayer(hidden_size, dropout) for _ in range(num_layers)]) # TransformerDecoder has its own attention mechanism. # Set up a separated copy attention layer, if needed. self._copy = False if copy_attn: self.copy_attn = onmt.modules.GlobalAttention( hidden_size, attn_type=attn_type) self._copy = True def forward(self, input, context, state): """ Forward through the TransformerDecoder. Args: input (LongTensor): a sequence of input tokens tensors of size (len x batch x nfeats). context (FloatTensor): output(tensor sequence) from the encoder of size (src_len x batch x hidden_size). state (FloatTensor): hidden state from the encoder RNN for initializing the decoder. Returns: outputs (FloatTensor): a Tensor sequence of output from the decoder of shape (len x batch x hidden_size). state (FloatTensor): final hidden state from the decoder. attns (dict of (str, FloatTensor)): a dictionary of different type of attention Tensor from the decoder of shape (src_len x batch). """ # CHECKS assert isinstance(state, TransformerDecoderState) input_len, input_batch, _ = input.size() contxt_len, contxt_batch, _ = context.size() aeq(input_batch, contxt_batch) if state.previous_input is not None: input = torch.cat([state.previous_input, input], 0) src = state.src src_words = src[:, :, 0].transpose(0, 1) tgt_words = input[:, :, 0].transpose(0, 1) src_batch, src_len = src_words.size() tgt_batch, tgt_len = tgt_words.size() aeq(input_batch, contxt_batch, src_batch, tgt_batch) aeq(contxt_len, src_len) # aeq(input_len, tgt_len) # END CHECKS # Initialize return variables. outputs = [] attns = {"std": []} if self._copy: attns["copy"] = [] # Run the forward pass of the TransformerDecoder. emb = self.embeddings(input) assert emb.dim() == 3 # len x batch x embedding_dim output = emb.transpose(0, 1).contiguous() src_context = context.transpose(0, 1).contiguous() padding_idx = self.embeddings.word_padding_idx src_pad_mask = src_words.data.eq(padding_idx).unsqueeze(1) \ .expand(src_batch, tgt_len, src_len) tgt_pad_mask = tgt_words.data.eq(padding_idx).unsqueeze(1) \ .expand(tgt_batch, tgt_len, tgt_len) for i in range(self.num_layers): output, attn \ = self.transformer_layers[i](output, src_context, src_pad_mask, tgt_pad_mask) # Process the result and update the attentions. outputs = output.transpose(0, 1).contiguous() if state.previous_input is not None: outputs = outputs[state.previous_input.size(0):] attn = attn[:, state.previous_input.size(0):].squeeze() attn = torch.stack([attn]) attns["std"] = attn if self._copy: attns["copy"] = attn # Update the state. state.update_state(input) return outputs, state, attns def init_decoder_state(self, src, context, enc_hidden): return TransformerDecoderState(src) class TransformerDecoderState(DecoderState): def __init__(self, src): """ Args: src (FloatTensor): a sequence of source words tensors with optional feature tensors, of size (len x batch). """ self.src = src self.previous_input = None @property def _all(self): """ Contains attributes that need to be updated in self.beam_update(). """ return (self.previous_input, self.src) def update_state(self, input): """ Called for every decoder forward pass. """ self.previous_input = input def repeat_beam_size_times(self, beam_size): """ Repeat beam_size times along batch dimension. """ self.src = Variable(self.src.data.repeat(1, beam_size, 1), volatile=True)

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/OpenNMT/onmt/modules/Embeddings.py

import torch import torch.nn as nn from torch.autograd import Variable from onmt.modules import BottleLinear, Elementwise from onmt.Utils import aeq class PositionalEncoding(nn.Module): def __init__(self, dropout, dim, max_len=5000): pe = torch.arange(0, max_len).unsqueeze(1).expand(max_len, dim) div_term = 1 / torch.pow(10000, torch.arange(0, dim * 2, 2) / dim) pe = pe * div_term.expand_as(pe) pe[:, 0::2] = torch.sin(pe[:, 0::2]) pe[:, 1::2] = torch.cos(pe[:, 1::2]) pe = pe.unsqueeze(1) super(PositionalEncoding, self).__init__() self.register_buffer('pe', pe) self.dropout = nn.Dropout(p=dropout) def forward(self, emb): # We must wrap the self.pe in Variable to compute, not the other # way - unwrap emb(i.e. emb.data). Otherwise the computation # wouldn't be watched to build the compute graph. emb = emb + Variable(self.pe[:emb.size(0), :1, :emb.size(2)] .expand_as(emb), requires_grad=False) emb = self.dropout(emb) return emb class Embeddings(nn.Module): """ Words embeddings dictionary for encoder/decoder. Args: word_vec_size (int): size of the dictionary of embeddings. position_encoding (bool): use a sin to mark relative words positions. feat_merge (string): merge action for the features embeddings: concat, sum or mlp. feat_vec_exponent (float): when using '-feat_merge concat', feature embedding size is N^feat_dim_exponent, where N is the number of values of feature takes. feat_vec_size (int): embedding dimension for features when using '-feat_merge mlp' dropout (float): dropout probability. word_padding_idx (int): padding index for words in the embeddings. feats_padding_idx ([int]): padding index for a list of features in the embeddings. word_vocab_size (int): size of dictionary of embeddings for words. feat_vocab_sizes ([int], optional): list of size of dictionary of embeddings for each feature. """ def __init__(self, word_vec_size, position_encoding, feat_merge, feat_vec_exponent, feat_vec_size, dropout, word_padding_idx, feat_padding_idx, word_vocab_size, feat_vocab_sizes=[]): self.word_padding_idx = word_padding_idx # Dimensions and padding for constructing the word embedding matrix vocab_sizes = [word_vocab_size] emb_dims = [word_vec_size] pad_indices = [word_padding_idx] # Dimensions and padding for feature embedding matrices # (these have no effect if feat_vocab_sizes is empty) if feat_merge == 'sum': feat_dims = [word_vec_size] * len(feat_vocab_sizes) elif feat_vec_size > 0: feat_dims = [feat_vec_size] * len(feat_vocab_sizes) else: feat_dims = [int(vocab ** feat_vec_exponent) for vocab in feat_vocab_sizes] vocab_sizes.extend(feat_vocab_sizes) emb_dims.extend(feat_dims) pad_indices.extend(feat_padding_idx) # The embedding matrix look-up tables. The first look-up table # is for words. Subsequent ones are for features, if any exist. emb_params = zip(vocab_sizes, emb_dims, pad_indices) embeddings = [nn.Embedding(vocab, dim, padding_idx=pad) for vocab, dim, pad in emb_params] emb_luts = Elementwise(feat_merge, embeddings) # The final output size of word + feature vectors. This can vary # from the word vector size if and only if features are defined. # This is the attribute you should access if you need to know # how big your embeddings are going to be. self.embedding_size = (sum(emb_dims) if feat_merge == 'concat' else word_vec_size) # The sequence of operations that converts the input sequence # into a sequence of embeddings. At minimum this consists of # looking up the embeddings for each word and feature in the # input. Model parameters may require the sequence to contain # additional operations as well. super(Embeddings, self).__init__() self.make_embedding = nn.Sequential() self.make_embedding.add_module('emb_luts', emb_luts) if feat_merge == 'mlp': in_dim = sum(emb_dims) out_dim = word_vec_size mlp = nn.Sequential(BottleLinear(in_dim, out_dim), nn.ReLU()) self.make_embedding.add_module('mlp', mlp) if position_encoding: pe = PositionalEncoding(dropout, self.embedding_size) self.make_embedding.add_module('pe', pe) @property def word_lut(self): return self.make_embedding[0][0] @property def emb_luts(self): return self.make_embedding[0] def load_pretrained_vectors(self, emb_file, fixed): if emb_file: pretrained = torch.load(emb_file) self.word_lut.weight.data.copy_(pretrained) if fixed: self.word_lut.weight.requires_grad = False def forward(self, input): """ Return the embeddings for words, and features if there are any. Args: input (LongTensor): len x batch x nfeat Return: emb (FloatTensor): len x batch x self.embedding_size """ in_length, in_batch, nfeat = input.size() aeq(nfeat, len(self.emb_luts)) emb = self.make_embedding(input) out_length, out_batch, emb_size = emb.size() aeq(in_length, out_length) aeq(in_batch, out_batch) aeq(emb_size, self.embedding_size) return emb

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/OpenNMT/onmt/modules/CopyGenerator.py

import torch.nn as nn import torch.nn.functional as F import torch import torch.cuda import onmt from onmt.Utils import aeq class CopyGenerator(nn.Module): """ Generator module that additionally considers copying words directly from the source. """ def __init__(self, opt, src_dict, tgt_dict): super(CopyGenerator, self).__init__() self.linear = nn.Linear(opt.rnn_size, len(tgt_dict)) self.linear_copy = nn.Linear(opt.rnn_size, 1) self.src_dict = src_dict self.tgt_dict = tgt_dict def forward(self, hidden, attn, src_map): """ Computes p(w) = p(z=1) p_{copy}(w|z=0) + p(z=0) * p_{softmax}(w|z=0) """ # CHECKS batch_by_tlen, _ = hidden.size() batch_by_tlen_, slen = attn.size() slen_, batch, cvocab = src_map.size() aeq(batch_by_tlen, batch_by_tlen_) aeq(slen, slen_) # Original probabilities. logits = self.linear(hidden) logits[:, self.tgt_dict.stoi[onmt.IO.PAD_WORD]] = -float('inf') prob = F.softmax(logits) # Probability of copying p(z=1) batch. copy = F.sigmoid(self.linear_copy(hidden)) # Probibility of not copying: p_{word}(w) * (1 - p(z)) out_prob = torch.mul(prob, 1 - copy.expand_as(prob)) mul_attn = torch.mul(attn, copy.expand_as(attn)) copy_prob = torch.bmm(mul_attn.view(-1, batch, slen) .transpose(0, 1), src_map.transpose(0, 1)).transpose(0, 1) copy_prob = copy_prob.contiguous().view(-1, cvocab) return torch.cat([out_prob, copy_prob], 1) class CopyGeneratorCriterion(object): def __init__(self, vocab_size, force_copy, pad, eps=1e-20): self.force_copy = force_copy self.eps = eps self.offset = vocab_size self.pad = pad def __call__(self, scores, align, target): align = align.view(-1) # Copy prob. out = scores.gather(1, align.view(-1, 1) + self.offset) \ .view(-1).mul(align.ne(0).float()) tmp = scores.gather(1, target.view(-1, 1)).view(-1) # Regular prob (no unks and unks that can't be copied) if not self.force_copy: out = out + self.eps + tmp.mul(target.ne(0).float()) + \ tmp.mul(align.eq(0).float()).mul(target.eq(0).float()) else: # Forced copy. out = out + self.eps + tmp.mul(align.eq(0).float()) # Drop padding. loss = -out.log().mul(target.ne(self.pad).float()).sum() return loss class CopyGeneratorLossCompute(onmt.Loss.LossComputeBase): """ Copy Generator Loss Computation. """ def __init__(self, generator, tgt_vocab, dataset, force_copy, eps=1e-20): super(CopyGeneratorLossCompute, self).__init__(generator, tgt_vocab) self.dataset = dataset self.force_copy = force_copy self.criterion = CopyGeneratorCriterion(len(tgt_vocab), force_copy, self.padding_idx) def make_shard_state(self, batch, output, range_, attns): """ See base class for args description. """ if getattr(batch, "alignment", None) is None: raise AssertionError("using -copy_attn you need to pass in " "-dynamic_dict during preprocess stage.") return { "output": output, "target": batch.tgt[range_[0] + 1: range_[1]], "copy_attn": attns.get("copy"), "align": batch.alignment[range_[0] + 1: range_[1]] } def compute_loss(self, batch, output, target, copy_attn, align): """ Compute the loss. The args must match self.make_shard_state(). Args: batch: the current batch. output: the predict output from the model. target: the validate target to compare output with. copy_attn: the copy attention value. align: the align info. """ target = target.view(-1) align = align.view(-1) scores = self.generator(self.bottle(output), self.bottle(copy_attn), batch.src_map) loss = self.criterion(scores, align, target) scores_data = scores.data.clone() scores_data = self.dataset.collapse_copy_scores( self.unbottle(scores_data, batch.batch_size), batch, self.tgt_vocab) scores_data = self.bottle(scores_data) # Correct target is copy when only option. # TODO: replace for loop with masking or boolean indexing target_data = target.data.clone() for i in range(target_data.size(0)): if target_data[i] == 0 and align.data[i] != 0: target_data[i] = align.data[i] + len(self.tgt_vocab) # Coverage loss term. loss_data = loss.data.clone() stats = self.stats(loss_data, scores_data, target_data) return loss, stats

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/OpenNMT/onmt/modules/StackedRNN.py

import torch import torch.nn as nn class StackedLSTM(nn.Module): """ Our own implementation of stacked LSTM. Needed for the decoder, because we do input feeding. """ def __init__(self, num_layers, input_size, rnn_size, dropout): super(StackedLSTM, self).__init__() self.dropout = nn.Dropout(dropout) self.num_layers = num_layers self.layers = nn.ModuleList() for i in range(num_layers): self.layers.append(nn.LSTMCell(input_size, rnn_size)) input_size = rnn_size def forward(self, input, hidden): h_0, c_0 = hidden h_1, c_1 = [], [] for i, layer in enumerate(self.layers): h_1_i, c_1_i = layer(input, (h_0[i], c_0[i])) input = h_1_i if i + 1 != self.num_layers: input = self.dropout(input) h_1 += [h_1_i] c_1 += [c_1_i] h_1 = torch.stack(h_1) c_1 = torch.stack(c_1) return input, (h_1, c_1) class StackedGRU(nn.Module): def __init__(self, num_layers, input_size, rnn_size, dropout): super(StackedGRU, self).__init__() self.dropout = nn.Dropout(dropout) self.num_layers = num_layers self.layers = nn.ModuleList() for i in range(num_layers): self.layers.append(nn.GRUCell(input_size, rnn_size)) input_size = rnn_size def forward(self, input, hidden): h_1 = [] for i, layer in enumerate(self.layers): h_1_i = layer(input, hidden[0][i]) input = h_1_i if i + 1 != self.num_layers: input = self.dropout(input) h_1 += [h_1_i] h_1 = torch.stack(h_1) return input, (h_1,)

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/OpenNMT/onmt/modules/MultiHeadedAttn.py

import math import torch import torch.nn as nn from torch.autograd import Variable from onmt.Utils import aeq from onmt.modules.UtilClass import BottleLinear, \ BottleLayerNorm, BottleSoftmax class MultiHeadedAttention(nn.Module): ''' Multi-Head Attention module from "Attention is All You Need". ''' def __init__(self, head_count, model_dim, p=0.1): """ Args: head_count(int): number of parallel heads. model_dim(int): the dimension of keys/values/queries in this MultiHeadedAttention, must be divisible by head_count. """ assert model_dim % head_count == 0 self.dim_per_head = model_dim // head_count self.model_dim = model_dim super(MultiHeadedAttention, self).__init__() self.head_count = head_count self.linear_keys = BottleLinear(model_dim, head_count * self.dim_per_head, bias=False) self.linear_values = BottleLinear(model_dim, head_count * self.dim_per_head, bias=False) self.linear_query = BottleLinear(model_dim, head_count * self.dim_per_head, bias=False) self.sm = BottleSoftmax() self.activation = nn.ReLU() self.layer_norm = BottleLayerNorm(model_dim) self.dropout = nn.Dropout(p) self.res_dropout = nn.Dropout(p) def forward(self, key, value, query, mask=None): # CHECKS batch, k_len, d = key.size() batch_, k_len_, d_ = value.size() aeq(batch, batch_) aeq(k_len, k_len_) aeq(d, d_) batch_, q_len, d_ = query.size() aeq(batch, batch_) aeq(d, d_) aeq(self.model_dim % 8, 0) if mask is not None: batch_, q_len_, k_len_ = mask.size() aeq(batch_, batch) aeq(k_len_, k_len) aeq(q_len_ == q_len) # END CHECKS def shape_projection(x): b, l, d = x.size() return x.view(b, l, self.head_count, self.dim_per_head) \ .transpose(1, 2).contiguous() \ .view(b * self.head_count, l, self.dim_per_head) def unshape_projection(x, q): b, l, d = q.size() return x.view(b, self.head_count, l, self.dim_per_head) \ .transpose(1, 2).contiguous() \ .view(b, l, self.head_count * self.dim_per_head) residual = query key_up = shape_projection(self.linear_keys(key)) value_up = shape_projection(self.linear_values(value)) query_up = shape_projection(self.linear_query(query)) scaled = torch.bmm(query_up, key_up.transpose(1, 2)) scaled = scaled / math.sqrt(self.dim_per_head) bh, l, dim_per_head = scaled.size() b = bh // self.head_count if mask is not None: scaled = scaled.view(b, self.head_count, l, dim_per_head) mask = mask.unsqueeze(1).expand_as(scaled) scaled = scaled.masked_fill(Variable(mask), -float('inf')) \ .view(bh, l, dim_per_head) attn = self.sm(scaled) # Return one attn top_attn = attn \ .view(b, self.head_count, l, dim_per_head)[:, 0, :, :] \ .contiguous() drop_attn = self.dropout(self.sm(scaled)) # values : (batch * 8) x qlen x dim out = unshape_projection(torch.bmm(drop_attn, value_up), residual) # Residual and layer norm res = self.res_dropout(out) + residual ret = self.layer_norm(res) # CHECK batch_, q_len_, d_ = ret.size() aeq(q_len, q_len_) aeq(batch, batch_) aeq(d, d_) # END CHECK return ret, top_attn

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/OpenNMT/onmt/modules/Gate.py

""" Context gate is a decoder module that takes as input the previous word embedding, the current decoder state and the attention state, and produces a gate. The gate can be used to select the input from the target side context (decoder state), from the source context (attention state) or both. """ import torch import torch.nn as nn def ContextGateFactory(type, embeddings_size, decoder_size, attention_size, output_size): """Returns the correct ContextGate class""" gate_types = {'source': SourceContextGate, 'target': TargetContextGate, 'both': BothContextGate} assert type in gate_types, "Not valid ContextGate type: {0}".format(type) return gate_types[type](embeddings_size, decoder_size, attention_size, output_size) class ContextGate(nn.Module): """Implement up to the computation of the gate""" def __init__(self, embeddings_size, decoder_size, attention_size, output_size): super(ContextGate, self).__init__() input_size = embeddings_size + decoder_size + attention_size self.gate = nn.Linear(input_size, output_size, bias=True) self.sig = nn.Sigmoid() self.source_proj = nn.Linear(attention_size, output_size) self.target_proj = nn.Linear(embeddings_size + decoder_size, output_size) def forward(self, prev_emb, dec_state, attn_state): input_tensor = torch.cat((prev_emb, dec_state, attn_state), dim=1) z = self.sig(self.gate(input_tensor)) proj_source = self.source_proj(attn_state) proj_target = self.target_proj( torch.cat((prev_emb, dec_state), dim=1)) return z, proj_source, proj_target class SourceContextGate(nn.Module): """Apply the context gate only to the source context""" def __init__(self, embeddings_size, decoder_size, attention_size, output_size): super(SourceContextGate, self).__init__() self.context_gate = ContextGate(embeddings_size, decoder_size, attention_size, output_size) self.tanh = nn.Tanh() def forward(self, prev_emb, dec_state, attn_state): z, source, target = self.context_gate( prev_emb, dec_state, attn_state) return self.tanh(target + z * source) class TargetContextGate(nn.Module): """Apply the context gate only to the target context""" def __init__(self, embeddings_size, decoder_size, attention_size, output_size): super(TargetContextGate, self).__init__() self.context_gate = ContextGate(embeddings_size, decoder_size, attention_size, output_size) self.tanh = nn.Tanh() def forward(self, prev_emb, dec_state, attn_state): z, source, target = self.context_gate(prev_emb, dec_state, attn_state) return self.tanh(z * target + source) class BothContextGate(nn.Module): """Apply the context gate to both contexts""" def __init__(self, embeddings_size, decoder_size, attention_size, output_size): super(BothContextGate, self).__init__() self.context_gate = ContextGate(embeddings_size, decoder_size, attention_size, output_size) self.tanh = nn.Tanh() def forward(self, prev_emb, dec_state, attn_state): z, source, target = self.context_gate(prev_emb, dec_state, attn_state) return self.tanh((1. - z) * target + z * source)

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/OpenNMT/onmt/modules/UtilClass.py

import torch import torch.nn as nn class Bottle(nn.Module): def forward(self, input): if len(input.size()) <= 2: return super(Bottle, self).forward(input) size = input.size()[:2] out = super(Bottle, self).forward(input.view(size[0]*size[1], -1)) return out.contiguous().view(size[0], size[1], -1) class Bottle2(nn.Module): def forward(self, input): if len(input.size()) <= 3: return super(Bottle2, self).forward(input) size = input.size() out = super(Bottle2, self).forward(input.view(size[0]*size[1], size[2], size[3])) return out.contiguous().view(size[0], size[1], size[2], size[3]) class LayerNorm(nn.Module): ''' Layer normalization module ''' def __init__(self, d_hid, eps=1e-3): super(LayerNorm, self).__init__() self.eps = eps self.a_2 = nn.Parameter(torch.ones(d_hid), requires_grad=True) self.b_2 = nn.Parameter(torch.zeros(d_hid), requires_grad=True) def forward(self, z): if z.size(1) == 1: return z mu = torch.mean(z, dim=1) sigma = torch.std(z, dim=1) # HACK. PyTorch is changing behavior if mu.dim() == 1: mu = mu.unsqueeze(1) sigma = sigma.unsqueeze(1) ln_out = (z - mu.expand_as(z)) / (sigma.expand_as(z) + self.eps) ln_out = ln_out.mul(self.a_2.expand_as(ln_out)) \ + self.b_2.expand_as(ln_out) return ln_out class BottleLinear(Bottle, nn.Linear): pass class BottleLayerNorm(Bottle, LayerNorm): pass class BottleSoftmax(Bottle, nn.Softmax): pass class Elementwise(nn.ModuleList): """ A simple network container. Parameters are a list of modules. Inputs are a 3d Variable whose last dimension is the same length as the list. Outputs are the result of applying modules to inputs elementwise. An optional merge parameter allows the outputs to be reduced to a single Variable. """ def __init__(self, merge=None, *args): assert merge in [None, 'first', 'concat', 'sum', 'mlp'] self.merge = merge super(Elementwise, self).__init__(*args) def forward(self, input): inputs = [feat.squeeze(2) for feat in input.split(1, dim=2)] assert len(self) == len(inputs) outputs = [f(x) for f, x in zip(self, inputs)] if self.merge == 'first': return outputs[0] elif self.merge == 'concat' or self.merge == 'mlp': return torch.cat(outputs, 2) elif self.merge == 'sum': return sum(outputs) else: return outputs

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/OpenNMT/onmt/modules/StructuredAttention.py

import torch.nn as nn import torch import torch.cuda from torch.autograd import Variable class MatrixTree(nn.Module): """Implementation of the matrix-tree theorem for computing marginals of non-projective dependency parsing. This attention layer is used in the paper "Learning Structured Text Representations." """ def __init__(self, eps=1e-5): self.eps = eps super(MatrixTree, self).__init__() def forward(self, input): laplacian = input.exp() + self.eps output = input.clone() for b in range(input.size(0)): lap = laplacian[b].masked_fill( Variable(torch.eye(input.size(1)).cuda().ne(0)), 0) lap = -lap + torch.diag(lap.sum(0)) # store roots on diagonal lap[0] = input[b].diag().exp() inv_laplacian = lap.inverse() factor = inv_laplacian.diag().unsqueeze(1)\ .expand_as(input[b]).transpose(0, 1) term1 = input[b].exp().mul(factor).clone() term2 = input[b].exp().mul(inv_laplacian.transpose(0, 1)).clone() term1[:, 0] = 0 term2[0] = 0 output[b] = term1 - term2 roots_output = input[b].diag().exp().mul( inv_laplacian.transpose(0, 1)[0]) output[b] = output[b] + torch.diag(roots_output) return output if __name__ == "__main__": dtree = MatrixTree() q = torch.rand(1, 5, 5).cuda() marg = dtree.forward(Variable(q)) print(marg.sum(1))

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/OpenNMT/onmt/modules/Conv2Conv.py

""" Implementation of "Convolutional Sequence to Sequence Learning" """ import torch import torch.nn as nn import torch.nn.init as init import torch.nn.functional as F from torch.autograd import Variable import onmt.modules from onmt.modules.WeightNorm import WeightNormConv2d from onmt.Models import EncoderBase from onmt.Models import DecoderState from onmt.Utils import aeq SCALE_WEIGHT = 0.5 ** 0.5 def shape_transform(x): """ Tranform the size of the tensors to fit for conv input. """ return torch.unsqueeze(torch.transpose(x, 1, 2), 3) class GatedConv(nn.Module): def __init__(self, input_size, width=3, dropout=0.2, nopad=False): super(GatedConv, self).__init__() self.conv = WeightNormConv2d(input_size, 2 * input_size, kernel_size=(width, 1), stride=(1, 1), padding=(width // 2 * (1 - nopad), 0)) init.xavier_uniform(self.conv.weight, gain=(4 * (1 - dropout))**0.5) self.dropout = nn.Dropout(dropout) def forward(self, x_var, hidden=None): x_var = self.dropout(x_var) x_var = self.conv(x_var) out, gate = x_var.split(int(x_var.size(1) / 2), 1) out = out * F.sigmoid(gate) return out class StackedCNN(nn.Module): def __init__(self, num_layers, input_size, cnn_kernel_width=3, dropout=0.2): super(StackedCNN, self).__init__() self.dropout = dropout self.num_layers = num_layers self.layers = nn.ModuleList() for i in range(num_layers): self.layers.append( GatedConv(input_size, cnn_kernel_width, dropout)) def forward(self, x, hidden=None): for conv in self.layers: x = x + conv(x) x *= SCALE_WEIGHT return x class CNNEncoder(EncoderBase): """ Encoder built on CNN. """ def __init__(self, num_layers, hidden_size, cnn_kernel_width, dropout, embeddings): super(CNNEncoder, self).__init__() self.embeddings = embeddings input_size = embeddings.embedding_size self.linear = nn.Linear(input_size, hidden_size) self.cnn = StackedCNN(num_layers, hidden_size, cnn_kernel_width, dropout) def forward(self, input, lengths=None, hidden=None): """ See EncoderBase.forward() for description of args and returns.""" self._check_args(input, lengths, hidden) emb = self.embeddings(input) s_len, batch, emb_dim = emb.size() emb = emb.transpose(0, 1).contiguous() emb_reshape = emb.view(emb.size(0) * emb.size(1), -1) emb_remap = self.linear(emb_reshape) emb_remap = emb_remap.view(emb.size(0), emb.size(1), -1) emb_remap = shape_transform(emb_remap) out = self.cnn(emb_remap) return emb_remap.squeeze(3).transpose(0, 1).contiguous(),\ out.squeeze(3).transpose(0, 1).contiguous() class CNNDecoder(nn.Module): """ Decoder built on CNN, which consists of resduial convolutional layers, with ConvMultiStepAttention. """ def __init__(self, num_layers, hidden_size, attn_type, copy_attn, cnn_kernel_width, dropout, embeddings): super(CNNDecoder, self).__init__() # Basic attributes. self.decoder_type = 'cnn' self.num_layers = num_layers self.hidden_size = hidden_size self.cnn_kernel_width = cnn_kernel_width self.embeddings = embeddings self.dropout = dropout # Build the CNN. input_size = self.embeddings.embedding_size self.linear = nn.Linear(input_size, self.hidden_size) self.conv_layers = nn.ModuleList() for i in range(self.num_layers): self.conv_layers.append( GatedConv(self.hidden_size, self.cnn_kernel_width, self.dropout, True)) self.attn_layers = nn.ModuleList() for i in range(self.num_layers): self.attn_layers.append( onmt.modules.ConvMultiStepAttention(self.hidden_size)) # CNNDecoder has its own attention mechanism. # Set up a separated copy attention layer, if needed. self._copy = False if copy_attn: self.copy_attn = onmt.modules.GlobalAttention( hidden_size, attn_type=attn_type) self._copy = True def forward(self, input, context, state): """ Forward through the CNNDecoder. Args: input (LongTensor): a sequence of input tokens tensors of size (len x batch x nfeats). context (FloatTensor): output(tensor sequence) from the encoder CNN of size (src_len x batch x hidden_size). state (FloatTensor): hidden state from the encoder CNN for initializing the decoder. Returns: outputs (FloatTensor): a Tensor sequence of output from the decoder of shape (len x batch x hidden_size). state (FloatTensor): final hidden state from the decoder. attns (dict of (str, FloatTensor)): a dictionary of different type of attention Tensor from the decoder of shape (src_len x batch). """ # CHECKS assert isinstance(state, CNNDecoderState) input_len, input_batch, _ = input.size() contxt_len, contxt_batch, _ = context.size() aeq(input_batch, contxt_batch) # END CHECKS if state.previous_input is not None: input = torch.cat([state.previous_input, input], 0) # Initialize return variables. outputs = [] attns = {"std": []} assert not self._copy, "Copy mechanism not yet tested in conv2conv" if self._copy: attns["copy"] = [] emb = self.embeddings(input) assert emb.dim() == 3 # len x batch x embedding_dim tgt_emb = emb.transpose(0, 1).contiguous() # The output of CNNEncoder. src_context_t = context.transpose(0, 1).contiguous() # The combination of output of CNNEncoder and source embeddings. src_context_c = state.init_src.transpose(0, 1).contiguous() # Run the forward pass of the CNNDecoder. emb_reshape = tgt_emb.contiguous().view( tgt_emb.size(0) * tgt_emb.size(1), -1) linear_out = self.linear(emb_reshape) x = linear_out.view(tgt_emb.size(0), tgt_emb.size(1), -1) x = shape_transform(x) pad = Variable(torch.zeros(x.size(0), x.size(1), self.cnn_kernel_width - 1, 1)) pad = pad.type_as(x) base_target_emb = x for conv, attention in zip(self.conv_layers, self.attn_layers): new_target_input = torch.cat([pad, x], 2) out = conv(new_target_input) c, attn = attention(base_target_emb, out, src_context_t, src_context_c) x = (x + (c + out) * SCALE_WEIGHT) * SCALE_WEIGHT output = x.squeeze(3).transpose(1, 2) # Process the result and update the attentions. outputs = output.transpose(0, 1).contiguous() if state.previous_input is not None: outputs = outputs[state.previous_input.size(0):] attn = attn[:, state.previous_input.size(0):].squeeze() attn = torch.stack([attn]) attns["std"] = attn if self._copy: attns["copy"] = attn # Update the state. state.update_state(input) return outputs, state, attns def init_decoder_state(self, src, context, enc_hidden): return CNNDecoderState(context, enc_hidden) class CNNDecoderState(DecoderState): def __init__(self, context, enc_hidden): self.init_src = (context + enc_hidden) * SCALE_WEIGHT self.previous_input = None @property def _all(self): """ Contains attributes that need to be updated in self.beam_update(). """ return (self.previous_input,) def update_state(self, input): """ Called for every decoder forward pass. """ self.previous_input = input def repeat_beam_size_times(self, beam_size): """ Repeat beam_size times along batch dimension. """ self.init_src = Variable( self.init_src.data.repeat(1, beam_size, 1), volatile=True)

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/OpenNMT/onmt/modules/GlobalAttention.py

import torch import torch.nn as nn from onmt.modules.UtilClass import BottleLinear from onmt.Utils import aeq class GlobalAttention(nn.Module): """ Luong Attention. Global attention takes a matrix and a query vector. It then computes a parameterized convex combination of the matrix based on the input query. H_1 H_2 H_3 ... H_n q q q q | | | | \ | | / ..... \ | / a Constructs a unit mapping. $$(H_1 + H_n, q) => (a)$$ Where H is of `batch x n x dim` and q is of `batch x dim`. Luong Attention (dot, general): The full function is $$\tanh(W_2 [(softmax((W_1 q + b_1) H) H), q] + b_2)$$. * dot: $$score(h_t,{\overline{h}}_s) = h_t^T{\overline{h}}_s$$ * general: $$score(h_t,{\overline{h}}_s) = h_t^T W_a {\overline{h}}_s$$ Bahdanau Attention (mlp): $$c = \sum_{j=1}^{SeqLength}\a_jh_j$$. The Alignment-function $$a$$ computes an alignment as: $$a_j = softmax(v_a^T \tanh(W_a q + U_a h_j) )$$. """ def __init__(self, dim, coverage=False, attn_type="dot"): super(GlobalAttention, self).__init__() self.dim = dim self.attn_type = attn_type assert (self.attn_type in ["dot", "general", "mlp"]), ( "Please select a valid attention type.") if self.attn_type == "general": self.linear_in = nn.Linear(dim, dim, bias=False) elif self.attn_type == "mlp": self.linear_context = BottleLinear(dim, dim, bias=False) self.linear_query = nn.Linear(dim, dim, bias=True) self.v = BottleLinear(dim, 1, bias=False) # mlp wants it with bias out_bias = self.attn_type == "mlp" self.linear_out = nn.Linear(dim*2, dim, bias=out_bias) self.sm = nn.Softmax() self.tanh = nn.Tanh() self.mask = None if coverage: self.linear_cover = nn.Linear(1, dim, bias=False) def applyMask(self, mask): self.mask = mask def score(self, h_t, h_s): """ h_t (FloatTensor): batch x tgt_len x dim h_s (FloatTensor): batch x src_len x dim returns scores (FloatTensor): batch x tgt_len x src_len: raw attention scores for each src index """ # Check input sizes src_batch, src_len, src_dim = h_s.size() tgt_batch, tgt_len, tgt_dim = h_t.size() aeq(src_batch, tgt_batch) aeq(src_dim, tgt_dim) aeq(self.dim, src_dim) if self.attn_type in ["general", "dot"]: if self.attn_type == "general": h_t_ = h_t.view(tgt_batch*tgt_len, tgt_dim) h_t_ = self.linear_in(h_t_) h_t = h_t_.view(tgt_batch, tgt_len, tgt_dim) h_s_ = h_s.transpose(1, 2) # (batch, t_len, d) x (batch, d, s_len) --> (batch, t_len, s_len) return torch.bmm(h_t, h_s_) else: dim = self.dim wq = self.linear_query(h_t.view(-1, dim)) wq = wq.view(tgt_batch, tgt_len, 1, dim) wq = wq.expand(tgt_batch, tgt_len, src_len, dim) uh = self.linear_context(h_s.contiguous().view(-1, dim)) uh = uh.view(src_batch, 1, src_len, dim) uh = uh.expand(src_batch, tgt_len, src_len, dim) # (batch, t_len, s_len, d) wquh = self.tanh(wq + uh) return self.v(wquh.view(-1, dim)).view(tgt_batch, tgt_len, src_len) def forward(self, input, context, coverage=None): """ input (FloatTensor): batch x tgt_len x dim: decoder's rnn's output. context (FloatTensor): batch x src_len x dim: src hidden states coverage (FloatTensor): None (not supported yet) """ # one step input if input.dim() == 2: one_step = True input = input.unsqueeze(1) else: one_step = False batch, sourceL, dim = context.size() batch_, targetL, dim_ = input.size() aeq(batch, batch_) aeq(dim, dim_) aeq(self.dim, dim) if coverage is not None: batch_, sourceL_ = coverage.size() aeq(batch, batch_) aeq(sourceL, sourceL_) if self.mask is not None: beam_, batch_, sourceL_ = self.mask.size() aeq(batch, batch_*beam_) aeq(sourceL, sourceL_) if coverage is not None: cover = coverage.view(-1).unsqueeze(1) context += self.linear_cover(cover).view_as(context) context = self.tanh(context) # compute attention scores, as in Luong et al. align = self.score(input, context) if self.mask is not None: mask_ = self.mask.view(batch, 1, sourceL) # make it broardcastable align.data.masked_fill_(mask_, -float('inf')) # Softmax to normalize attention weights align_vectors = self.sm(align.view(batch*targetL, sourceL)) align_vectors = align_vectors.view(batch, targetL, sourceL) # each context vector c_t is the weighted average # over all the source hidden states c = torch.bmm(align_vectors, context) # concatenate concat_c = torch.cat([c, input], 2).view(batch*targetL, dim*2) attn_h = self.linear_out(concat_c).view(batch, targetL, dim) if self.attn_type in ["general", "dot"]: attn_h = self.tanh(attn_h) if one_step: attn_h = attn_h.squeeze(1) align_vectors = align_vectors.squeeze(1) # Check output sizes batch_, dim_ = attn_h.size() aeq(batch, batch_) aeq(dim, dim_) batch_, sourceL_ = align_vectors.size() aeq(batch, batch_) aeq(sourceL, sourceL_) else: attn_h = attn_h.transpose(0, 1).contiguous() align_vectors = align_vectors.transpose(0, 1).contiguous() # Check output sizes targetL_, batch_, dim_ = attn_h.size() aeq(targetL, targetL_) aeq(batch, batch_) aeq(dim, dim_) targetL_, batch_, sourceL_ = align_vectors.size() aeq(targetL, targetL_) aeq(batch, batch_) aeq(sourceL, sourceL_) return attn_h, align_vectors

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/OpenNMT/onmt/modules/SRU.py

""" Implementation of "Training RNNs as Fast as CNNs". TODO: turn to pytorch's implementation when it is available. This implementation is adpoted from the author of the paper: https://github.com/taolei87/sru/blob/master/cuda_functional.py. """ import subprocess import platform import os import re import argparse import torch import torch.nn as nn from torch.autograd import Function, Variable from collections import namedtuple # For command-line option parsing class CheckSRU(argparse.Action): def __init__(self, option_strings, dest, **kwargs): super(CheckSRU, self).__init__(option_strings, dest, **kwargs) def __call__(self, parser, namespace, values, option_string=None): if values == 'SRU': check_sru_requirement(abort=True) # Check pass, set the args. setattr(namespace, self.dest, values) # This SRU version implements its own cuda-level optimization, # so it requires that: # 1. `cupy` and `pynvrtc` python package installed. # 2. pytorch is built with cuda support. # 3. library path set: export LD_LIBRARY_PATH=<cuda lib path>. def check_sru_requirement(abort=False): """ Return True if check pass; if check fails and abort is True, raise an Exception, othereise return False. """ # Check 1. try: if platform.system() == 'Windows': subprocess.check_output('pip freeze | findstr cupy', shell=True) subprocess.check_output('pip freeze | findstr pynvrtc', shell=True) else: # Unix-like systems subprocess.check_output('pip freeze | grep -w cupy', shell=True) subprocess.check_output('pip freeze | grep -w pynvrtc', shell=True) except subprocess.CalledProcessError: if not abort: return False raise AssertionError("Using SRU requires 'cupy' and 'pynvrtc' " "python packages installed.") # Check 2. if torch.cuda.is_available() is False: if not abort: return False raise AssertionError("Using SRU requires pytorch built with cuda.") # Check 3. pattern = re.compile(".*cuda/lib.*") ld_path = os.getenv('LD_LIBRARY_PATH', "") if re.match(pattern, ld_path) is None: if not abort: return False raise AssertionError("Using SRU requires setting cuda lib path, e.g. " "export LD_LIBRARY_PATH=/usr/local/cuda/lib64.") return True SRU_CODE = """ extern "C" { __forceinline__ __device__ float sigmoidf(float x) { return 1.f / (1.f + expf(-x)); } __forceinline__ __device__ float reluf(float x) { return (x > 0.f) ? x : 0.f; } __global__ void sru_fwd(const float * __restrict__ u, const float * __restrict__ x, const float * __restrict__ bias, const float * __restrict__ init, const float * __restrict__ mask_h, const int len, const int batch, const int d, const int k, float * __restrict__ h, float * __restrict__ c, const int activation_type) { assert ((k == 3) || (x == NULL)); int ncols = batch*d; int col = blockIdx.x * blockDim.x + threadIdx.x; if (col >= ncols) return; int ncols_u = ncols*k; int ncols_x = (k == 3) ? ncols : ncols_u; const float bias1 = *(bias + (col%d)); const float bias2 = *(bias + (col%d) + d); const float mask = (mask_h == NULL) ? 1.0 : (*(mask_h + col)); float cur = *(init + col); const float *up = u + (col*k); const float *xp = (k == 3) ? (x + col) : (up + 3); float *cp = c + col; float *hp = h + col; for (int row = 0; row < len; ++row) { float g1 = sigmoidf((*(up+1))+bias1); float g2 = sigmoidf((*(up+2))+bias2); cur = (cur-(*up))*g1 + (*up); *cp = cur; float val = (activation_type == 1) ? tanh(cur) : ( (activation_type == 2) ? reluf(cur) : cur ); *hp = (val*mask-(*xp))*g2 + (*xp); up += ncols_u; xp += ncols_x; cp += ncols; hp += ncols; } } __global__ void sru_bwd(const float * __restrict__ u, const float * __restrict__ x, const float * __restrict__ bias, const float * __restrict__ init, const float * __restrict__ mask_h, const float * __restrict__ c, const float * __restrict__ grad_h, const float * __restrict__ grad_last, const int len, const int batch, const int d, const int k, float * __restrict__ grad_u, float * __restrict__ grad_x, float * __restrict__ grad_bias, float * __restrict__ grad_init, int activation_type) { assert((k == 3) || (x == NULL)); assert((k == 3) || (grad_x == NULL)); int ncols = batch*d; int col = blockIdx.x * blockDim.x + threadIdx.x; if (col >= ncols) return; int ncols_u = ncols*k; int ncols_x = (k == 3) ? ncols : ncols_u; const float bias1 = *(bias + (col%d)); const float bias2 = *(bias + (col%d) + d); const float mask = (mask_h == NULL) ? 1.0 : (*(mask_h + col)); float gbias1 = 0; float gbias2 = 0; float cur = *(grad_last + col); const float *up = u + (col*k) + (len-1)*ncols_u; const float *xp = (k == 3) ? (x + col + (len-1)*ncols) : (up + 3); const float *cp = c + col + (len-1)*ncols; const float *ghp = grad_h + col + (len-1)*ncols; float *gup = grad_u + (col*k) + (len-1)*ncols_u; float *gxp = (k == 3) ? (grad_x + col + (len-1)*ncols) : (gup + 3); for (int row = len-1; row >= 0; --row) { const float g1 = sigmoidf((*(up+1))+bias1); const float g2 = sigmoidf((*(up+2))+bias2); const float c_val = (activation_type == 1) ? tanh(*cp) : ( (activation_type == 2) ? reluf(*cp) : (*cp) ); const float x_val = *xp; const float u_val = *up; const float prev_c_val = (row>0) ? (*(cp-ncols)) : (*(init+col)); const float gh_val = *ghp; // h = c*g2 + x*(1-g2) = (c-x)*g2 + x // c = c'*g1 + g0*(1-g1) = (c'-g0)*g1 + g0 // grad wrt x *gxp = gh_val*(1-g2); // grad wrt g2, u2 and bias2 float gg2 = gh_val*(c_val*mask-x_val)*(g2*(1-g2)); *(gup+2) = gg2; gbias2 += gg2; // grad wrt c const float tmp = (activation_type == 1) ? (g2*(1-c_val*c_val)) : ( ((activation_type == 0) || (c_val > 0)) ? g2 : 0.f ); const float gc = gh_val*mask*tmp + cur; // grad wrt u0 *gup = gc*(1-g1); // grad wrt g1, u1, and bias1 float gg1 = gc*(prev_c_val-u_val)*(g1*(1-g1)); *(gup+1) = gg1; gbias1 += gg1; // grad wrt c' cur = gc*g1; up -= ncols_u; xp -= ncols_x; cp -= ncols; gup -= ncols_u; gxp -= ncols_x; ghp -= ncols; } *(grad_bias + col) = gbias1; *(grad_bias + col + ncols) = gbias2; *(grad_init +col) = cur; } __global__ void sru_bi_fwd(const float * __restrict__ u, const float * __restrict__ x, const float * __restrict__ bias, const float * __restrict__ init, const float * __restrict__ mask_h, const int len, const int batch, const int d, const int k, float * __restrict__ h, float * __restrict__ c, const int activation_type) { assert ((k == 3) || (x == NULL)); assert ((k == 3) || (k == 4)); int ncols = batch*d*2; int col = blockIdx.x * blockDim.x + threadIdx.x; if (col >= ncols) return; int ncols_u = ncols*k; int ncols_x = (k == 3) ? ncols : ncols_u; const float mask = (mask_h == NULL) ? 1.0 : (*(mask_h + col)); float cur = *(init + col); const int d2 = d*2; const bool flip = (col%d2) >= d; const float bias1 = *(bias + (col%d2)); const float bias2 = *(bias + (col%d2) + d2); const float *up = u + (col*k); const float *xp = (k == 3) ? (x + col) : (up + 3); float *cp = c + col; float *hp = h + col; if (flip) { up += (len-1)*ncols_u; xp += (len-1)*ncols_x; cp += (len-1)*ncols; hp += (len-1)*ncols; } int ncols_u_ = flip ? -ncols_u : ncols_u; int ncols_x_ = flip ? -ncols_x : ncols_x; int ncols_ = flip ? -ncols : ncols; for (int cnt = 0; cnt < len; ++cnt) { float g1 = sigmoidf((*(up+1))+bias1); float g2 = sigmoidf((*(up+2))+bias2); cur = (cur-(*up))*g1 + (*up); *cp = cur; float val = (activation_type == 1) ? tanh(cur) : ( (activation_type == 2) ? reluf(cur) : cur ); *hp = (val*mask-(*xp))*g2 + (*xp); up += ncols_u_; xp += ncols_x_; cp += ncols_; hp += ncols_; } } __global__ void sru_bi_bwd(const float * __restrict__ u, const float * __restrict__ x, const float * __restrict__ bias, const float * __restrict__ init, const float * __restrict__ mask_h, const float * __restrict__ c, const float * __restrict__ grad_h, const float * __restrict__ grad_last, const int len, const int batch, const int d, const int k, float * __restrict__ grad_u, float * __restrict__ grad_x, float * __restrict__ grad_bias, float * __restrict__ grad_init, int activation_type) { assert((k == 3) || (x == NULL)); assert((k == 3) || (grad_x == NULL)); assert((k == 3) || (k == 4)); int ncols = batch*d*2; int col = blockIdx.x * blockDim.x + threadIdx.x; if (col >= ncols) return; int ncols_u = ncols*k; int ncols_x = (k == 3) ? ncols : ncols_u; const float mask = (mask_h == NULL) ? 1.0 : (*(mask_h + col)); float gbias1 = 0; float gbias2 = 0; float cur = *(grad_last + col); const int d2 = d*2; const bool flip = ((col%d2) >= d); const float bias1 = *(bias + (col%d2)); const float bias2 = *(bias + (col%d2) + d2); const float *up = u + (col*k); const float *xp = (k == 3) ? (x + col) : (up + 3); const float *cp = c + col; const float *ghp = grad_h + col; float *gup = grad_u + (col*k); float *gxp = (k == 3) ? (grad_x + col) : (gup + 3); if (!flip) { up += (len-1)*ncols_u; xp += (len-1)*ncols_x; cp += (len-1)*ncols; ghp += (len-1)*ncols; gup += (len-1)*ncols_u; gxp += (len-1)*ncols_x; } int ncols_u_ = flip ? -ncols_u : ncols_u; int ncols_x_ = flip ? -ncols_x : ncols_x; int ncols_ = flip ? -ncols : ncols; for (int cnt = 0; cnt < len; ++cnt) { const float g1 = sigmoidf((*(up+1))+bias1); const float g2 = sigmoidf((*(up+2))+bias2); const float c_val = (activation_type == 1) ? tanh(*cp) : ( (activation_type == 2) ? reluf(*cp) : (*cp) ); const float x_val = *xp; const float u_val = *up; const float prev_c_val = (cnt<len-1)?(*(cp-ncols_)):(*(init+col)); const float gh_val = *ghp; // h = c*g2 + x*(1-g2) = (c-x)*g2 + x // c = c'*g1 + g0*(1-g1) = (c'-g0)*g1 + g0 // grad wrt x *gxp = gh_val*(1-g2); // grad wrt g2, u2 and bias2 float gg2 = gh_val*(c_val*mask-x_val)*(g2*(1-g2)); *(gup+2) = gg2; gbias2 += gg2; // grad wrt c const float tmp = (activation_type == 1) ? (g2*(1-c_val*c_val)) : ( ((activation_type == 0) || (c_val > 0)) ? g2 : 0.f ); const float gc = gh_val*mask*tmp + cur; // grad wrt u0 *gup = gc*(1-g1); // grad wrt g1, u1, and bias1 float gg1 = gc*(prev_c_val-u_val)*(g1*(1-g1)); *(gup+1) = gg1; gbias1 += gg1; // grad wrt c' cur = gc*g1; up -= ncols_u_; xp -= ncols_x_; cp -= ncols_; gup -= ncols_u_; gxp -= ncols_x_; ghp -= ncols_; } *(grad_bias + col) = gbias1; *(grad_bias + col + ncols) = gbias2; *(grad_init +col) = cur; } } """ if check_sru_requirement(): from cupy.cuda import function from pynvrtc.compiler import Program # This cuda() is important, it sets up device to use. tmp_ = torch.rand(1, 1).cuda() sru_prog = Program(SRU_CODE.encode('utf-8'), 'sru_prog.cu'.encode('utf-8')) sru_ptx = sru_prog.compile() sru_mod = function.Module() sru_mod.load(bytes(sru_ptx.encode())) SRU_FWD_FUNC = sru_mod.get_function('sru_fwd') SRU_BWD_FUNC = sru_mod.get_function('sru_bwd') SRU_BiFWD_FUNC = sru_mod.get_function('sru_bi_fwd') SRU_BiBWD_FUNC = sru_mod.get_function('sru_bi_bwd') stream = namedtuple('Stream', ['ptr']) SRU_STREAM = stream(ptr=torch.cuda.current_stream().cuda_stream) class SRU_Compute(Function): def __init__(self, activation_type, d_out, bidirectional=False): super(SRU_Compute, self).__init__() self.activation_type = activation_type self.d_out = d_out self.bidirectional = bidirectional def forward(self, u, x, bias, init=None, mask_h=None): bidir = 2 if self.bidirectional else 1 length = x.size(0) if x.dim() == 3 else 1 batch = x.size(-2) d = self.d_out k = u.size(-1) // d k_ = k // 2 if self.bidirectional else k ncols = batch * d * bidir thread_per_block = min(512, ncols) num_block = (ncols-1) // thread_per_block+1 init_ = x.new(ncols).zero_() if init is None else init size = (length, batch, d*bidir) if x.dim() == 3 else (batch, d*bidir) c = x.new(*size) h = x.new(*size) FUNC = SRU_FWD_FUNC if not self.bidirectional else SRU_BiFWD_FUNC FUNC(args=[ u.contiguous().data_ptr(), x.contiguous().data_ptr() if k_ == 3 else 0, bias.data_ptr(), init_.contiguous().data_ptr(), mask_h.data_ptr() if mask_h is not None else 0, length, batch, d, k_, h.data_ptr(), c.data_ptr(), self.activation_type], block=(thread_per_block, 1, 1), grid=(num_block, 1, 1), stream=SRU_STREAM ) self.save_for_backward(u, x, bias, init, mask_h) self.intermediate = c if x.dim() == 2: last_hidden = c elif self.bidirectional: # -> directions x batch x dim last_hidden = torch.stack((c[-1, :, :d], c[0, :, d:])) else: last_hidden = c[-1] return h, last_hidden def backward(self, grad_h, grad_last): if self.bidirectional: grad_last = torch.cat((grad_last[0], grad_last[1]), 1) bidir = 2 if self.bidirectional else 1 u, x, bias, init, mask_h = self.saved_tensors c = self.intermediate length = x.size(0) if x.dim() == 3 else 1 batch = x.size(-2) d = self.d_out k = u.size(-1) // d k_ = k//2 if self.bidirectional else k ncols = batch*d*bidir thread_per_block = min(512, ncols) num_block = (ncols-1) // thread_per_block+1 init_ = x.new(ncols).zero_() if init is None else init grad_u = u.new(*u.size()) grad_bias = x.new(2, batch, d*bidir) grad_init = x.new(batch, d*bidir) # For DEBUG # size = (length, batch, x.size(-1)) \ # if x.dim() == 3 else (batch, x.size(-1)) # grad_x = x.new(*x.size()) if k_ == 3 else x.new(*size).zero_() # Normal use grad_x = x.new(*x.size()) if k_ == 3 else None FUNC = SRU_BWD_FUNC if not self.bidirectional else SRU_BiBWD_FUNC FUNC(args=[ u.contiguous().data_ptr(), x.contiguous().data_ptr() if k_ == 3 else 0, bias.data_ptr(), init_.contiguous().data_ptr(), mask_h.data_ptr() if mask_h is not None else 0, c.data_ptr(), grad_h.contiguous().data_ptr(), grad_last.contiguous().data_ptr(), length, batch, d, k_, grad_u.data_ptr(), grad_x.data_ptr() if k_ == 3 else 0, grad_bias.data_ptr(), grad_init.data_ptr(), self.activation_type], block=(thread_per_block, 1, 1), grid=(num_block, 1, 1), stream=SRU_STREAM ) return grad_u, grad_x, grad_bias.sum(1).view(-1), grad_init, None class SRUCell(nn.Module): def __init__(self, n_in, n_out, dropout=0, rnn_dropout=0, bidirectional=False, use_tanh=1, use_relu=0): super(SRUCell, self).__init__() self.n_in = n_in self.n_out = n_out self.rnn_dropout = rnn_dropout self.dropout = dropout self.bidirectional = bidirectional self.activation_type = 2 if use_relu else (1 if use_tanh else 0) out_size = n_out*2 if bidirectional else n_out k = 4 if n_in != out_size else 3 self.size_per_dir = n_out*k self.weight = nn.Parameter(torch.Tensor( n_in, self.size_per_dir*2 if bidirectional else self.size_per_dir )) self.bias = nn.Parameter(torch.Tensor( n_out*4 if bidirectional else n_out*2 )) self.init_weight() def init_weight(self): val_range = (3.0/self.n_in)**0.5 self.weight.data.uniform_(-val_range, val_range) self.bias.data.zero_() def set_bias(self, bias_val=0): n_out = self.n_out if self.bidirectional: self.bias.data[n_out*2:].zero_().add_(bias_val) else: self.bias.data[n_out:].zero_().add_(bias_val) def forward(self, input, c0=None): assert input.dim() == 2 or input.dim() == 3 n_in, n_out = self.n_in, self.n_out batch = input.size(-2) if c0 is None: c0 = Variable(input.data.new( batch, n_out if not self.bidirectional else n_out*2 ).zero_()) if self.training and (self.rnn_dropout > 0): mask = self.get_dropout_mask_((batch, n_in), self.rnn_dropout) x = input * mask.expand_as(input) else: x = input x_2d = x if x.dim() == 2 else x.contiguous().view(-1, n_in) u = x_2d.mm(self.weight) if self.training and (self.dropout > 0): bidir = 2 if self.bidirectional else 1 mask_h = self.get_dropout_mask_((batch, n_out*bidir), self.dropout) h, c = SRU_Compute(self.activation_type, n_out, self.bidirectional)( u, input, self.bias, c0, mask_h ) else: h, c = SRU_Compute(self.activation_type, n_out, self.bidirectional)( u, input, self.bias, c0 ) return h, c def get_dropout_mask_(self, size, p): w = self.weight.data return Variable(w.new(*size).bernoulli_(1-p).div_(1-p)) class SRU(nn.Module): def __init__(self, input_size, hidden_size, num_layers=2, dropout=0, rnn_dropout=0, bidirectional=False, use_tanh=1, use_relu=0): # An entry check here, will catch on train side and translate side # if requirements are not satisfied. check_sru_requirement(abort=True) super(SRU, self).__init__() self.n_in = input_size self.n_out = hidden_size self.depth = num_layers self.dropout = dropout self.rnn_dropout = rnn_dropout self.rnn_lst = nn.ModuleList() self.bidirectional = bidirectional self.out_size = hidden_size*2 if bidirectional else hidden_size for i in range(num_layers): sru_cell = SRUCell( n_in=self.n_in if i == 0 else self.out_size, n_out=self.n_out, dropout=dropout if i+1 != num_layers else 0, rnn_dropout=rnn_dropout, bidirectional=bidirectional, use_tanh=use_tanh, use_relu=use_relu, ) self.rnn_lst.append(sru_cell) def set_bias(self, bias_val=0): for l in self.rnn_lst: l.set_bias(bias_val) def forward(self, input, c0=None, return_hidden=True): assert input.dim() == 3 # (len, batch, n_in) dir_ = 2 if self.bidirectional else 1 if c0 is None: zeros = Variable(input.data.new( input.size(1), self.n_out*dir_ ).zero_()) c0 = [zeros for i in range(self.depth)] else: if isinstance(c0, tuple): # RNNDecoderState wraps hidden as a tuple. c0 = c0[0] assert c0.dim() == 3 # (depth, batch, dir_*n_out) c0 = [h.squeeze(0) for h in c0.chunk(self.depth, 0)] prevx = input lstc = [] for i, rnn in enumerate(self.rnn_lst): h, c = rnn(prevx, c0[i]) prevx = h lstc.append(c) if self.bidirectional: # fh -> (layers*directions) x batch x dim fh = torch.cat(lstc) else: fh = torch.stack(lstc) if return_hidden: return prevx, fh else: return prevx

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/OpenNMT/onmt/modules/WeightNorm.py

""" Implementation of "Weight Normalization: A Simple Reparameterization to Accelerate Training of Deep Neural Networks" As a reparameterization method, weight normalization is same as BatchNormalization, but it doesn't depend on minibatch. """ import torch import torch.nn as nn import torch.nn.functional as F from torch.nn import Parameter from torch.autograd import Variable def get_var_maybe_avg(namespace, var_name, training, polyak_decay): # utility for retrieving polyak averaged params # Update average v = getattr(namespace, var_name) v_avg = getattr(namespace, var_name + '_avg') v_avg -= (1 - polyak_decay) * (v_avg - v.data) if training: return v else: return Variable(v_avg) def get_vars_maybe_avg(namespace, var_names, training, polyak_decay): # utility for retrieving polyak averaged params vars = [] for vn in var_names: vars.append(get_var_maybe_avg( namespace, vn, training, polyak_decay)) return vars class WeightNormLinear(nn.Linear): def __init__(self, in_features, out_features, init_scale=1., polyak_decay=0.9995): super(WeightNormLinear, self).__init__( in_features, out_features, bias=True) self.V = self.weight self.g = Parameter(torch.Tensor(out_features)) self.b = self.bias self.register_buffer( 'V_avg', torch.zeros(out_features, in_features)) self.register_buffer('g_avg', torch.zeros(out_features)) self.register_buffer('b_avg', torch.zeros(out_features)) self.init_scale = init_scale self.polyak_decay = polyak_decay self.reset_parameters() def reset_parameters(self): return def forward(self, x, init=False): if init is True: # out_features * in_features self.V.data.copy_(torch.randn(self.V.data.size()).type_as( self.V.data) * 0.05) # norm is out_features * 1 V_norm = self.V.data / \ self.V.data.norm(2, 1).expand_as(self.V.data) # batch_size * out_features x_init = F.linear(x, Variable(V_norm)).data # out_features m_init, v_init = x_init.mean(0).squeeze( 0), x_init.var(0).squeeze(0) # out_features scale_init = self.init_scale / \ torch.sqrt(v_init + 1e-10) self.g.data.copy_(scale_init) self.b.data.copy_(-m_init * scale_init) x_init = scale_init.view(1, -1).expand_as(x_init) \ * (x_init - m_init.view(1, -1).expand_as(x_init)) self.V_avg.copy_(self.V.data) self.g_avg.copy_(self.g.data) self.b_avg.copy_(self.b.data) return Variable(x_init) else: V, g, b = get_vars_maybe_avg(self, ['V', 'g', 'b'], self.training, polyak_decay=self.polyak_decay) # batch_size * out_features x = F.linear(x, V) scalar = g / torch.norm(V, 2, 1).squeeze(1) x = scalar.view(1, -1).expand_as(x) * x + \ b.view(1, -1).expand_as(x) return x class WeightNormConv2d(nn.Conv2d): def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, init_scale=1., polyak_decay=0.9995): super(WeightNormConv2d, self).__init__(in_channels, out_channels, kernel_size, stride, padding, dilation, groups) self.V = self.weight self.g = Parameter(torch.Tensor(out_channels)) self.b = self.bias self.register_buffer('V_avg', torch.zeros(self.V.size())) self.register_buffer('g_avg', torch.zeros(out_channels)) self.register_buffer('b_avg', torch.zeros(out_channels)) self.init_scale = init_scale self.polyak_decay = polyak_decay self.reset_parameters() def reset_parameters(self): return def forward(self, x, init=False): if init is True: # out_channels, in_channels // groups, * kernel_size self.V.data.copy_(torch.randn(self.V.data.size() ).type_as(self.V.data) * 0.05) V_norm = self.V.data / self.V.data.view(self.out_channels, -1)\ .norm(2, 1).view(self.out_channels, *( [1] * (len(self.kernel_size) + 1))).expand_as(self.V.data) x_init = F.conv2d(x, Variable(V_norm), None, self.stride, self.padding, self.dilation, self.groups).data t_x_init = x_init.transpose(0, 1).contiguous().view( self.out_channels, -1) m_init, v_init = t_x_init.mean(1).squeeze( 1), t_x_init.var(1).squeeze(1) # out_features scale_init = self.init_scale / \ torch.sqrt(v_init + 1e-10) self.g.data.copy_(scale_init) self.b.data.copy_(-m_init * scale_init) scale_init_shape = scale_init.view( 1, self.out_channels, *([1] * (len(x_init.size()) - 2))) m_init_shape = m_init.view( 1, self.out_channels, *([1] * (len(x_init.size()) - 2))) x_init = scale_init_shape.expand_as( x_init) * (x_init - m_init_shape.expand_as(x_init)) self.V_avg.copy_(self.V.data) self.g_avg.copy_(self.g.data) self.b_avg.copy_(self.b.data) return Variable(x_init) else: V, g, b = get_vars_maybe_avg( self, ['V', 'g', 'b'], self.training, polyak_decay=self.polyak_decay) scalar = torch.norm(V.view(self.out_channels, -1), 2, 1) if len(scalar.size()) == 2: scalar = g / scalar.squeeze(1) else: scalar = g / scalar W = scalar.view(self.out_channels, * ([1] * (len(V.size()) - 1))).expand_as(V) * V x = F.conv2d(x, W, b, self.stride, self.padding, self.dilation, self.groups) return x class WeightNormConvTranspose2d(nn.ConvTranspose2d): def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0, output_padding=0, groups=1, init_scale=1., polyak_decay=0.9995): super(WeightNormConvTranspose2d, self).__init__( in_channels, out_channels, kernel_size, stride, padding, output_padding, groups) # in_channels, out_channels, *kernel_size self.V = self.weight self.g = Parameter(torch.Tensor(out_channels)) self.b = self.bias self.register_buffer('V_avg', torch.zeros(self.V.size())) self.register_buffer('g_avg', torch.zeros(out_channels)) self.register_buffer('b_avg', torch.zeros(out_channels)) self.init_scale = init_scale self.polyak_decay = polyak_decay self.reset_parameters() def reset_parameters(self): return def forward(self, x, init=False): if init is True: # in_channels, out_channels, *kernel_size self.V.data.copy_(torch.randn(self.V.data.size()).type_as( self.V.data) * 0.05) V_norm = self.V.data / self.V.data.transpose(0, 1).contiguous() \ .view(self.out_channels, -1).norm(2, 1).view( self.in_channels, self.out_channels, *([1] * len(self.kernel_size))).expand_as(self.V.data) x_init = F.conv_transpose2d( x, Variable(V_norm), None, self.stride, self.padding, self.output_padding, self.groups).data # self.out_channels, 1 t_x_init = x_init.tranpose(0, 1).contiguous().view( self.out_channels, -1) # out_features m_init, v_init = t_x_init.mean(1).squeeze( 1), t_x_init.var(1).squeeze(1) # out_features scale_init = self.init_scale / \ torch.sqrt(v_init + 1e-10) self.g.data.copy_(scale_init) self.b.data.copy_(-m_init * scale_init) scale_init_shape = scale_init.view( 1, self.out_channels, *([1] * (len(x_init.size()) - 2))) m_init_shape = m_init.view( 1, self.out_channels, *([1] * (len(x_init.size()) - 2))) x_init = scale_init_shape.expand_as(x_init)\ * (x_init - m_init_shape.expand_as(x_init)) self.V_avg.copy_(self.V.data) self.g_avg.copy_(self.g.data) self.b_avg.copy_(self.b.data) return Variable(x_init) else: V, g, b = get_vars_maybe_avg( self, ['V', 'g', 'b'], self.training, polyak_decay=self.polyak_decay) scalar = g / \ torch.norm(V.transpose(0, 1).contiguous().view( self.out_channels, -1), 2, 1).squeeze(1) W = scalar.view(self.in_channels, self.out_channels, *([1] * (len(V.size()) - 2))).expand_as(V) * V x = F.conv_transpose2d(x, W, b, self.stride, self.padding, self.output_padding, self.groups) return x

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/OpenNMT/onmt/modules/ImageEncoder.py

import torch.nn as nn import torch.nn.functional as F import torch import torch.cuda from torch.autograd import Variable class ImageEncoder(nn.Module): """ Encoder recurrent neural network for Images. """ def __init__(self, num_layers, bidirectional, rnn_size, dropout): """ Args: num_layers (int): number of encoder layers. bidirectional (bool): bidirectional encoder. rnn_size (int): size of hidden states of the rnn. dropout (float): dropout probablity. """ super(ImageEncoder, self).__init__() self.num_layers = num_layers self.num_directions = 2 if bidirectional else 1 self.hidden_size = rnn_size self.layer1 = nn.Conv2d(3, 64, kernel_size=(3, 3), padding=(1, 1), stride=(1, 1)) self.layer2 = nn.Conv2d(64, 128, kernel_size=(3, 3), padding=(1, 1), stride=(1, 1)) self.layer3 = nn.Conv2d(128, 256, kernel_size=(3, 3), padding=(1, 1), stride=(1, 1)) self.layer4 = nn.Conv2d(256, 256, kernel_size=(3, 3), padding=(1, 1), stride=(1, 1)) self.layer5 = nn.Conv2d(256, 512, kernel_size=(3, 3), padding=(1, 1), stride=(1, 1)) self.layer6 = nn.Conv2d(512, 512, kernel_size=(3, 3), padding=(1, 1), stride=(1, 1)) self.batch_norm1 = nn.BatchNorm2d(256) self.batch_norm2 = nn.BatchNorm2d(512) self.batch_norm3 = nn.BatchNorm2d(512) input_size = 512 self.rnn = nn.LSTM(input_size, rnn_size, num_layers=num_layers, dropout=dropout, bidirectional=bidirectional) self.pos_lut = nn.Embedding(1000, input_size) def load_pretrained_vectors(self, opt): # Pass in needed options only when modify function definition. pass def forward(self, input, lengths=None): batchSize = input.size(0) # (batch_size, 64, imgH, imgW) # layer 1 input = F.relu(self.layer1(input[:, :, :, :]-0.5), True) # (batch_size, 64, imgH/2, imgW/2) input = F.max_pool2d(input, kernel_size=(2, 2), stride=(2, 2)) # (batch_size, 128, imgH/2, imgW/2) # layer 2 input = F.relu(self.layer2(input), True) # (batch_size, 128, imgH/2/2, imgW/2/2) input = F.max_pool2d(input, kernel_size=(2, 2), stride=(2, 2)) # (batch_size, 256, imgH/2/2, imgW/2/2) # layer 3 # batch norm 1 input = F.relu(self.batch_norm1(self.layer3(input)), True) # (batch_size, 256, imgH/2/2, imgW/2/2) # layer4 input = F.relu(self.layer4(input), True) # (batch_size, 256, imgH/2/2/2, imgW/2/2) input = F.max_pool2d(input, kernel_size=(1, 2), stride=(1, 2)) # (batch_size, 512, imgH/2/2/2, imgW/2/2) # layer 5 # batch norm 2 input = F.relu(self.batch_norm2(self.layer5(input)), True) # (batch_size, 512, imgH/2/2/2, imgW/2/2/2) input = F.max_pool2d(input, kernel_size=(2, 1), stride=(2, 1)) # (batch_size, 512, imgH/2/2/2, imgW/2/2/2) input = F.relu(self.batch_norm3(self.layer6(input)), True) # # (batch_size, 512, H, W) # # (batch_size, H, W, 512) all_outputs = [] for row in range(input.size(2)): inp = input[:, :, row, :].transpose(0, 2)\ .transpose(1, 2) pos_emb = self.pos_lut( Variable(torch.cuda.LongTensor(batchSize).fill_(row))) with_pos = torch.cat( (pos_emb.view(1, pos_emb.size(0), pos_emb.size(1)), inp), 0) outputs, hidden_t = self.rnn(with_pos) all_outputs.append(outputs) out = torch.cat(all_outputs, 0) return hidden_t, out

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/source/Generation/seq2seq_translation_tutorial.py

# -*- coding: utf-8 -*- """ Translation with a Sequence to Sequence Network and Attention ************************************************************* **Author**: `Sean Robertson <https://github.com/spro/practical-pytorch>`_ In this project we will be teaching a neural network to translate from French to English. :: [KEY: > input, = target, < output] > il est en train de peindre un tableau . = he is painting a picture . < he is painting a picture . > pourquoi ne pas essayer ce vin delicieux ? = why not try that delicious wine ? < why not try that delicious wine ? > elle n est pas poete mais romanciere . = she is not a poet but a novelist . < she not not a poet but a novelist . > vous etes trop maigre . = you re too skinny . < you re all alone . ... to varying degrees of success. This is made possible by the simple but powerful idea of the `sequence to sequence network <http://arxiv.org/abs/1409.3215>`__, in which two recurrent neural networks work together to transform one sequence to another. An encoder network condenses an input sequence into a vector, and a decoder network unfolds that vector into a new sequence. .. figure:: /_static/img/seq-seq-images/seq2seq.png :alt: To improve upon this model we'll use an `attention mechanism <https://arxiv.org/abs/1409.0473>`__, which lets the decoder learn to focus over a specific range of the input sequence. **Recommended Reading:** I assume you have at least installed PyTorch, know Python, and understand Tensors: - http://pytorch.org/ For installation instructions - :doc:`/beginner/deep_learning_60min_blitz` to get started with PyTorch in general - :doc:`/beginner/pytorch_with_examples` for a wide and deep overview - :doc:`/beginner/former_torchies_tutorial` if you are former Lua Torch user It would also be useful to know about Sequence to Sequence networks and how they work: - `Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation <http://arxiv.org/abs/1406.1078>`__ - `Sequence to Sequence Learning with Neural Networks <http://arxiv.org/abs/1409.3215>`__ - `Neural Machine Translation by Jointly Learning to Align and Translate <https://arxiv.org/abs/1409.0473>`__ - `A Neural Conversational Model <http://arxiv.org/abs/1506.05869>`__ You will also find the previous tutorials on :doc:`/intermediate/char_rnn_classification_tutorial` and :doc:`/intermediate/char_rnn_generation_tutorial` helpful as those concepts are very similar to the Encoder and Decoder models, respectively. And for more, read the papers that introduced these topics: - `Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation <http://arxiv.org/abs/1406.1078>`__ - `Sequence to Sequence Learning with Neural Networks <http://arxiv.org/abs/1409.3215>`__ - `Neural Machine Translation by Jointly Learning to Align and Translate <https://arxiv.org/abs/1409.0473>`__ - `A Neural Conversational Model <http://arxiv.org/abs/1506.05869>`__ **Requirements** """ from __future__ import unicode_literals, print_function, division from io import open import unicodedata import string import re import random import torch import torch.nn as nn from torch.autograd import Variable from torch import optim import torch.nn.functional as F use_cuda = torch.cuda.is_available() ###################################################################### # Loading data files # ================== # # The data for this project is a set of many thousands of English to # French translation pairs. # # `This question on Open Data Stack # Exchange <http://opendata.stackexchange.com/questions/3888/dataset-of-sentences-translated-into-many-languages>`__ # pointed me to the open translation site http://tatoeba.org/ which has # downloads available at http://tatoeba.org/eng/downloads - and better # yet, someone did the extra work of splitting language pairs into # individual text files here: http://www.manythings.org/anki/ # # The English to French pairs are too big to include in the repo, so # download to ``data/eng-fra.txt`` before continuing. The file is a tab # separated list of translation pairs: # # :: # # I am cold. Je suis froid. # # .. Note:: # Download the data from # `here <https://download.pytorch.org/tutorial/data.zip>`_ # and extract it to the current directory. ###################################################################### # Similar to the character encoding used in the character-level RNN # tutorials, we will be representing each word in a language as a one-hot # vector, or giant vector of zeros except for a single one (at the index # of the word). Compared to the dozens of characters that might exist in a # language, there are many many more words, so the encoding vector is much # larger. We will however cheat a bit and trim the data to only use a few # thousand words per language. # # .. figure:: /_static/img/seq-seq-images/word-encoding.png # :alt: # # ###################################################################### # We'll need a unique index per word to use as the inputs and targets of # the networks later. To keep track of all this we will use a helper class # called ``Lang`` which has word → index (``word2index``) and index → word # (``index2word``) dictionaries, as well as a count of each word # ``word2count`` to use to later replace rare words. # SOS_token = 0 EOS_token = 1 class Lang: def __init__(self, name): self.name = name self.word2index = {} self.word2count = {} self.index2word = {0: "SOS", 1: "EOS"} self.n_words = 2 # Count SOS and EOS def addSentence(self, sentence): for word in sentence.split(' '): self.addWord(word) def addWord(self, word): if word not in self.word2index: self.word2index[word] = self.n_words self.word2count[word] = 1 self.index2word[self.n_words] = word self.n_words += 1 else: self.word2count[word] += 1 ###################################################################### # The files are all in Unicode, to simplify we will turn Unicode # characters to ASCII, make everything lowercase, and trim most # punctuation. # # Turn a Unicode string to plain ASCII, thanks to # http://stackoverflow.com/a/518232/2809427 def unicodeToAscii(s): return ''.join( c for c in unicodedata.normalize('NFD', s) if unicodedata.category(c) != 'Mn' ) # Lowercase, trim, and remove non-letter characters def normalizeString(s): s = unicodeToAscii(s.lower().strip()) s = re.sub(r"([.!?])", r" \1", s) s = re.sub(r"[^a-zA-Z.!?]+", r" ", s) return s ###################################################################### # To read the data file we will split the file into lines, and then split # lines into pairs. The files are all English → Other Language, so if we # want to translate from Other Language → English I added the ``reverse`` # flag to reverse the pairs. # def readLangs(lang1, lang2, reverse=False): print("Reading lines...") # Read the file and split into lines lines = open('data/%s-%s.txt' % (lang1, lang2), encoding='utf-8').\ read().strip().split('\n') # Split every line into pairs and normalize pairs = [[normalizeString(s) for s in l.split('\t')] for l in lines] # Reverse pairs, make Lang instances if reverse: pairs = [list(reversed(p)) for p in pairs] input_lang = Lang(lang2) output_lang = Lang(lang1) else: input_lang = Lang(lang1) output_lang = Lang(lang2) return input_lang, output_lang, pairs ###################################################################### # Since there are a *lot* of example sentences and we want to train # something quickly, we'll trim the data set to only relatively short and # simple sentences. Here the maximum length is 10 words (that includes # ending punctuation) and we're filtering to sentences that translate to # the form "I am" or "He is" etc. (accounting for apostrophes replaced # earlier). # MAX_LENGTH = 10 eng_prefixes = ( "i am ", "i m ", "he is", "he s ", "she is", "she s", "you are", "you re ", "we are", "we re ", "they are", "they re " ) def filterPair(p): return len(p[0].split(' ')) < MAX_LENGTH and \ len(p[1].split(' ')) < MAX_LENGTH and \ p[1].startswith(eng_prefixes) def filterPairs(pairs): return [pair for pair in pairs if filterPair(pair)] ###################################################################### # The full process for preparing the data is: # # - Read text file and split into lines, split lines into pairs # - Normalize text, filter by length and content # - Make word lists from sentences in pairs # def prepareData(lang1, lang2, reverse=False): input_lang, output_lang, pairs = readLangs(lang1, lang2, reverse) print("Read %s sentence pairs" % len(pairs)) pairs = filterPairs(pairs) print("Trimmed to %s sentence pairs" % len(pairs)) print("Counting words...") for pair in pairs: input_lang.addSentence(pair[0]) output_lang.addSentence(pair[1]) print("Counted words:") print(input_lang.name, input_lang.n_words) print(output_lang.name, output_lang.n_words) return input_lang, output_lang, pairs input_lang, output_lang, pairs = prepareData('eng', 'fra', True) print(random.choice(pairs)) ###################################################################### # The Seq2Seq Model # ================= # # A Recurrent Neural Network, or RNN, is a network that operates on a # sequence and uses its own output as input for subsequent steps. # # A `Sequence to Sequence network <http://arxiv.org/abs/1409.3215>`__, or # seq2seq network, or `Encoder Decoder # network <https://arxiv.org/pdf/1406.1078v3.pdf>`__, is a model # consisting of two RNNs called the encoder and decoder. The encoder reads # an input sequence and outputs a single vector, and the decoder reads # that vector to produce an output sequence. # # .. figure:: /_static/img/seq-seq-images/seq2seq.png # :alt: # # Unlike sequence prediction with a single RNN, where every input # corresponds to an output, the seq2seq model frees us from sequence # length and order, which makes it ideal for translation between two # languages. # # Consider the sentence "Je ne suis pas le chat noir" → "I am not the # black cat". Most of the words in the input sentence have a direct # translation in the output sentence, but are in slightly different # orders, e.g. "chat noir" and "black cat". Because of the "ne/pas" # construction there is also one more word in the input sentence. It would # be difficult to produce a correct translation directly from the sequence # of input words. # # With a seq2seq model the encoder creates a single vector which, in the # ideal case, encodes the "meaning" of the input sequence into a single # vector — a single point in some N dimensional space of sentences. # ###################################################################### # The Encoder # ----------- # # The encoder of a seq2seq network is a RNN that outputs some value for # every word from the input sentence. For every input word the encoder # outputs a vector and a hidden state, and uses the hidden state for the # next input word. # # .. figure:: /_static/img/seq-seq-images/encoder-network.png # :alt: # # class EncoderRNN(nn.Module): def __init__(self, input_size, hidden_size, n_layers=1): super(EncoderRNN, self).__init__() self.n_layers = n_layers self.hidden_size = hidden_size self.embedding = nn.Embedding(input_size, hidden_size) self.gru = nn.GRU(hidden_size, hidden_size) def forward(self, input, hidden): embedded = self.embedding(input).view(1, 1, -1) output = embedded for i in range(self.n_layers): output, hidden = self.gru(output, hidden) return output, hidden def initHidden(self): result = Variable(torch.zeros(1, 1, self.hidden_size)) if use_cuda: return result.cuda() else: return result ###################################################################### # The Decoder # ----------- # # The decoder is another RNN that takes the encoder output vector(s) and # outputs a sequence of words to create the translation. # ###################################################################### # Simple Decoder # ^^^^^^^^^^^^^^ # # In the simplest seq2seq decoder we use only last output of the encoder. # This last output is sometimes called the *context vector* as it encodes # context from the entire sequence. This context vector is used as the # initial hidden state of the decoder. # # At every step of decoding, the decoder is given an input token and # hidden state. The initial input token is the start-of-string ``<SOS>`` # token, and the first hidden state is the context vector (the encoder's # last hidden state). # # .. figure:: /_static/img/seq-seq-images/decoder-network.png # :alt: # # class DecoderRNN(nn.Module): def __init__(self, hidden_size, output_size, n_layers=1): super(DecoderRNN, self).__init__() self.n_layers = n_layers self.hidden_size = hidden_size self.embedding = nn.Embedding(output_size, hidden_size) self.gru = nn.GRU(hidden_size, hidden_size) self.out = nn.Linear(hidden_size, output_size) self.softmax = nn.LogSoftmax() def forward(self, input, hidden): output = self.embedding(input).view(1, 1, -1) for i in range(self.n_layers): output = F.relu(output) output, hidden = self.gru(output, hidden) output = self.softmax(self.out(output[0])) return output, hidden def initHidden(self): result = Variable(torch.zeros(1, 1, self.hidden_size)) if use_cuda: return result.cuda() else: return result ###################################################################### # I encourage you to train and observe the results of this model, but to # save space we'll be going straight for the gold and introducing the # Attention Mechanism. # ###################################################################### # Attention Decoder # ^^^^^^^^^^^^^^^^^ # # If only the context vector is passed betweeen the encoder and decoder, # that single vector carries the burden of encoding the entire sentence. # # Attention allows the decoder network to "focus" on a different part of # the encoder's outputs for every step of the decoder's own outputs. First # we calculate a set of *attention weights*. These will be multiplied by # the encoder output vectors to create a weighted combination. The result # (called ``attn_applied`` in the code) should contain information about # that specific part of the input sequence, and thus help the decoder # choose the right output words. # # .. figure:: https://i.imgur.com/1152PYf.png # :alt: # # Calculating the attention weights is done with another feed-forward # layer ``attn``, using the decoder's input and hidden state as inputs. # Because there are sentences of all sizes in the training data, to # actually create and train this layer we have to choose a maximum # sentence length (input length, for encoder outputs) that it can apply # to. Sentences of the maximum length will use all the attention weights, # while shorter sentences will only use the first few. # # .. figure:: /_static/img/seq-seq-images/attention-decoder-network.png # :alt: # # class AttnDecoderRNN(nn.Module): def __init__(self, hidden_size, output_size, n_layers=1, dropout_p=0.1, max_length=MAX_LENGTH): super(AttnDecoderRNN, self).__init__() self.hidden_size = hidden_size self.output_size = output_size self.n_layers = n_layers self.dropout_p = dropout_p self.max_length = max_length self.embedding = nn.Embedding(self.output_size, self.hidden_size) self.attn = nn.Linear(self.hidden_size * 2, self.max_length) self.attn_combine = nn.Linear(self.hidden_size * 2, self.hidden_size) self.dropout = nn.Dropout(self.dropout_p) self.gru = nn.GRU(self.hidden_size, self.hidden_size) self.out = nn.Linear(self.hidden_size, self.output_size) def forward(self, input, hidden, encoder_output, encoder_outputs): embedded = self.embedding(input).view(1, 1, -1) embedded = self.dropout(embedded) attn_weights = F.softmax( self.attn(torch.cat((embedded[0], hidden[0]), 1))) attn_applied = torch.bmm(attn_weights.unsqueeze(0), encoder_outputs.unsqueeze(0)) output = torch.cat((embedded[0], attn_applied[0]), 1) output = self.attn_combine(output).unsqueeze(0) for i in range(self.n_layers): output = F.relu(output) output, hidden = self.gru(output, hidden) output = F.log_softmax(self.out(output[0])) return output, hidden, attn_weights def initHidden(self): result = Variable(torch.zeros(1, 1, self.hidden_size)) if use_cuda: return result.cuda() else: return result ###################################################################### # .. note:: There are other forms of attention that work around the length # limitation by using a relative position approach. Read about "local # attention" in `Effective Approaches to Attention-based Neural Machine # Translation <https://arxiv.org/abs/1508.04025>`__. # # Training # ======== # # Preparing Training Data # ----------------------- # # To train, for each pair we will need an input tensor (indexes of the # words in the input sentence) and target tensor (indexes of the words in # the target sentence). While creating these vectors we will append the # EOS token to both sequences. # def indexesFromSentence(lang, sentence): return [lang.word2index[word] for word in sentence.split(' ')] def variableFromSentence(lang, sentence): indexes = indexesFromSentence(lang, sentence) indexes.append(EOS_token) result = Variable(torch.LongTensor(indexes).view(-1, 1)) if use_cuda: return result.cuda() else: return result def variablesFromPair(pair): input_variable = variableFromSentence(input_lang, pair[0]) target_variable = variableFromSentence(output_lang, pair[1]) return (input_variable, target_variable) ###################################################################### # Training the Model # ------------------ # # To train we run the input sentence through the encoder, and keep track # of every output and the latest hidden state. Then the decoder is given # the ``<SOS>`` token as its first input, and the last hidden state of the # encoder as its first hidden state. # # "Teacher forcing" is the concept of using the real target outputs as # each next input, instead of using the decoder's guess as the next input. # Using teacher forcing causes it to converge faster but `when the trained # network is exploited, it may exhibit # instability <http://minds.jacobs-university.de/sites/default/files/uploads/papers/ESNTutorialRev.pdf>`__. # # You can observe outputs of teacher-forced networks that read with # coherent grammar but wander far from the correct translation - # intuitively it has learned to represent the output grammar and can "pick # up" the meaning once the teacher tells it the first few words, but it # has not properly learned how to create the sentence from the translation # in the first place. # # Because of the freedom PyTorch's autograd gives us, we can randomly # choose to use teacher forcing or not with a simple if statement. Turn # ``teacher_forcing_ratio`` up to use more of it. # teacher_forcing_ratio = 0.5 def train(input_variable, target_variable, encoder, decoder, encoder_optimizer, decoder_optimizer, criterion, max_length=MAX_LENGTH): encoder_hidden = encoder.initHidden() encoder_optimizer.zero_grad() decoder_optimizer.zero_grad() input_length = input_variable.size()[0] target_length = target_variable.size()[0] encoder_outputs = Variable(torch.zeros(max_length, encoder.hidden_size)) encoder_outputs = encoder_outputs.cuda() if use_cuda else encoder_outputs loss = 0 for ei in range(input_length): encoder_output, encoder_hidden = encoder(input_variable[ei], encoder_hidden) encoder_outputs[ei] = encoder_output[0][0] decoder_input = Variable(torch.LongTensor([[SOS_token]])) decoder_input = decoder_input.cuda() if use_cuda else decoder_input decoder_hidden = encoder_hidden use_teacher_forcing = True if random.random() < teacher_forcing_ratio else False if use_teacher_forcing: # Teacher forcing: Feed the target as the next input for di in range(target_length): decoder_output, decoder_hidden, decoder_attention = decoder( decoder_input, decoder_hidden, encoder_output, encoder_outputs) loss += criterion(decoder_output, target_variable[di]) decoder_input = target_variable[di] # Teacher forcing else: # Without teacher forcing: use its own predictions as the next input for di in range(target_length): decoder_output, decoder_hidden, decoder_attention = decoder( decoder_input, decoder_hidden, encoder_output, encoder_outputs) topv, topi = decoder_output.data.topk(1) ni = topi[0][0] decoder_input = Variable(torch.LongTensor([[ni]])) decoder_input = decoder_input.cuda() if use_cuda else decoder_input loss += criterion(decoder_output, target_variable[di]) if ni == EOS_token: break loss.backward() encoder_optimizer.step() decoder_optimizer.step() return loss.data[0] / target_length ###################################################################### # This is a helper function to print time elapsed and estimated time # remaining given the current time and progress %. # import time import math def asMinutes(s): m = math.floor(s / 60) s -= m * 60 return '%dm %ds' % (m, s) def timeSince(since, percent): now = time.time() s = now - since es = s / (percent) rs = es - s return '%s (- %s)' % (asMinutes(s), asMinutes(rs)) ###################################################################### # The whole training process looks like this: # # - Start a timer # - Initialize optimizers and criterion # - Create set of training pairs # - Start empty losses array for plotting # # Then we call ``train`` many times and occasionally print the progress (% # of examples, time so far, estimated time) and average loss. # def trainIters(encoder, decoder, n_iters, print_every=1000, plot_every=100, learning_rate=0.01): start = time.time() plot_losses = [] print_loss_total = 0 # Reset every print_every plot_loss_total = 0 # Reset every plot_every encoder_optimizer = optim.SGD(encoder.parameters(), lr=learning_rate) decoder_optimizer = optim.SGD(decoder.parameters(), lr=learning_rate) training_pairs = [variablesFromPair(random.choice(pairs)) for i in range(n_iters)] criterion = nn.NLLLoss() for iter in range(1, n_iters + 1): training_pair = training_pairs[iter - 1] input_variable = training_pair[0] target_variable = training_pair[1] loss = train(input_variable, target_variable, encoder, decoder, encoder_optimizer, decoder_optimizer, criterion) print_loss_total += loss plot_loss_total += loss if iter % print_every == 0: print_loss_avg = print_loss_total / print_every print_loss_total = 0 print('%s (%d %d%%) %.4f' % (timeSince(start, iter / n_iters), iter, iter / n_iters * 100, print_loss_avg)) if iter % plot_every == 0: plot_loss_avg = plot_loss_total / plot_every plot_losses.append(plot_loss_avg) plot_loss_total = 0 showPlot(plot_losses) ###################################################################### # Plotting results # ---------------- # # Plotting is done with matplotlib, using the array of loss values # ``plot_losses`` saved while training. # import matplotlib.pyplot as plt import matplotlib.ticker as ticker import numpy as np def showPlot(points): plt.figure() fig, ax = plt.subplots() # this locator puts ticks at regular intervals loc = ticker.MultipleLocator(base=0.2) ax.yaxis.set_major_locator(loc) plt.plot(points) ###################################################################### # Evaluation # ========== # # Evaluation is mostly the same as training, but there are no targets so # we simply feed the decoder's predictions back to itself for each step. # Every time it predicts a word we add it to the output string, and if it # predicts the EOS token we stop there. We also store the decoder's # attention outputs for display later. # def evaluate(encoder, decoder, sentence, max_length=MAX_LENGTH): input_variable = variableFromSentence(input_lang, sentence) input_length = input_variable.size()[0] encoder_hidden = encoder.initHidden() encoder_outputs = Variable(torch.zeros(max_length, encoder.hidden_size)) encoder_outputs = encoder_outputs.cuda() if use_cuda else encoder_outputs for ei in range(input_length): encoder_output, encoder_hidden = encoder(input_variable[ei], encoder_hidden) encoder_outputs[ei] = encoder_outputs[ei] + encoder_output[0][0] decoder_input = Variable(torch.LongTensor([[SOS_token]])) # SOS decoder_input = decoder_input.cuda() if use_cuda else decoder_input decoder_hidden = encoder_hidden decoded_words = [] decoder_attentions = torch.zeros(max_length, max_length) for di in range(max_length): decoder_output, decoder_hidden, decoder_attention = decoder( decoder_input, decoder_hidden, encoder_output, encoder_outputs) decoder_attentions[di] = decoder_attention.data topv, topi = decoder_output.data.topk(1) ni = topi[0][0] if ni == EOS_token: decoded_words.append('<EOS>') break else: decoded_words.append(output_lang.index2word[ni]) decoder_input = Variable(torch.LongTensor([[ni]])) decoder_input = decoder_input.cuda() if use_cuda else decoder_input return decoded_words, decoder_attentions[:di + 1] ###################################################################### # We can evaluate random sentences from the training set and print out the # input, target, and output to make some subjective quality judgements: # def evaluateRandomly(encoder, decoder, n=10): for i in range(n): pair = random.choice(pairs) print('>', pair[0]) print('=', pair[1]) output_words, attentions = evaluate(encoder, decoder, pair[0]) output_sentence = ' '.join(output_words) print('<', output_sentence) print('') ###################################################################### # Training and Evaluating # ======================= # # With all these helper functions in place (it looks like extra work, but # it's easier to run multiple experiments easier) we can actually # initialize a network and start training. # # Remember that the input sentences were heavily filtered. For this small # dataset we can use relatively small networks of 256 hidden nodes and a # single GRU layer. After about 40 minutes on a MacBook CPU we'll get some # reasonable results. # # .. Note:: # If you run this notebook you can train, interrupt the kernel, # evaluate, and continue training later. Comment out the lines where the # encoder and decoder are initialized and run ``trainIters`` again. # hidden_size = 256 encoder1 = EncoderRNN(input_lang.n_words, hidden_size) attn_decoder1 = AttnDecoderRNN(hidden_size, output_lang.n_words, 1, dropout_p=0.1) if use_cuda: encoder1 = encoder1.cuda() attn_decoder1 = attn_decoder1.cuda() trainIters(encoder1, attn_decoder1, 75000, print_every=5000) ###################################################################### # evaluateRandomly(encoder1, attn_decoder1) ###################################################################### # Visualizing Attention # --------------------- # # A useful property of the attention mechanism is its highly interpretable # outputs. Because it is used to weight specific encoder outputs of the # input sequence, we can imagine looking where the network is focused most # at each time step. # # You could simply run ``plt.matshow(attentions)`` to see attention output # displayed as a matrix, with the columns being input steps and rows being # output steps: # output_words, attentions = evaluate( encoder1, attn_decoder1, "je suis trop froid .") plt.matshow(attentions.numpy()) ###################################################################### # For a better viewing experience we will do the extra work of adding axes # and labels: # def showAttention(input_sentence, output_words, attentions): # Set up figure with colorbar fig = plt.figure() ax = fig.add_subplot(111) cax = ax.matshow(attentions.numpy(), cmap='bone') fig.colorbar(cax) # Set up axes ax.set_xticklabels([''] + input_sentence.split(' ') + ['<EOS>'], rotation=90) ax.set_yticklabels([''] + output_words) # Show label at every tick ax.xaxis.set_major_locator(ticker.MultipleLocator(1)) ax.yaxis.set_major_locator(ticker.MultipleLocator(1)) plt.show() def evaluateAndShowAttention(input_sentence): output_words, attentions = evaluate( encoder1, attn_decoder1, input_sentence) print('input =', input_sentence) print('output =', ' '.join(output_words)) showAttention(input_sentence, output_words, attentions) evaluateAndShowAttention("elle a cinq ans de moins que moi .") evaluateAndShowAttention("elle est trop petit .") evaluateAndShowAttention("je ne crains pas de mourir .") evaluateAndShowAttention("c est un jeune directeur plein de talent .") ###################################################################### # Exercises # ========= # # - Try with a different dataset # # - Another language pair # - Human → Machine (e.g. IOT commands) # - Chat → Response # - Question → Answer # # - Replace the embeddings with pre-trained word embeddings such as word2vec or # GloVe # - Try with more layers, more hidden units, and more sentences. Compare # the training time and results. # - If you use a translation file where pairs have two of the same phrase # (``I am test \t I am test``), you can use this as an autoencoder. Try # this: # # - Train as an autoencoder # - Save only the Encoder network # - Train a new Decoder for translation from there #

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/source/Generation/Evaluation.py

# coding: utf-8 ###################################################################### # Evaluation # ========== # # Evaluation is mostly the same as training, but there are no targets so # we simply feed the decoder's predictions back to itself for each step. # Every time it predicts a word we add it to the output string, and if it # predicts the EOS token we stop there. We also store the decoder's # attention outputs for display later. # import random import torch from torch.autograd import Variable from source.AuxiliaryTools.nn_tool import variable_from_sentence import matplotlib.pyplot as plt import matplotlib.ticker as ticker SOS_token = 0 EOS_token = 1 teacher_forcing_ratio = 0.5 def evaluate(encoder, decoder, sentence, input_word_table, output_word_table, max_length, use_cuda): input_variable = variable_from_sentence(input_word_table, sentence) input_length = input_variable.size()[0] encoder_hidden = encoder.initHidden() encoder_outputs = Variable(torch.zeros(max_length, encoder.hidden_size)) encoder_outputs = encoder_outputs.cuda() if use_cuda else encoder_outputs for ei in range(input_length): encoder_output, encoder_hidden = encoder(input_variable[ei], encoder_hidden) encoder_outputs[ei] = encoder_outputs[ei] + encoder_output[0][0] decoder_input = Variable(torch.LongTensor([[SOS_token]])) # SOS decoder_input = decoder_input.cuda() if use_cuda else decoder_input decoder_hidden = encoder_hidden decoded_words = [] decoder_attentions = torch.zeros(max_length, max_length) last_time_best = 0 for di in range(max_length): decoder_output, decoder_hidden, decoder_attention = decoder( decoder_input, decoder_hidden, encoder_output, encoder_outputs) decoder_attentions[di] = decoder_attention.data topv, topi = decoder_output.data.topk(1) print('wait to check the shape') ni = topi[0][0] if ni == EOS_token: decoded_words.append('<EOS>') break else: decoded_words.append(output_word_table.index2word[ni]) decoder_input = Variable(torch.LongTensor([[ni]])) decoder_input = decoder_input.cuda() if use_cuda else decoder_input return decoded_words, decoder_attentions[:di + 1] ###################################################################### # We can evaluate random sentences from the training set and print out the # input, target, and output to make some subjective quality judgements: # def evaluate_randomly(all_pairs, encoder, decoder, input_word_table, output_word_table, max_length, use_cuda, n=10): for i in range(n): pair = random.choice(all_pairs) print('>', pair[0]) print('=', pair[1]) output_words, attentions = evaluate(encoder, decoder, pair[0], input_word_table, output_word_table, max_length, use_cuda) output_sentence = ' '.join(output_words) print('<', output_sentence) print('') ###################################################################### # For a better viewing experience we will do the extra work of adding axes # and labels: # def show_attention(input_sentence, output_words, attentions): # Set up figure with colorbar fig = plt.figure() ax = fig.add_subplot(111) cax = ax.matshow(attentions.numpy(), cmap='bone') fig.colorbar(cax) # Set up axes ax.set_xticklabels([''] + input_sentence.split(' ') + ['<EOS>'], rotation=90) ax.set_yticklabels([''] + output_words) # Show label at every tick ax.xaxis.set_major_locator(ticker.MultipleLocator(1)) ax.yaxis.set_major_locator(ticker.MultipleLocator(1)) plt.show() def evaluate_and_show_attention(input_sentence, encoder, attn_decoder, input_word_table, output_word_table): output_words, attentions = evaluate(encoder, attn_decoder, input_sentence, input_word_table, output_word_table) print('input =', input_sentence) print('output =', ' '.join(output_words)) show_attention(input_sentence, output_words, attentions)

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/source/Generation/Seq2SeqModel.py

# coding: utf-8 import torch import torch.nn as nn from torch.autograd import Variable import torch.nn.functional as F class EncoderRNN(nn.Module): def __init__(self, input_size, hidden_size, use_cuda, n_layers=1): super(EncoderRNN, self).__init__() self.n_layers = n_layers self.hidden_size = hidden_size self.use_cuda = use_cuda self.embedding = nn.Embedding(input_size, hidden_size) self.gru = nn.GRU(hidden_size, hidden_size) def forward(self, input, hidden): embedded = self.embedding(input).view(1, 1, -1) output = embedded for i in range(self.n_layers): output, hidden = self.gru(output, hidden) return output, hidden def initHidden(self): result = Variable(torch.zeros(1, 1, self.hidden_size)) if self.use_cuda: return result.cuda() else: return result class DecoderRNN(nn.Module): def __init__(self, hidden_size, output_size, use_cuda, n_layers=1): super(DecoderRNN, self).__init__() self.n_layers = n_layers self.hidden_size = hidden_size self.use_cuda = use_cuda self.embedding = nn.Embedding(output_size, hidden_size) self.gru = nn.GRU(hidden_size, hidden_size) self.out = nn.Linear(hidden_size, output_size) self.softmax = nn.LogSoftmax() def forward(self, input, hidden): output = self.embedding(input).view(1, 1, -1) for i in range(self.n_layers): output = F.relu(output) output, hidden = self.gru(output, hidden) output = self.softmax(self.out(output[0])) return output, hidden def initHidden(self): result = Variable(torch.zeros(1, 1, self.hidden_size)) if self.use_cuda: return result.cuda() else: return result class AttnDecoderRNN(nn.Module): def __init__(self, hidden_size, output_size, max_length, use_cuda, n_layers=1, dropout_p=0.1): super(AttnDecoderRNN, self).__init__() self.hidden_size = hidden_size self.output_size = output_size self.n_layers = n_layers self.dropout_p = dropout_p self.max_length = max_length self.use_cuda = use_cuda self.embedding = nn.Embedding(self.output_size, self.hidden_size) self.attn = nn.Linear(self.hidden_size * 2, self.max_length) self.attn_combine = nn.Linear(self.hidden_size * 2, self.hidden_size) self.dropout = nn.Dropout(self.dropout_p) self.gru = nn.GRU(self.hidden_size, self.hidden_size) self.out = nn.Linear(self.hidden_size, self.output_size) def forward(self, input, hidden, encoder_output, encoder_outputs): embedded = self.embedding(input).view(1, 1, -1) embedded = self.dropout(embedded) attn_weights = F.softmax(self.attn(torch.cat((embedded[0], hidden[0]), 1))) # print('debug', embedded.size(), attn_weights.unsqueeze(0).size(), encoder_outputs.unsqueeze(0).size()) attn_applied = torch.bmm(attn_weights.unsqueeze(0), encoder_outputs.unsqueeze(0)) output = torch.cat((embedded[0], attn_applied[0]), 1) output = self.attn_combine(output).unsqueeze(0) for i in range(self.n_layers): output = F.relu(output) output, hidden = self.gru(output, hidden) output = F.log_softmax(self.out(output[0])) return output, hidden, attn_weights def initHidden(self): result = Variable(torch.zeros(1, 1, self.hidden_size)) if self.use_cuda: return result.cuda() else: return result

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/source/Generation/Training.py

import time import math import random import torch import torch.nn as nn from torch.autograd import Variable from torch import optim from source.AuxiliaryTools.nn_tool import show_plot, variables_from_pair SOS_token = 0 EOS_token = 1 teacher_forcing_ratio = 0.5 def train(input_variable, target_variable, encoder, decoder, encoder_optimizer, decoder_optimizer, criterion, max_length, use_cuda): encoder_hidden = encoder.initHidden() encoder_optimizer.zero_grad() decoder_optimizer.zero_grad() input_length = input_variable.size()[0] target_length = target_variable.size()[0] encoder_outputs = Variable(torch.zeros(max_length, encoder.hidden_size)) encoder_outputs = encoder_outputs.cuda() if use_cuda else encoder_outputs loss = 0 for ei in range(input_length): encoder_output, encoder_hidden = encoder(input_variable[ei], encoder_hidden) encoder_outputs[ei] = encoder_output[0][0] decoder_input = Variable(torch.LongTensor([[SOS_token]])) decoder_input = decoder_input.cuda() if use_cuda else decoder_input decoder_hidden = encoder_hidden use_teacher_forcing = True if random.random() < teacher_forcing_ratio else False if use_teacher_forcing: # Teacher forcing: Feed the target as the next input for di in range(target_length): decoder_output, decoder_hidden, decoder_attention = decoder( decoder_input, decoder_hidden, encoder_output, encoder_outputs) loss += criterion(decoder_output, target_variable[di]) decoder_input = target_variable[di] # Teacher forcing else: # Without teacher forcing: use its own predictions as the next input for di in range(target_length): decoder_output, decoder_hidden, decoder_attention = decoder( decoder_input, decoder_hidden, encoder_output, encoder_outputs) topv, topi = decoder_output.data.topk(1) ni = topi[0][0] decoder_input = Variable(torch.LongTensor([[ni]])) decoder_input = decoder_input.cuda() if use_cuda else decoder_input loss += criterion(decoder_output, target_variable[di]) if ni == EOS_token: break loss.backward() encoder_optimizer.step() decoder_optimizer.step() return loss.data[0] / target_length def as_minutes(s): m = math.floor(s / 60) s -= m * 60 return '%dm %ds' % (m, s) def time_since(since, percent): now = time.time() s = now - since es = s / (percent) rs = es - s return '%s (- %s)' % (as_minutes(s), as_minutes(rs)) def train_iters(encoder, decoder, n_iters, pairs, input_word_table, output_word_table, max_length, use_cuda, print_every=1000, plot_every=100, learning_rate=0.01): start = time.time() plot_losses = [] print_loss_total = 0 # Reset every print_every plot_loss_total = 0 # Reset every plot_every encoder_optimizer = optim.SGD(encoder.parameters(), lr=learning_rate) decoder_optimizer = optim.SGD(decoder.parameters(), lr=learning_rate) training_pairs = [variables_from_pair(random.choice(pairs), input_word_table, output_word_table) for i in range(n_iters)] criterion = nn.NLLLoss() for iter in range(1, n_iters + 1): training_pair = training_pairs[iter - 1] input_variable = training_pair[0].cuda() if use_cuda else training_pair[0] target_variable = training_pair[1].cuda() if use_cuda else training_pair[1] loss = train(input_variable, target_variable, encoder, decoder, encoder_optimizer, decoder_optimizer, criterion, max_length, use_cuda) print_loss_total += loss plot_loss_total += loss if iter % print_every == 0: print_loss_avg = print_loss_total / print_every print_loss_total = 0 print('%s (%d %d%%) %.4f' % (time_since(start, iter / n_iters), iter, iter / n_iters * 100, print_loss_avg)) if iter % plot_every == 0: plot_loss_avg = plot_loss_total / plot_every plot_losses.append(plot_loss_avg) plot_loss_total = 0 # show_plot(plot_losses)

py

Seq2SeqDataAugmentationForLU

Seq2SeqDataAugmentationForLU-master/source/AuxiliaryTools/nn_tool.py

# coding: utf-8 from __future__ import unicode_literals, print_function, division import torch from torch.autograd import Variable import matplotlib.pyplot as plt import matplotlib.ticker as ticker SOS_token = 0 EOS_token = 1 teacher_forcing_ratio = 0.5 MAX_LENGTH = 10 def show_plot(points): plt.figure() fig, ax = plt.subplots() # this locator puts ticks at regular intervals loc = ticker.MultipleLocator(base=0.2) ax.yaxis.set_major_locator(loc) plt.plot(points) def indexes_from_sentence(word_table, sentence): return [word_table.word2index[word] for word in sentence] def variable_from_sentence(word_table, sentence): indexes = indexes_from_sentence(word_table, sentence) indexes.append(EOS_token) result = Variable(torch.LongTensor(indexes).view(-1, 1)) return result def variables_from_pair(pair, input_word_table, output_word_table): input_variable = variable_from_sentence(input_word_table, pair[0]) target_variable = variable_from_sentence(output_word_table, pair[1]) return input_variable, target_variable

py

EmpTransfo

EmpTransfo-master/train_full.py

# Copyright (c) 2019-present, HuggingFace Inc. # All rights reserved. This source code is licensed under the BSD-style license found in the LICENSE file in the root directory of this source tree. import os import math import logging from pprint import pformat from argparse import ArgumentParser from collections import defaultdict from itertools import chain import torch from torch.nn.parallel import DistributedDataParallel from torch.utils.data import DataLoader, TensorDataset from ignite.engine import Engine, Events from ignite.handlers import ModelCheckpoint from ignite.metrics import Accuracy, Loss, MetricsLambda, RunningAverage from ignite.contrib.handlers import ProgressBar, PiecewiseLinear from config import Config from ignite.contrib.handlers.tensorboard_logger import TensorboardLogger, OutputHandler, OptimizerParamsHandler from pytorch_pretrained_bert import (OpenAIAdam, OpenAIGPTDoubleHeadsModel, OpenAIGPTTokenizer, GPT2DoubleHeadsModel, GPT2Tokenizer, WEIGHTS_NAME, CONFIG_NAME, BertModel, BertTokenizer) from utils import get_dataset, get_dataset_for_daily_dialog SPECIAL_TOKENS = ["<bos>", "<eos>", "<speaker1>", "<speaker2>", "<no_emotion>", "<happiness>", "<surprise>", "<sadness>", "<disgust>", "<anger>", "<fear>", "<directive>", "<inform>", "<commissive>", "<question>", "<pad>"] MODEL_INPUTS = ["input_ids", "mc_token_ids", "lm_labels", "mc_labels", "token_type_ids", "token_emotion_ids"] PADDED_INPUTS = ["input_ids", "lm_labels", "token_type_ids", "token_emotion_ids"] logger = logging.getLogger(__file__) def average_distributed_scalar(scalar, config): """ Average a scalar over the nodes if we are in distributed training. We use this for distributed evaluation. """ if config.local_rank == -1: return scalar scalar_t = torch.tensor(scalar, dtype=torch.float, device=config.device) / torch.distributed.get_world_size() torch.distributed.all_reduce(scalar_t, op=torch.distributed.ReduceOp.SUM) return scalar_t.item() def pad_dataset(dataset, padding=0): """ Pad the dataset. This could be optimized by defining a Dataset class and padd only batches but this is simpler. """ max_l = max(len(x) for x in dataset["input_ids"]) for name in PADDED_INPUTS: dataset[name] = [x + [padding if name != "lm_labels" else -1] * (max_l - len(x)) for x in dataset[name]] return dataset def build_input_from_segments(history, emotions, reply, candidate_emotion, tokenizer, lm_labels=False, with_eos=True): """ Build a sequence of input from 3 segments: persona, history and last reply """ bos, eos, speaker1, speaker2 = tokenizer.convert_tokens_to_ids(SPECIAL_TOKENS[:4]) instance = {} #sequence = [[bos] + history[0] + list(chain(*history[1:]))] + [reply + ([eos] if with_eos else [])] #seq = [personas, history, reply] concatenate all persona sentences sequence = [[bos] + history[0]] + history[1:] +[reply +([eos] if with_eos else [])] sequence = [[speaker2 if (len(sequence)-i) % 2 else speaker1] + s for i, s in enumerate(sequence)] all_emotions = emotions + [candidate_emotion] sequence = [[all_emotions[i]] + s for i, s in enumerate(sequence)] instance["input_ids"] = list(chain(*sequence)) instance["token_type_ids"] = [speaker2 if i % 2 else speaker1 for i, s in enumerate(sequence) for _ in s] # the last for is for repeating the speaker1 and speaker2 for all tokens instance["token_emotion_ids"] = [emotions[i] for i, s in enumerate(sequence[:-1]) for _ in s]+[candidate_emotion]*len(sequence[-1]) instance["mc_token_ids"] = len(instance["input_ids"]) - 1 instance["lm_labels"] = [-1] * len(instance["input_ids"]) if lm_labels: instance["lm_labels"] = ([-1] * sum(len(s) for s in sequence[:-1])) + [-1] + sequence[-1][1:] #all -1 except for reply, reply is just the ids return instance, sequence def get_data_loaders(config, tokenizer): """ Prepare the dataset for training and evaluation """ personachat = get_dataset_for_daily_dialog(tokenizer, config.dataset_path, config.dataset_cache, SPECIAL_TOKENS) # personachat["train"] = personachat["train"][:100] # personachat["valid"] = personachat["valid"][:10] logger.info("Build inputs and labels") datasets = {"train": defaultdict(list), "valid": defaultdict(list)} gpu_max_length = 310 for dataset_name, dataset in personachat.items(): num_candidates = len(dataset[0]["utterances"][0]["candidates"]) if config.num_candidates > 0 and dataset_name == 'train': num_candidates = min(config.num_candidates, num_candidates) for dialog in dataset: for utterance in dialog["utterances"]: history = utterance["history"][-(2*config.max_history+1):] emotions = utterance["emotion"][-(2 * config.max_history + 1):] for j, candidate in enumerate(utterance["candidates"][-num_candidates:]): lm_labels = bool(j == num_candidates-1) #the true label is always the last one in list of candidates candidate_emotion = utterance['candidates_emotions'][j] instance, _ = build_input_from_segments(history, emotions, candidate, candidate_emotion, tokenizer, lm_labels) #print(len(instance["input_ids"])) if len(instance["input_ids"]) > gpu_max_length: truncated_history = [hist[:10] for hist in history] truncated_candidate = candidate[:10] instance, _ = build_input_from_segments(truncated_history, emotions, truncated_candidate, candidate_emotion, tokenizer, lm_labels) for input_name, input_array in instance.items(): datasets[dataset_name][input_name].append(input_array) datasets[dataset_name]["mc_labels"].append(num_candidates - 1) datasets[dataset_name]["n_candidates"] = num_candidates logger.info("Pad inputs and convert to Tensor") tensor_datasets = {"train": [], "valid": []} for dataset_name, dataset in datasets.items(): dataset = pad_dataset(dataset, padding=tokenizer.convert_tokens_to_ids(SPECIAL_TOKENS[-1])) for input_name in MODEL_INPUTS: tensor = torch.tensor(dataset[input_name]) if input_name != "mc_labels": tensor = tensor.view((-1, datasets[dataset_name]["n_candidates"]) + tensor.shape[1:]) tensor_datasets[dataset_name].append(tensor) logger.info("Build train and validation dataloaders") train_dataset, valid_dataset = TensorDataset(*tensor_datasets["train"]), TensorDataset(*tensor_datasets["valid"]) train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset) if config.distributed else None valid_sampler = torch.utils.data.distributed.DistributedSampler(valid_dataset) if config.distributed else None train_loader = DataLoader(train_dataset, sampler=train_sampler, batch_size=config.train_batch_size, shuffle=False) valid_loader = DataLoader(valid_dataset, sampler=valid_sampler, batch_size=config.valid_batch_size, shuffle=False) logger.info("Train dataset (Batch, Candidates, Seq length): {}".format(train_dataset.tensors[0].shape)) logger.info("Valid dataset (Batch, Candidates, Seq length): {}".format(valid_dataset.tensors[0].shape)) return train_loader, valid_loader, train_sampler, valid_sampler def train(): config_file = "configs/train_full_config.json" config = Config.from_json_file(config_file) # logging is set to INFO (resp. WARN) for main (resp. auxiliary) process. logger.info => log main process only, logger.warning => log all processes logging.basicConfig(level=logging.INFO if config.local_rank in [-1, 0] else logging.WARN) logger.warning("Running process %d", config.local_rank) # This is a logger.warning: it will be printed by all distributed processes logger.info("Arguments: %s", pformat(config)) # Initialize distributed training if needed config.distributed = (config.local_rank != -1) if config.distributed: torch.cuda.set_device(config.local_rank) config.device = torch.device("cuda", config.local_rank) torch.distributed.init_process_group(backend='nccl', init_method='env://') logger.info("Prepare tokenizer, pretrained model and optimizer - add special tokens for fine-tuning") tokenizer_class = GPT2Tokenizer if "gpt2" in config.model_checkpoint else OpenAIGPTTokenizer tokenizer = tokenizer_class.from_pretrained(config.model_checkpoint) model_class = GPT2DoubleHeadsModel if "gpt2" in config.model_checkpoint else OpenAIGPTDoubleHeadsModel model = model_class.from_pretrained(config.model_checkpoint) tokenizer.set_special_tokens(SPECIAL_TOKENS) model.set_num_special_tokens(len(SPECIAL_TOKENS)) model.to(config.device) optimizer = OpenAIAdam(model.parameters(), lr=config.lr) # Prepare model for FP16 and distributed training if needed (order is important, distributed should be the last) if config.fp16: from apex import amp # Apex is only required if we use fp16 training model, optimizer = amp.initialize(model, optimizer, opt_level=config.fp16) if config.distributed: model = DistributedDataParallel(model, device_ids=[config.local_rank], output_device=config.local_rank) logger.info("Prepare datasets") train_loader, val_loader, train_sampler, valid_sampler = get_data_loaders(config, tokenizer) # Training function and trainer def update(engine, batch): model.train() input_ids, mc_token_ids, lm_labels, mc_labels, token_type_ids, token_emotion_ids = tuple(input_tensor.to(config.device) for input_tensor in batch) lm_loss, mc_loss = model(input_ids, mc_token_ids, lm_labels, mc_labels, token_type_ids, token_emotion_ids) loss = (lm_loss * config.lm_coef + mc_loss * config.mc_coef) / config.gradient_accumulation_steps if config.fp16: with amp.scale_loss(loss, optimizer) as scaled_loss: scaled_loss.backward() torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), config.max_norm) else: loss.backward() torch.nn.utils.clip_grad_norm_(model.parameters(), config.max_norm) if engine.state.iteration % config.gradient_accumulation_steps == 0: optimizer.step() optimizer.zero_grad() return loss.item() trainer = Engine(update) # Evaluation function and evaluator (evaluator output is the input of the metrics) def inference(engine, batch): model.eval() with torch.no_grad(): batch = tuple(input_tensor.to(config.device) for input_tensor in batch) input_ids, mc_token_ids, lm_labels, mc_labels, token_type_ids, token_emotion_ids = batch #logger.info(tokenizer.decode(input_ids[0, -1, :].tolist())) model_outputs = model(input_ids, mc_token_ids, token_type_ids=token_type_ids, token_emotion_ids=token_emotion_ids) lm_logits, mc_logits = model_outputs[0], model_outputs[1] # So we can also use GPT2 outputs lm_logits_flat_shifted = lm_logits[..., :-1, :].contiguous().view(-1, lm_logits.size(-1)) lm_labels_flat_shifted = lm_labels[..., 1:].contiguous().view(-1) return (lm_logits_flat_shifted, mc_logits), (lm_labels_flat_shifted, mc_labels) evaluator = Engine(inference) # Attach evaluation to trainer: we evaluate when we start the training and at the end of each epoch trainer.add_event_handler(Events.EPOCH_COMPLETED, lambda _: evaluator.run(val_loader)) if config.n_epochs < 1: trainer.add_event_handler(Events.COMPLETED, lambda _: evaluator.run(val_loader)) if config.eval_before_start: trainer.add_event_handler(Events.STARTED, lambda _: evaluator.run(val_loader)) # Make sure distributed data samplers split the dataset nicely between the distributed processes if config.distributed: trainer.add_event_handler(Events.EPOCH_STARTED, lambda engine: train_sampler.set_epoch(engine.state.epoch)) evaluator.add_event_handler(Events.EPOCH_STARTED, lambda engine: valid_sampler.set_epoch(engine.state.epoch)) # Linearly decrease the learning rate from lr to zero scheduler = PiecewiseLinear(optimizer, "lr", [(0, config.lr), (config.n_epochs * len(train_loader), 0.0)]) trainer.add_event_handler(Events.ITERATION_STARTED, scheduler) # Prepare metrics - note how we compute distributed metrics RunningAverage(output_transform=lambda x: x).attach(trainer, "loss") metrics = {"nll": Loss(torch.nn.CrossEntropyLoss(ignore_index=-1), output_transform=lambda x: (x[0][0], x[1][0])), "accuracy": Accuracy(output_transform=lambda x: (x[0][1], x[1][1]))} metrics.update({"average_nll": MetricsLambda(average_distributed_scalar, metrics["nll"], config), "average_accuracy": MetricsLambda(average_distributed_scalar, metrics["accuracy"], config)}) metrics["average_ppl"] = MetricsLambda(math.exp, metrics["average_nll"]) for name, metric in metrics.items(): metric.attach(evaluator, name) # On the main process: add progress bar, tensorboard, checkpoints and save model, configuration and tokenizer before we start to train if config.local_rank in [-1, 0]: pbar = ProgressBar(persist=True) pbar.attach(trainer, metric_names=["loss"]) evaluator.add_event_handler(Events.COMPLETED, lambda _: pbar.log_message("Validation: %s" % pformat(evaluator.state.metrics))) tb_logger = TensorboardLogger(log_dir=config.log_dir) tb_logger.attach(trainer, log_handler=OutputHandler(tag="training", metric_names=["loss"]), event_name=Events.ITERATION_COMPLETED) tb_logger.attach(trainer, log_handler=OptimizerParamsHandler(optimizer), event_name=Events.ITERATION_STARTED) tb_logger.attach(evaluator, log_handler=OutputHandler(tag="validation", metric_names=list(metrics.keys()), another_engine=trainer), event_name=Events.EPOCH_COMPLETED) checkpoint_handler = ModelCheckpoint(tb_logger.writer.log_dir, 'checkpoint', save_interval=1, n_saved=3) trainer.add_event_handler(Events.EPOCH_COMPLETED, checkpoint_handler, {'mymodel': getattr(model, 'module', model)}) # "getattr" take care of distributed encapsulation torch.save(config, tb_logger.writer.log_dir + '/model_training_args.bin') getattr(model, 'module', model).config.to_json_file(os.path.join(tb_logger.writer.log_dir, CONFIG_NAME)) tokenizer.save_vocabulary(tb_logger.writer.log_dir) # Run the training trainer.run(train_loader, max_epochs=config.n_epochs) # On the main process: close tensorboard logger and rename the last checkpoint (for easy re-loading with OpenAIGPTModel.from_pretrained method) if config.local_rank in [-1, 0] and config.n_epochs > 0: os.rename(checkpoint_handler._saved[-1][1][-1], os.path.join(tb_logger.writer.log_dir, WEIGHTS_NAME)) # TODO: PR in ignite to have better access to saved file paths (cleaner) tb_logger.close() if __name__ == "__main__": train()

py

EmpTransfo

EmpTransfo-master/evaluate.py

# # Copyright (c) 2019-present, HuggingFace Inc. # All rights reserved. # This source code is licensed under the BSD-style license found in the # LICENSE file in the root directory of this source tree. import logging import random from argparse import ArgumentParser from itertools import chain from pprint import pformat import numpy as np import torch import torch.nn.functional as F from tqdm import tqdm from config import InteractConfig from pytorch_pretrained_bert import OpenAIGPTLMHeadModel, OpenAIGPTTokenizer, GPT2LMHeadModel, GPT2Tokenizer from utils import download_pretrained_model, get_dataset, _bleu, _f1_score def build_input_from_segments(persona, history, reply, tokenizer, SPECIAL_TOKENS, lm_labels=False, with_eos=True): """ Build a sequence of input from 3 segments: persona, history and last reply """ bos, eos, speaker1, speaker2 = tokenizer.convert_tokens_to_ids(SPECIAL_TOKENS[:-1]) instance = {} sequence = [[bos] + list(chain(*persona))] + history + [ reply + ([eos] if with_eos else [])] # seq = [personas, history, reply] concatenate all persona sentences sequence = [sequence[0]] + [[speaker2 if (len(sequence) - i) % 2 else speaker1] + s for i, s in enumerate(sequence[1:])] instance["input_ids"] = list(chain(*sequence)) instance["token_type_ids"] = [speaker2 if i % 2 else speaker1 for i, s in enumerate(sequence) for _ in s] # the last for is for repeating the speaker1 and speaker2 for all tokens instance["mc_token_ids"] = len(instance["input_ids"]) - 1 instance["lm_labels"] = [-1] * len(instance["input_ids"]) if lm_labels: instance["lm_labels"] = ([-1] * sum(len(s) for s in sequence[:-1])) + [-1] + sequence[-1][1:] # all -1 except for reply, reply is just the ids return instance, sequence def top_filtering(logits, top_k=0, top_p=0.0, threshold=-float('Inf'), filter_value=-float('Inf')): """ Filter a distribution of logits using top-k, top-p (nucleus) and/or threshold filtering Args: logits: logits distribution shape (..., vocabulary size) top_k: <=0: no filtering, >0: keep only top k tokens with highest probability. top_p: <=0.0: no filtering, >0.0: keep only a subset S of candidates, where S is the smallest subset whose total probability mass is greater than or equal to the threshold top_p. In practice, we select the highest probability tokens whose cumulative probability mass exceeds the threshold top_p. threshold: a minimal threshold to keep logits """ top_k = min(top_k, logits.size(-1)) if top_k > 0: # Remove all tokens with a probability less than the last token in the top-k tokens indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None] logits[indices_to_remove] = filter_value if top_p > 0.0: # Compute cumulative probabilities of sorted tokens sorted_logits, sorted_indices = torch.sort(logits, descending=True) cumulative_probabilities = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1) # Remove tokens with cumulative probability above the threshold sorted_indices_to_remove = cumulative_probabilities > top_p # Shift the indices to the right to keep also the first token above the threshold sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone() sorted_indices_to_remove[..., 0] = 0 # Back to unsorted indices and set them to -infinity indices_to_remove = sorted_indices[sorted_indices_to_remove] logits[indices_to_remove] = filter_value indices_to_remove = logits < threshold logits[indices_to_remove] = filter_value return logits def get_emotions(dataset): for data in tqdm(dataset['valid']): utterances = data['utterances'] for utterance in utterances: true_emotion = utterance["emotion"] def calculate_metrics(args, model, tokenizer, dataset, special_tokens): special_tokens_ids = tokenizer.convert_tokens_to_ids(special_tokens) all_blues = [] all_f1_scores = [] all_true_sentences = [] all_predicted_sentences = [] for data in tqdm(dataset['valid']): personality = data['personality'] utterances = data['utterances'] #utterance = utterances[-1] #only the longest conversaion for utterance in utterances: true_label = utterance['candidates'][-1] history = utterance['history'] predicted_output = [] for i in range(args.max_length): instance, _ = build_input_from_segments(personality, history, predicted_output, tokenizer, special_tokens, with_eos=False) try: if len(instance["input_ids"]) > 310: truncated_history = [hist[:5] for hist in history] instance, _ = build_input_from_segments(personality, truncated_history, predicted_output, tokenizer, special_tokens, with_eos=False) input_ids = torch.tensor(instance["input_ids"], device=args.device).unsqueeze(0) token_type_ids = torch.tensor(instance["token_type_ids"], device=args.device).unsqueeze(0) logits = model(input_ids, token_type_ids=token_type_ids) except: print("exception") continue if "gpt2" == args.model: logits = logits[0] logits = logits[0, -1, :] / args.temperature logits = top_filtering(logits, top_k=args.top_k, top_p=args.top_p) probs = F.softmax(logits, dim=-1) prev = torch.topk(probs, 1)[1] if args.no_sample else torch.multinomial(probs, 1) # if i < args.min_length and prev.item() in special_tokens_ids: # k=0 # while prev.item() in special_tokens_ids and k < 100: # prev = torch.multinomial(probs, num_samples=1) # k+=1 if i < args.min_length: prev = torch.multinomial(probs, num_samples=1) # if prev.item() in special_tokens_ids: # break predicted_output.append(prev.item()) predicted_sentence = tokenizer.decode(predicted_output, skip_special_tokens=True) true_sentence = tokenizer.decode(true_label, skip_special_tokens=True) #looks like zero gives the best results all_predicted_sentences.append(predicted_sentence) all_true_sentences.append(true_sentence) bleus = [_bleu(predicted_sentence, [true_sentence], method="method"+str(i)) for i in [0,1,2,3,5]] #bleu = _bleu(predicted_sentence, [true_sentence]) f1_score = _f1_score(predicted_sentence, [true_sentence]) #print(f1_score) all_blues.append(bleus) all_f1_scores.append(f1_score) #compare predicted and label with bleu print("avg bleu", np.array(all_blues).mean(axis=0)) print("avg f1 score", np.mean(all_f1_scores)) print("max bleu", np.array(all_blues).max(axis=0)) def run(): config_file = "configs/interact_config.json" config = InteractConfig.from_json_file(config_file) logging.basicConfig(level=logging.INFO) logger = logging.getLogger(__file__) logger.info(pformat(config)) if config.model_checkpoint == "": config.model_checkpoint = download_pretrained_model() random.seed(config.seed) torch.random.manual_seed(config.seed) torch.cuda.manual_seed(config.seed) logger.info("Get pretrained model and tokenizer") tokenizer_class = GPT2Tokenizer if "gpt2" == config.model else OpenAIGPTTokenizer tokenizer = tokenizer_class.from_pretrained(config.model_checkpoint) model_class = GPT2LMHeadModel if "gpt2" == config.model else OpenAIGPTLMHeadModel model = model_class.from_pretrained(config.model_checkpoint) model.to(config.device) model.eval() dataset = get_dataset(tokenizer, config.dataset_path, config.dataset_cache) special_tokens = ["<bos>", "<eos>", "<speaker1>", "<speaker2>", "<pad>"] calculate_metrics(config, model, tokenizer, dataset, special_tokens) if __name__ == "__main__": run()

py

EmpTransfo

EmpTransfo-master/utils.py

# Copyright (c) 2019-present, HuggingFace Inc. # All rights reserved. This source code is licensed under the BSD-style license found in the # LICENSE file in the root directory of this source tree. import json import logging import os import tarfile import tempfile import re import torch from pytorch_pretrained_bert import cached_path from collections import Counter try: from nltk.translate import bleu_score as nltkbleu except ImportError: # User doesn't have nltk installed, so we can't use it for bleu # We'll just turn off things, but we might want to warn the user nltkbleu = None PERSONACHAT_URL = "https://s3.amazonaws.com/datasets.huggingface.co/personachat/personachat_self_original.json" HF_FINETUNED_MODEL = "https://s3.amazonaws.com/models.huggingface.co/transfer-learning-chatbot/finetuned_chatbot_gpt.tar.gz" logger = logging.getLogger(__file__) re_art = re.compile(r'\b(a|an|the)\b') re_punc = re.compile(r'[!"#$%&()*+,-./:;<=>?@\[\]\\^`{|}~_\']') def normalize_answer(s): """Lower text and remove punctuation, articles and extra whitespace.""" def remove_articles(text): return re_art.sub(' ', text) def white_space_fix(text): return ' '.join(text.split()) def remove_punc(text): return re_punc.sub(' ', text) # convert punctuation to spaces def lower(text): return text.lower() return white_space_fix(remove_articles(remove_punc(lower(s)))) def download_pretrained_model(): """ Download and extract finetuned model from S3 """ resolved_archive_file = cached_path(HF_FINETUNED_MODEL) tempdir = tempfile.mkdtemp() logger.info("extracting archive file {} to temp dir {}".format(resolved_archive_file, tempdir)) with tarfile.open(resolved_archive_file, 'r:gz') as archive: archive.extractall(tempdir) return tempdir def get_dataset(tokenizer, dataset_path, dataset_cache=None): """ Get PERSONACHAT from S3 """ dataset_path = dataset_path or PERSONACHAT_URL dataset_cache = dataset_cache + '_' + type(tokenizer).__name__ # Do avoid using GPT cache for GPT-2 and vice-versa if dataset_cache and os.path.isfile(dataset_cache): logger.info("Load tokenized dataset from cache at %s", dataset_cache) dataset = torch.load(dataset_cache) else: logger.info("Download dataset from %s", dataset_path) personachat_file = cached_path(dataset_path) with open(personachat_file, "r", encoding="utf-8") as f: dataset = json.loads(f.read()) logger.info("Tokenize and encode the dataset") def tokenize(obj): if isinstance(obj, str): return tokenizer.convert_tokens_to_ids(tokenizer.tokenize(obj)) if isinstance(obj, dict): return dict((n, tokenize(o)) for n, o in obj.items()) return list(tokenize(o) for o in obj) dataset = tokenize(dataset) if dataset_cache: torch.save(dataset, dataset_cache) return dataset def get_dataset_for_daily_dialog(tokenizer, dataset_path, dataset_cache=None, special_tokens=None): """ Get PERSONACHAT from S3 """ dataset_path = dataset_path or PERSONACHAT_URL dataset_cache = dataset_cache + '_' + type(tokenizer).__name__ # Do avoid using GPT cache for GPT-2 and vice-versa if dataset_cache and os.path.isfile(dataset_cache): logger.info("Load tokenized dataset from cache at %s", dataset_cache) dataset = torch.load(dataset_cache) else: logger.info("Download dataset from %s", dataset_path) personachat_file = cached_path(dataset_path) with open(personachat_file, "r", encoding="utf-8") as f: dataset = json.loads(f.read()) logger.info("Tokenize and encode the dataset") def tokenize(obj): if isinstance(obj, str): if obj in special_tokens: return tokenizer.convert_tokens_to_ids(obj) else: return tokenizer.convert_tokens_to_ids(tokenizer.tokenize(obj)) if isinstance(obj, dict): return dict((n, tokenize(o)) for n, o in obj.items()) return list(tokenize(o) for o in obj) dataset = tokenize(dataset) if dataset_cache: torch.save(dataset, dataset_cache) return dataset def get_dataset_personalities(tokenizer, dataset_path, dataset_cache=None): """ Get personalities from PERSONACHAT """ dataset_path = dataset_path or PERSONACHAT_URL dataset_cache = dataset_cache + '_' + type(tokenizer).__name__ # Do avoid using GPT cache for GPT-2 and vice-versa if os.path.isfile(dataset_cache): logger.info("Load tokenized dataset from cache at %s", dataset_cache) personachat = torch.load(dataset_cache) else: logger.info("Download PERSONACHAT dataset from %s", dataset_path) personachat_file = cached_path(dataset_path) with open(personachat_file, "r", encoding="utf-8") as f: personachat = json.loads(f.read()) logger.info("Tokenize and encode the dataset") def tokenize(obj): if isinstance(obj, str): return tokenizer.convert_tokens_to_ids(tokenizer.tokenize(obj)) if isinstance(obj, dict): return dict((n, tokenize(o)) for n, o in obj.items()) return list(tokenize(o) for o in obj) personachat = tokenize(personachat) torch.save(personachat, dataset_cache) logger.info("Filter personalities") personalities = [] for dataset in personachat.values(): for dialog in dataset: personalities.append(dialog["personality"]) logger.info("Gathered {} personalities".format(len(personalities))) return personalities def _prec_recall_f1_score(pred_items, gold_items): """ Compute precision, recall and f1 given a set of gold and prediction items. :param pred_items: iterable of predicted values :param gold_items: iterable of gold values :return: tuple (p, r, f1) for precision, recall, f1 """ common = Counter(gold_items) & Counter(pred_items) num_same = sum(common.values()) if num_same == 0: return 0, 0, 0 precision = 1.0 * num_same / len(pred_items) recall = 1.0 * num_same / len(gold_items) f1 = (2 * precision * recall) / (precision + recall) return precision, recall, f1 def _f1_score(guess, answers): """Return the max F1 score between the guess and *any* answer.""" if guess is None or answers is None: return 0 g_tokens = normalize_answer(guess).split() scores = [ _prec_recall_f1_score(g_tokens, normalize_answer(a).split())for a in answers ] return max(f1 for p, r, f1 in scores) def _bleu(guess, answers, method=None): """Compute approximate BLEU score between guess and a set of answers.""" if nltkbleu is None: # bleu library not installed, just return a default value return None # Warning: BLEU calculation *should* include proper tokenization and # punctuation etc. We're using the normalize_answer for everything though, # so we're over-estimating our BLEU scores. Also note that NLTK's bleu is # going to be slower than fairseq's (which is written in C), but fairseq's # requires that everything be in arrays of ints (i.e. as tensors). NLTK's # works with strings, which is better suited for this module. if method == "method0": smoothing_func = nltkbleu.SmoothingFunction(epsilon=1e-12).method0 elif method == "method1": smoothing_func = nltkbleu.SmoothingFunction(epsilon=1e-12).method1 elif method == "method2": smoothing_func = nltkbleu.SmoothingFunction(epsilon=1e-12).method2 elif method == "method3": smoothing_func = nltkbleu.SmoothingFunction(epsilon=1e-12).method3 elif method == "method4": smoothing_func = nltkbleu.SmoothingFunction(epsilon=1e-12).method4 elif method == "method5": smoothing_func = nltkbleu.SmoothingFunction(epsilon=1e-12).method5 elif method == "method6": smoothing_func = nltkbleu.SmoothingFunction(epsilon=1e-12).method6 elif method == "method7": smoothing_func = nltkbleu.SmoothingFunction(epsilon=1e-12).method7 else: smoothing_func = nltkbleu.SmoothingFunction(epsilon=1e-12).method3 return nltkbleu.sentence_bleu( [normalize_answer(a).split(" ") for a in answers], normalize_answer(guess).split(" "), smoothing_function=smoothing_func, ) class AttrDict(dict): def __init__(self, *args, **kwargs): super(AttrDict, self).__init__(*args, **kwargs) self.__dict__ = self

py

EmpTransfo

EmpTransfo-master/train_emotion_recognition.py

# Copyright (c) 2019-present, HuggingFace Inc. # All rights reserved. This source code is licensed under the BSD-style license found in the LICENSE file in the root directory of this source tree. import os import math import logging from pprint import pformat from argparse import ArgumentParser from collections import defaultdict from itertools import chain import torch from torch.nn.parallel import DistributedDataParallel from torch.utils.data import DataLoader, TensorDataset from ignite.engine import Engine, Events from ignite.handlers import ModelCheckpoint from ignite.metrics import Accuracy, Recall, Loss, MetricsLambda, RunningAverage, Precision, ConfusionMatrix from ignite.contrib.handlers import ProgressBar, PiecewiseLinear from ignite.contrib.handlers.tensorboard_logger import TensorboardLogger, OutputHandler, OptimizerParamsHandler from config import Config from pytorch_pretrained_bert import (OpenAIAdam, OpenAIGPTDoubleHeadLMEmotionRecognitionModel, OpenAIGPTTokenizer, GPT2DoubleHeadsModel, GPT2Tokenizer, WEIGHTS_NAME, CONFIG_NAME) from utils import get_dataset, get_dataset_for_daily_dialog SPECIAL_TOKENS = ["<bos>", "<eos>", "<speaker1>", "<speaker2>", "<no_emotion>", "<happiness>", "<surprise>", "<sadness>", "<disgust>", "<anger>", "<fear>", "<directive>", "<inform>", "<commissive>", "<question>", "<pad>"] MODEL_INPUTS = ["input_ids", "mc_token_ids", "lm_labels", "mc_labels", "token_type_ids", "token_emotion_ids"] PADDED_INPUTS = ["input_ids", "lm_labels", "token_type_ids", "token_emotion_ids"] logger = logging.getLogger(__file__) def average_distributed_scalar(scalar, config): """ Average a scalar over the nodes if we are in distributed training. We use this for distributed evaluation. """ if config.local_rank == -1: return scalar scalar_t = torch.tensor(scalar, dtype=torch.float, device=config.device) / torch.distributed.get_world_size() torch.distributed.all_reduce(scalar_t, op=torch.distributed.ReduceOp.SUM) return scalar_t.item() def pad_dataset(dataset, padding=0): """ Pad the dataset. This could be optimized by defining a Dataset class and padd only batches but this is simpler. """ max_l = max(len(x) for x in dataset["input_ids"]) for name in PADDED_INPUTS: dataset[name] = [x + [padding if name != "lm_labels" else -1] * (max_l - len(x)) for x in dataset[name]] return dataset def get_emotion_label(tokenizer, candidate_emotion): _, _, _, _, no_emotion_id, happiness_id, surprise_id, sadness_id, disgust_id, anger_id, fear_id, _, _, _, _, _ = tokenizer.convert_tokens_to_ids(SPECIAL_TOKENS) if candidate_emotion == happiness_id: return 0 elif candidate_emotion == surprise_id: return 1 elif candidate_emotion == sadness_id: return 2 elif candidate_emotion == disgust_id: return 3 elif candidate_emotion == anger_id: return 4 elif candidate_emotion == fear_id: return 5 elif candidate_emotion == no_emotion_id: return 6 def build_input_from_segments(history, emotions, reply, true_emotion, tokenizer, with_eos=True): """ Build a sequence of input from 3 segments: persona, history and last reply """ bos, eos, speaker1, speaker2 = tokenizer.convert_tokens_to_ids(SPECIAL_TOKENS[:4]) #tokenizer.convert_tokens_to_ids(SPECIAL_TOKENS[-1]) instance = {} # sequence = [[bos] + history[0] + list(chain(*history[1:]))] + [reply + ([eos] if with_eos else [])] #seq = [personas, history, reply] concatenate all persona sentences sequence = [[bos] + history[0]] + history[1:] + [reply + ([eos] if with_eos else [])] sequence = [[speaker2 if (len(sequence)-i) % 2 else speaker1] + s for i, s in enumerate(sequence)] instance["input_ids"] = list(chain(*sequence)) instance["token_type_ids"] = [speaker2 if i % 2 else speaker1 for i, s in enumerate(sequence) for _ in s] # the last for is for repeating the speaker1 and speaker2 for all tokens #instance["token_emotion_ids"] = [emotions[i] for i, s in enumerate(sequence[:-1]) for _ in s] + [true_emotion] * len(sequence[-1]) instance["token_emotion_ids"] = [emotions[i] for i, s in enumerate(sequence[:-1]) for _ in s] instance["mc_token_ids"] = len(instance["input_ids"]) - 1 instance["mc_labels"] = get_emotion_label(tokenizer, true_emotion) instance["lm_labels"] = ([-1] * sum(len(s) for s in sequence[:-1])) + [-1] + sequence[-1][1:] #all -1 except for reply, reply is just the ids return instance, sequence def get_data_loaders(config, tokenizer): """ Prepare the dataset for training and evaluation """ personachat = get_dataset_for_daily_dialog(tokenizer, config.dataset_path, config.dataset_cache, SPECIAL_TOKENS) # personachat["train"] = personachat["train"][:100] # personachat["valid"] = personachat["valid"][:10] logger.info("Build inputs and labels") datasets = {"train": defaultdict(list), "valid": defaultdict(list)} gpu_max_length = 310 for dataset_name, dataset in personachat.items(): num_candidates = 2#len(dataset[0]["utterances"][0]["candidates"]) if config.num_candidates > 0 and dataset_name == 'train': num_candidates = min(config.num_candidates, num_candidates) for dialog in dataset: for utterance in dialog["utterances"]: history = utterance["history"][-(2 * config.max_history + 1):] emotions = utterance["emotion"][-(2 * config.max_history + 1):] reply = utterance["candidates"][-1] true_emotion = utterance['candidates_emotions'][-1] if true_emotion == tokenizer.convert_tokens_to_ids(SPECIAL_TOKENS)[4]: continue instance, _ = build_input_from_segments(history, emotions, reply, true_emotion, tokenizer) if len(instance["input_ids"]) > gpu_max_length: truncated_history = [hist[:10] for hist in history] truncated_candidate = reply[:10] true_emotion = utterance['candidates_emotions'][-1] instance, _ = build_input_from_segments(truncated_history, emotions, truncated_candidate, true_emotion, tokenizer) for input_name, input_array in instance.items(): datasets[dataset_name][input_name].append(input_array) datasets[dataset_name]["n_candidates"] = num_candidates logger.info("Pad inputs and convert to Tensor") tensor_datasets = {"train": [], "valid": []} for dataset_name, dataset in datasets.items(): dataset = pad_dataset(dataset, padding=tokenizer.convert_tokens_to_ids(SPECIAL_TOKENS[-1])) for input_name in MODEL_INPUTS: tensor = torch.tensor(dataset[input_name]) #if input_name != "mc_labels": # tensor = tensor.view((-1, datasets[dataset_name]["n_candidates"]) + tensor.shape[1:]) tensor_datasets[dataset_name].append(tensor) logger.info("Build train and validation dataloaders") train_dataset, valid_dataset = TensorDataset(*tensor_datasets["train"]), TensorDataset(*tensor_datasets["valid"]) train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset) if config.distributed else None valid_sampler = torch.utils.data.distributed.DistributedSampler(valid_dataset) if config.distributed else None train_loader = DataLoader(train_dataset, sampler=train_sampler, batch_size=config.train_batch_size, shuffle=False) valid_loader = DataLoader(valid_dataset, sampler=valid_sampler, batch_size=config.valid_batch_size, shuffle=False) logger.info("Train dataset (Batch, Candidates, Seq length): {}".format(train_dataset.tensors[0].shape)) logger.info("Valid dataset (Batch, Candidates, Seq length): {}".format(valid_dataset.tensors[0].shape)) return train_loader, valid_loader, train_sampler, valid_sampler def train(): config_file = "configs/train_emotion_recognition_config.json" config = Config.from_json_file(config_file) # logging is set to INFO (resp. WARN) for main (resp. auxiliary) process. logger.info => log main process only, logger.warning => log all processes logging.basicConfig(level=logging.INFO if config.local_rank in [-1, 0] else logging.WARN) logger.warning("Running process %d", config.local_rank) # This is a logger.warning: it will be printed by all distributed processes logger.info("Arguments: %s", pformat(config)) # Initialize distributed training if needed config.distributed = (config.local_rank != -1) if config.distributed: torch.cuda.set_device(config.local_rank) config.device = torch.device("cuda", config.local_rank) torch.distributed.init_process_group(backend='nccl', init_method='env://') logger.info("Prepare tokenizer, pretrained model and optimizer - add special tokens for fine-tuning") tokenizer_class = GPT2Tokenizer if "gpt2" in config.model_checkpoint else OpenAIGPTTokenizer tokenizer = tokenizer_class.from_pretrained(config.model_checkpoint) model_class = OpenAIGPTDoubleHeadLMEmotionRecognitionModel model = model_class.from_pretrained(config.model_checkpoint) tokenizer.set_special_tokens(SPECIAL_TOKENS) model.set_num_special_tokens(len(SPECIAL_TOKENS)) model.to(config.device) optimizer = OpenAIAdam(model.parameters(), lr=config.lr) # Prepare model for FP16 and distributed training if needed (order is important, distributed should be the last) if config.fp16: from apex import amp # Apex is only required if we use fp16 training model, optimizer = amp.initialize(model, optimizer, opt_level=config.fp16) if config.distributed: model = DistributedDataParallel(model, device_ids=[config.local_rank], output_device=config.local_rank) logger.info("Prepare datasets") train_loader, val_loader, train_sampler, valid_sampler = get_data_loaders(config, tokenizer) # Training function and trainer def update(engine, batch): model.train() input_ids, mc_token_ids, lm_labels, mc_labels, token_type_ids, token_emotion_ids = tuple(input_tensor.to(config.device) for input_tensor in batch) #token_emotion_ids = None lm_loss, mc_loss = model(input_ids, mc_token_ids, lm_labels, mc_labels, token_type_ids, token_emotion_ids) loss = (lm_loss * config.lm_coef + mc_loss * config.mc_coef) / config.gradient_accumulation_steps if config.fp16: with amp.scale_loss(loss, optimizer) as scaled_loss: scaled_loss.backward() torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), config.max_norm) else: loss.backward() torch.nn.utils.clip_grad_norm_(model.parameters(), config.max_norm) if engine.state.iteration % config.gradient_accumulation_steps == 0: optimizer.step() optimizer.zero_grad() return loss.item() trainer = Engine(update) # Evaluation function and evaluator (evaluator output is the input of the metrics) def inference(engine, batch): model.eval() with torch.no_grad(): batch = tuple(input_tensor.to(config.device) for input_tensor in batch) input_ids, mc_token_ids, lm_labels, mc_labels, token_type_ids, token_emotion_ids = batch #token_emotion_ids = None model_outputs = model(input_ids, mc_token_ids, token_type_ids=token_type_ids, token_emotion_ids=token_emotion_ids) lm_logits, mc_logits = model_outputs[0], model_outputs[1] # So we can also use GPT2 outputs lm_logits_flat_shifted = lm_logits[..., :-1, :].contiguous().view(-1, lm_logits.size(-1)) lm_labels_flat_shifted = lm_labels[..., 1:].contiguous().view(-1) return (lm_logits_flat_shifted, mc_logits), (lm_labels_flat_shifted, mc_labels) evaluator = Engine(inference) # Attach evaluation to trainer: we evaluate when we start the training and at the end of each epoch trainer.add_event_handler(Events.EPOCH_COMPLETED, lambda _: evaluator.run(val_loader)) if config.n_epochs < 1: trainer.add_event_handler(Events.COMPLETED, lambda _: evaluator.run(val_loader)) if config.eval_before_start: trainer.add_event_handler(Events.STARTED, lambda _: evaluator.run(val_loader)) # Make sure distributed data samplers split the dataset nicely between the distributed processes if config.distributed: trainer.add_event_handler(Events.EPOCH_STARTED, lambda engine: train_sampler.set_epoch(engine.state.epoch)) evaluator.add_event_handler(Events.EPOCH_STARTED, lambda engine: valid_sampler.set_epoch(engine.state.epoch)) # Linearly decrease the learning rate from lr to zero scheduler = PiecewiseLinear(optimizer, "lr", [(0, config.lr), (config.n_epochs * len(train_loader), 0.0)]) trainer.add_event_handler(Events.ITERATION_STARTED, scheduler) # Prepare metrics - note how we compute distributed metrics RunningAverage(output_transform=lambda x: x).attach(trainer, "loss") metrics = {"nll": Loss(torch.nn.CrossEntropyLoss(ignore_index=-1), output_transform=lambda x: (x[0][0], x[1][0])), "accuracy": Accuracy(output_transform=lambda x: (x[0][1], x[1][1]))} metrics.update({"precision": Precision(output_transform=lambda x: (x[0][1], x[1][1])), "recall": Recall(output_transform=lambda x: (x[0][1], x[1][1]))}) metrics.update({"average_nll": MetricsLambda(average_distributed_scalar, metrics["nll"], config), "average_accuracy": MetricsLambda(average_distributed_scalar, metrics["accuracy"], config)}) metrics.update({"confusion_matrix": ConfusionMatrix(num_classes=6, output_transform=lambda x: (x[0][1], x[1][1]))}) metrics["average_ppl"] = MetricsLambda(math.exp, metrics["average_nll"]) for name, metric in metrics.items(): metric.attach(evaluator, name) # On the main process: add progress bar, tensorboard, checkpoints and save model, configuration and tokenizer before we start to train if config.local_rank in [-1, 0]: pbar = ProgressBar(persist=True) pbar.attach(trainer, metric_names=["loss"]) evaluator.add_event_handler(Events.COMPLETED, lambda _: pbar.log_message("Validation: %s" % pformat(evaluator.state.metrics))) tb_logger = TensorboardLogger(log_dir=config.log_dir) tb_logger.attach(trainer, log_handler=OutputHandler(tag="training", metric_names=["loss"]), event_name=Events.ITERATION_COMPLETED) tb_logger.attach(trainer, log_handler=OptimizerParamsHandler(optimizer), event_name=Events.ITERATION_STARTED) tb_logger.attach(evaluator, log_handler=OutputHandler(tag="validation", metric_names=list(metrics.keys()), another_engine=trainer), event_name=Events.EPOCH_COMPLETED) checkpoint_handler = ModelCheckpoint(tb_logger.writer.log_dir, 'checkpoint', save_interval=1, n_saved=3) trainer.add_event_handler(Events.EPOCH_COMPLETED, checkpoint_handler, {'mymodel': getattr(model, 'module', model)}) # "getattr" take care of distributed encapsulation torch.save(config, tb_logger.writer.log_dir + '/model_training_args.bin') getattr(model, 'module', model).config.to_json_file(os.path.join(tb_logger.writer.log_dir, CONFIG_NAME)) tokenizer.save_vocabulary(tb_logger.writer.log_dir) # Run the training trainer.run(train_loader, max_epochs=config.n_epochs) # On the main process: close tensorboard logger and rename the last checkpoint (for easy re-loading with OpenAIGPTModel.from_pretrained method) if config.local_rank in [-1, 0] and config.n_epochs > 0: os.rename(checkpoint_handler._saved[-1][1][-1], os.path.join(tb_logger.writer.log_dir, WEIGHTS_NAME)) # TODO: PR in ignite to have better access to saved file paths (cleaner) tb_logger.close() if __name__ == "__main__": train()

py

EmpTransfo

EmpTransfo-master/interact.py

# # Copyright (c) 2019-present, HuggingFace Inc. # All rights reserved. # This source code is licensed under the BSD-style license found in the # LICENSE file in the root directory of this source tree. import logging import random from argparse import ArgumentParser from itertools import chain from pprint import pformat import torch import torch.nn.functional as F from config import InteractConfig from pytorch_pretrained_bert import OpenAIGPTLMHeadModel, OpenAIGPTTokenizer, GPT2LMHeadModel, GPT2Tokenizer, \ BertTokenizer from pytorch_pretrained_bert.modeling import BertLMHeadModel from utils import get_dataset_personalities, download_pretrained_model, get_dataset def build_input_from_segments(history, reply, tokenizer, SPECIAL_TOKENS, lm_labels=False, with_eos=True): """ Build a sequence of input from 3 segments: persona, history and last reply """ bos, eos, speaker1, speaker2 = tokenizer.convert_tokens_to_ids(SPECIAL_TOKENS[:-1]) persona = [] instance = {} sequence = [[bos] + list(chain(*persona))] + history + [ reply + ([eos] if with_eos else [])] # seq = [personas, history, reply] concatenate all persona sentences sequence = [sequence[0]] + [[speaker2 if (len(sequence) - i) % 2 else speaker1] + s for i, s in enumerate(sequence[1:])] instance["input_ids"] = list(chain(*sequence)) instance["token_type_ids"] = [speaker2 if i % 2 else speaker1 for i, s in enumerate(sequence) for _ in s] # the last for is for repeating the speaker1 and speaker2 for all tokens instance["mc_token_ids"] = len(instance["input_ids"]) - 1 instance["lm_labels"] = [-1] * len(instance["input_ids"]) if lm_labels: instance["lm_labels"] = ([-1] * sum(len(s) for s in sequence[:-1])) + [-1] + sequence[-1][1:] # all -1 except for reply, reply is just the ids return instance, sequence def top_filtering(logits, top_k=0, top_p=0.0, threshold=-float('Inf'), filter_value=-float('Inf')): """ Filter a distribution of logits using top-k, top-p (nucleus) and/or threshold filtering Args: logits: logits distribution shape (..., vocabulary size) top_k: <=0: no filtering, >0: keep only top k tokens with highest probability. top_p: <=0.0: no filtering, >0.0: keep only a subset S of candidates, where S is the smallest subset whose total probability mass is greater than or equal to the threshold top_p. In practice, we select the highest probability tokens whose cumulative probability mass exceeds the threshold top_p. threshold: a minimal threshold to keep logits """ top_k = min(top_k, logits.size(-1)) if top_k > 0: # Remove all tokens with a probability less than the last token in the top-k tokens indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None] logits[indices_to_remove] = filter_value if top_p > 0.0: # Compute cumulative probabilities of sorted tokens sorted_logits, sorted_indices = torch.sort(logits, descending=True) cumulative_probabilities = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1) # Remove tokens with cumulative probability above the threshold sorted_indices_to_remove = cumulative_probabilities > top_p # Shift the indices to the right to keep also the first token above the threshold sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone() sorted_indices_to_remove[..., 0] = 0 # Back to unsorted indices and set them to -infinity indices_to_remove = sorted_indices[sorted_indices_to_remove] logits[indices_to_remove] = filter_value indices_to_remove = logits < threshold logits[indices_to_remove] = filter_value return logits def sample_sequence(history, tokenizer, model, args, SPECIAL_TOKENS, current_output=None): special_tokens_ids = tokenizer.convert_tokens_to_ids(SPECIAL_TOKENS) if current_output is None: current_output = [] for i in range(args.max_length): instance, sequence = build_input_from_segments(history, current_output, tokenizer, SPECIAL_TOKENS, with_eos=False) input_ids = torch.tensor(instance["input_ids"], device=args.device).unsqueeze(0) token_type_ids = torch.tensor(instance["token_type_ids"], device=args.device).unsqueeze(0) logits = model(input_ids, token_type_ids=token_type_ids) if "gpt2" == args.model: logits = logits[0] logits = logits[0, -1, :] / args.temperature logits = top_filtering(logits, top_k=args.top_k, top_p=args.top_p) probs = F.softmax(logits, dim=-1) prev = torch.topk(probs, 1)[1] if args.no_sample else torch.multinomial(probs, 1) if i < args.min_length and prev.item() in special_tokens_ids: while prev.item() in special_tokens_ids: prev = torch.multinomial(probs, num_samples=1) if prev.item() in special_tokens_ids: break current_output.append(prev.item()) return current_output def run(): config_file = "configs/interact_config.json" config = InteractConfig.from_json_file(config_file) logging.basicConfig(level=logging.INFO) logger = logging.getLogger(__file__) logger.info(pformat(config)) if config.model_checkpoint == "": config.model_checkpoint = download_pretrained_model() torch.random.manual_seed(config.seed) torch.cuda.manual_seed(config.seed) logger.info("Get pretrained model and tokenizer") if config.model == "bert": tokenizer_class = BertTokenizer model_class = BertLMHeadModel elif config.model == "gpt2": tokenizer_class = GPT2Tokenizer model_class = GPT2LMHeadModel else: tokenizer_class = OpenAIGPTTokenizer model_class = OpenAIGPTLMHeadModel SPECIAL_TOKENS = ["<bos>", "<eos>", "<speaker1>", "<speaker2>", "<pad>"] tokenizer = tokenizer_class.from_pretrained(config.model_checkpoint) model = model_class.from_pretrained(config.model_checkpoint) model.to(config.device) model.eval() history = [] while True: raw_text = input(">>> ") while not raw_text: print('Prompt should not be empty!') raw_text = input(">>> ") history.append(tokenizer.encode(raw_text)) with torch.no_grad(): out_ids = sample_sequence(history, tokenizer, model, config, SPECIAL_TOKENS) history.append(out_ids) history = history[-(2 * config.max_history + 1):] out_text = tokenizer.decode(out_ids, skip_special_tokens=True) print(out_text) if __name__ == "__main__": run()

py

EmpTransfo

EmpTransfo-master/train.py

# Copyright (c) 2019-present, HuggingFace Inc. # All rights reserved. This source code is licensed under the BSD-style license found in the LICENSE file in the root directory of this source tree. import os import math import logging from pprint import pformat from argparse import ArgumentParser from collections import defaultdict from itertools import chain from config import Config import torch from torch.nn.parallel import DistributedDataParallel from torch.utils.data import DataLoader, TensorDataset from ignite.engine import Engine, Events from ignite.handlers import ModelCheckpoint from ignite.metrics import Accuracy, Loss, MetricsLambda, RunningAverage from ignite.contrib.handlers import ProgressBar, PiecewiseLinear from ignite.contrib.handlers.tensorboard_logger import TensorboardLogger, OutputHandler, OptimizerParamsHandler from pytorch_pretrained_bert import (OpenAIAdam, OpenAIGPTDoubleHeadsModel, OpenAIGPTTokenizer, GPT2DoubleHeadsModel, GPT2Tokenizer, WEIGHTS_NAME, CONFIG_NAME, BertModel, BertTokenizer) from utils import get_dataset SPECIAL_TOKENS = ["<bos>", "<eos>", "<speaker1>", "<speaker2>", "<pad>"] MODEL_INPUTS = ["input_ids", "mc_token_ids", "lm_labels", "mc_labels", "token_type_ids"] PADDED_INPUTS = ["input_ids", "lm_labels", "token_type_ids"] logger = logging.getLogger(__file__) def average_distributed_scalar(scalar, config): """ Average a scalar over the nodes if we are in distributed training. We use this for distributed evaluation. """ if config.local_rank == -1: return scalar scalar_t = torch.tensor(scalar, dtype=torch.float, device=config.device) / torch.distributed.get_world_size() torch.distributed.all_reduce(scalar_t, op=torch.distributed.ReduceOp.SUM) return scalar_t.item() def pad_dataset(dataset, padding=0): """ Pad the dataset. This could be optimized by defining a Dataset class and padd only batches but this is simpler. """ max_l = max(len(x) for x in dataset["input_ids"]) for name in PADDED_INPUTS: dataset[name] = [x + [padding if name != "lm_labels" else -1] * (max_l - len(x)) for x in dataset[name]] return dataset def build_input_from_segments(history, reply, tokenizer, lm_labels=False, with_eos=True): """ Build a sequence of input from 3 segments: persona, history and last reply """ bos, eos, speaker1, speaker2 = tokenizer.convert_tokens_to_ids(SPECIAL_TOKENS[:-1]) instance = {} sequence = [[bos] + history[0]] + history[1:] +[reply +([eos] if with_eos else [])] sequence = [sequence[0]] + [[speaker2 if (len(sequence)-i) % 2 else speaker1] + s for i, s in enumerate(sequence[1:])] instance["input_ids"] = list(chain(*sequence)) instance["token_type_ids"] = [speaker2 if i % 2 else speaker1 for i, s in enumerate(sequence) for _ in s] # the last for is for repeating the speaker1 and speaker2 for all tokens instance["mc_token_ids"] = len(instance["input_ids"]) - 1 instance["lm_labels"] = [-1] * len(instance["input_ids"]) if lm_labels: instance["lm_labels"] = ([-1] * sum(len(s) for s in sequence[:-1])) + [-1] + sequence[-1][1:] #all -1 except for reply, reply is just the ids return instance, sequence def get_data_loaders(config, tokenizer): """ Prepare the dataset for training and evaluation """ personachat = get_dataset(tokenizer, config.dataset_path, config.dataset_cache) logger.info("Build inputs and labels") datasets = {"train": defaultdict(list), "valid": defaultdict(list)} gpu_max_length = 310 #this depends on the gpu memory size, using bigger gpu memory you can increase this to include longer inputs for dataset_name, dataset in personachat.items(): num_candidates = len(dataset[0]["utterances"][0]["candidates"]) if config.num_candidates > 0 and dataset_name == 'train': num_candidates = min(config.num_candidates, num_candidates) for dialog in dataset: for utterance in dialog["utterances"]: history = utterance["history"][-(2*config.max_history+1):] for j, candidate in enumerate(utterance["candidates"][-num_candidates:]): lm_labels = bool(j == num_candidates-1) #the true label is always the last one in list of candidates instance, _ = build_input_from_segments(history, candidate, tokenizer, lm_labels) #print(len(instance["input_ids"])) ## if len(instance["input_ids"]) > gpu_max_length: truncated_history = [hist[:10] for hist in history] truncated_candidate = candidate[:10] instance, _ = build_input_from_segments(truncated_history, truncated_candidate, tokenizer, lm_labels) for input_name, input_array in instance.items(): datasets[dataset_name][input_name].append(input_array) datasets[dataset_name]["mc_labels"].append(num_candidates - 1) datasets[dataset_name]["n_candidates"] = num_candidates logger.info("Pad inputs and convert to Tensor") tensor_datasets = {"train": [], "valid": []} for dataset_name, dataset in datasets.items(): dataset = pad_dataset(dataset, padding=tokenizer.convert_tokens_to_ids(SPECIAL_TOKENS[-1])) for input_name in MODEL_INPUTS: tensor = torch.tensor(dataset[input_name]) if input_name != "mc_labels": tensor = tensor.view((-1, datasets[dataset_name]["n_candidates"]) + tensor.shape[1:]) tensor_datasets[dataset_name].append(tensor) logger.info("Build train and validation dataloaders") train_dataset, valid_dataset = TensorDataset(*tensor_datasets["train"]), TensorDataset(*tensor_datasets["valid"]) train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset) if config.distributed else None valid_sampler = torch.utils.data.distributed.DistributedSampler(valid_dataset) if config.distributed else None train_loader = DataLoader(train_dataset, sampler=train_sampler, batch_size=config.train_batch_size, shuffle=False) valid_loader = DataLoader(valid_dataset, sampler=valid_sampler, batch_size=config.valid_batch_size, shuffle=False) logger.info("Train dataset (Batch, Candidates, Seq length): {}".format(train_dataset.tensors[0].shape)) logger.info("Valid dataset (Batch, Candidates, Seq length): {}".format(valid_dataset.tensors[0].shape)) return train_loader, valid_loader, train_sampler, valid_sampler def train(): config_file = "configs/train_full_config.json" config = Config.from_json_file(config_file) # logging is set to INFO (resp. WARN) for main (resp. auxiliary) process. logger.info => log main process only, logger.warning => log all processes logging.basicConfig(level=logging.INFO if config.local_rank in [-1, 0] else logging.WARN) logger.warning("Running process %d", config.local_rank) # This is a logger.warning: it will be printed by all distributed processes logger.info("Arguments: %s", pformat(config)) # Initialize distributed training if needed config.distributed = (config.local_rank != -1) if config.distributed: torch.cuda.set_device(config.local_rank) config.device = torch.device("cuda", config.local_rank) torch.distributed.init_process_group(backend='nccl', init_method='env://') logger.info("Prepare tokenizer, pretrained model and optimizer - add special tokens for fine-tuning") tokenizer_class = GPT2Tokenizer if "gpt2" in config.model_checkpoint else OpenAIGPTTokenizer tokenizer = tokenizer_class.from_pretrained(config.model_checkpoint) model_class = GPT2DoubleHeadsModel if "gpt2" in config.model_checkpoint else OpenAIGPTDoubleHeadsModel model = model_class.from_pretrained(config.model_checkpoint) tokenizer.set_special_tokens(SPECIAL_TOKENS) model.set_num_special_tokens(len(SPECIAL_TOKENS)) model.to(config.device) optimizer = OpenAIAdam(model.parameters(), lr=config.lr) # Prepare model for FP16 and distributed training if needed (order is important, distributed should be the last) if config.fp16: from apex import amp # Apex is only required if we use fp16 training model, optimizer = amp.initialize(model, optimizer, opt_level=config.fp16) if config.distributed: model = DistributedDataParallel(model, device_ids=[config.local_rank], output_device=config.local_rank) logger.info("Prepare datasets") train_loader, val_loader, train_sampler, valid_sampler = get_data_loaders(config, tokenizer) # Training function and trainer def update(engine, batch): model.train() batch = tuple(input_tensor.to(config.device) for input_tensor in batch) lm_loss, mc_loss = model(*batch) loss = (lm_loss * config.lm_coef + mc_loss * config.mc_coef) / config.gradient_accumulation_steps if config.fp16: with amp.scale_loss(loss, optimizer) as scaled_loss: scaled_loss.backward() torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), config.max_norm) else: loss.backward() torch.nn.utils.clip_grad_norm_(model.parameters(), config.max_norm) if engine.state.iteration % config.gradient_accumulation_steps == 0: optimizer.step() optimizer.zero_grad() return loss.item() trainer = Engine(update) # Evaluation function and evaluator (evaluator output is the input of the metrics) def inference(engine, batch): model.eval() with torch.no_grad(): batch = tuple(input_tensor.to(config.device) for input_tensor in batch) input_ids, mc_token_ids, lm_labels, mc_labels, token_type_ids = batch #logger.info(tokenizer.decode(input_ids[0, -1, :].tolist())) model_outputs = model(input_ids, mc_token_ids, token_type_ids=token_type_ids) lm_logits, mc_logits = model_outputs[0], model_outputs[1] # So we can also use GPT2 outputs lm_logits_flat_shifted = lm_logits[..., :-1, :].contiguous().view(-1, lm_logits.size(-1)) lm_labels_flat_shifted = lm_labels[..., 1:].contiguous().view(-1) return (lm_logits_flat_shifted, mc_logits), (lm_labels_flat_shifted, mc_labels) evaluator = Engine(inference) # Attach evaluation to trainer: we evaluate when we start the training and at the end of each epoch trainer.add_event_handler(Events.EPOCH_COMPLETED, lambda _: evaluator.run(val_loader)) if config.n_epochs < 1: trainer.add_event_handler(Events.COMPLETED, lambda _: evaluator.run(val_loader)) if config.eval_before_start: trainer.add_event_handler(Events.STARTED, lambda _: evaluator.run(val_loader)) # Make sure distributed data samplers split the dataset nicely between the distributed processes if config.distributed: trainer.add_event_handler(Events.EPOCH_STARTED, lambda engine: train_sampler.set_epoch(engine.state.epoch)) evaluator.add_event_handler(Events.EPOCH_STARTED, lambda engine: valid_sampler.set_epoch(engine.state.epoch)) # Linearly decrease the learning rate from lr to zero scheduler = PiecewiseLinear(optimizer, "lr", [(0, config.lr), (config.n_epochs * len(train_loader), 0.0)]) trainer.add_event_handler(Events.ITERATION_STARTED, scheduler) # Prepare metrics - note how we compute distributed metrics RunningAverage(output_transform=lambda x: x).attach(trainer, "loss") metrics = {"nll": Loss(torch.nn.CrossEntropyLoss(ignore_index=-1), output_transform=lambda x: (x[0][0], x[1][0])), "accuracy": Accuracy(output_transform=lambda x: (x[0][1], x[1][1]))} metrics.update({"average_nll": MetricsLambda(average_distributed_scalar, metrics["nll"], config), "average_accuracy": MetricsLambda(average_distributed_scalar, metrics["accuracy"], config)}) metrics["average_ppl"] = MetricsLambda(math.exp, metrics["average_nll"]) for name, metric in metrics.items(): metric.attach(evaluator, name) # On the main process: add progress bar, tensorboard, checkpoints and save model, configuration and tokenizer before we start to train if config.local_rank in [-1, 0]: pbar = ProgressBar(persist=True) pbar.attach(trainer, metric_names=["loss"]) evaluator.add_event_handler(Events.COMPLETED, lambda _: pbar.log_message("Validation: %s" % pformat(evaluator.state.metrics))) tb_logger = TensorboardLogger(log_dir=config.log_dir) tb_logger.attach(trainer, log_handler=OutputHandler(tag="training", metric_names=["loss"]), event_name=Events.ITERATION_COMPLETED) tb_logger.attach(trainer, log_handler=OptimizerParamsHandler(optimizer), event_name=Events.ITERATION_STARTED) tb_logger.attach(evaluator, log_handler=OutputHandler(tag="validation", metric_names=list(metrics.keys()), another_engine=trainer), event_name=Events.EPOCH_COMPLETED) checkpoint_handler = ModelCheckpoint(tb_logger.writer.log_dir, 'checkpoint', save_interval=1, n_saved=3) trainer.add_event_handler(Events.EPOCH_COMPLETED, checkpoint_handler, {'mymodel': getattr(model, 'module', model)}) # "getattr" take care of distributed encapsulation torch.save(config, tb_logger.writer.log_dir + '/model_training_args.bin') getattr(model, 'module', model).config.to_json_file(os.path.join(tb_logger.writer.log_dir, CONFIG_NAME)) tokenizer.save_vocabulary(tb_logger.writer.log_dir) # Run the training trainer.run(train_loader, max_epochs=config.n_epochs) # On the main process: close tensorboard logger and rename the last checkpoint (for easy re-loading with OpenAIGPTModel.from_pretrained method) if config.local_rank in [-1, 0] and config.n_epochs > 0: os.rename(checkpoint_handler._saved[-1][1][-1], os.path.join(tb_logger.writer.log_dir, WEIGHTS_NAME)) # TODO: PR in ignite to have better access to saved file paths (cleaner) tb_logger.close() if __name__ == "__main__": train()

py

EmpTransfo

EmpTransfo-master/train_multihead.py

# Copyright (c) 2019-present, HuggingFace Inc. # All rights reserved. This source code is licensed under the BSD-style license found in the LICENSE file in the root directory of this source tree. import os import math import logging from pprint import pformat from argparse import ArgumentParser from collections import defaultdict from itertools import chain import torch from torch.nn.parallel import DistributedDataParallel from torch.utils.data import DataLoader, TensorDataset from ignite.engine import Engine, Events from ignite.handlers import ModelCheckpoint from ignite.metrics import Accuracy, Loss, MetricsLambda, RunningAverage from ignite.contrib.handlers import ProgressBar, PiecewiseLinear from config import Config from ignite.contrib.handlers.tensorboard_logger import TensorboardLogger, OutputHandler, OptimizerParamsHandler from pytorch_pretrained_bert import (OpenAIAdam, OpenAIGPTMultiHeadModel, OpenAIGPTTokenizer, GPT2DoubleHeadsModel, GPT2Tokenizer, WEIGHTS_NAME, CONFIG_NAME, BertModel, BertTokenizer) from utils import get_dataset, get_dataset_for_daily_dialog SPECIAL_TOKENS = ["<bos>", "<eos>", "<speaker1>", "<speaker2>", "<no_emotion>", "<happiness>", "<surprise>", "<sadness>", "<disgust>", "<anger>", "<fear>", "<work>", "<finance>", "<relationship>", "<attitude_and_emotion>", "<culture_and_education>", "<school_life>", "<tourism>", "<ordinary_life>", "<politics>", "<health>", "<directive>", "<inform>", "<commissive>", "<question>", "<pad>"] MODEL_INPUTS = ["input_ids", "ec_token_ids", "sc_token_ids", "lm_labels", "ec_labels", "sc_labels", "token_type_ids", "token_emotion_ids", "token_action_ids"] PADDED_INPUTS = ["input_ids", "lm_labels", "token_type_ids", "token_emotion_ids", "token_action_ids"] logger = logging.getLogger(__file__) def average_distributed_scalar(scalar, config): """ Average a scalar over the nodes if we are in distributed training. We use this for distributed evaluation. """ if config.local_rank == -1: return scalar scalar_t = torch.tensor(scalar, dtype=torch.float, device=config.device) / torch.distributed.get_world_size() torch.distributed.all_reduce(scalar_t, op=torch.distributed.ReduceOp.SUM) return scalar_t.item() def pad_dataset(dataset, padding=0): """ Pad the dataset. This could be optimized by defining a Dataset class and padd only batches but this is simpler. """ max_l = max(len(x) for x in dataset["input_ids"]) for name in PADDED_INPUTS: dataset[name] = [x + [padding if name != "lm_labels" else -1] * (max_l - len(x)) for x in dataset[name]] return dataset def get_emotion_label(tokenizer, candidate_emotion): no_emotion_id, happiness_id, surprise_id, sadness_id, disgust_id, anger_id, fear_id = tokenizer.convert_tokens_to_ids(SPECIAL_TOKENS[4:11]) if candidate_emotion == no_emotion_id: return 0 elif candidate_emotion == happiness_id: return 1 elif candidate_emotion == surprise_id: return 2 elif candidate_emotion == sadness_id: return 3 elif candidate_emotion == disgust_id: return 4 elif candidate_emotion == anger_id: return 5 elif candidate_emotion == fear_id: return 6 def build_input_from_segments(topic, history, emotions, actions, reply, candidate_emotion, canidate_act, tokenizer, lm_labels=False, with_eos=True): """ Build a sequence of input from 3 segments: persona, history and last reply """ bos, eos, speaker1, speaker2, no_emotion = tokenizer.convert_tokens_to_ids(SPECIAL_TOKENS[:5]) inform = tokenizer.convert_tokens_to_ids(SPECIAL_TOKENS[-4]) emotions = [no_emotion] + emotions actions = [inform] + actions instance = {} sequence = [[bos] + [topic]] + history + [reply + ([eos] if with_eos else [])] sequence = [[speaker2 if (len(sequence) - i) % 2 else speaker1] + s for i, s in enumerate(sequence)] instance["input_ids"] = list(chain(*sequence)) instance["token_type_ids"] = [speaker2 if i % 2 else speaker1 for i, s in enumerate(sequence) for _ in s] # the last for is for repeating the speaker1 and speaker2 for all tokens instance["token_emotion_ids"] = [emotions[i] for i, s in enumerate(sequence[:-1]) for _ in s] + [ candidate_emotion] * len(sequence[-1]) instance["token_action_ids"] = [actions[i] for i, s in enumerate(sequence[:-1]) for _ in s] + [canidate_act] * len( sequence[-1]) instance["ec_token_ids"] = len(instance["input_ids"]) - 1 instance["sc_token_ids"] = len(instance["input_ids"]) - 2 instance["ec_labels"] = -1 instance["lm_labels"] = [-1] * len(instance["input_ids"]) if lm_labels: instance["lm_labels"] = ([-1] * sum(len(s) for s in sequence[:-1])) + [-1] + sequence[-1][ 1:] # all -1 except for reply, reply is just the ids instance["ec_labels"] = get_emotion_label(tokenizer, candidate_emotion) return instance, sequence def get_data_loaders(config, tokenizer): """ Prepare the dataset for training and evaluation """ personachat = get_dataset_for_daily_dialog(tokenizer, config.dataset_path, config.dataset_cache, SPECIAL_TOKENS) logger.info("Build inputs and labels") datasets = {"train": defaultdict(list), "valid": defaultdict(list)} gpu_max_length = 310 for dataset_name, dataset in personachat.items(): num_candidates = len(dataset[0]["utterances"][0]["candidates"]) if config.num_candidates > 0 and dataset_name == 'train': num_candidates = min(config.num_candidates, num_candidates) for dialog in dataset: topic = dialog["topic"] for utterance in dialog["utterances"]: history = utterance["history"][-(2 * config.max_history + 1):] emotions = utterance["emotion"][-(2 * config.max_history + 1):] actions = utterance["act"][-(2 * config.max_history + 1):] for j, candidate in enumerate(utterance["candidates"][-num_candidates:]): lm_labels = bool( j == num_candidates - 1) # the true label is always the last one in list of candidates candidate_emotion = utterance['candidates_emotions'][j] candidate_act = utterance['candidates_acts'][j] instance, _ = build_input_from_segments(topic, history, emotions, actions, candidate, candidate_emotion, candidate_act, tokenizer, lm_labels) if len(instance["input_ids"]) > gpu_max_length: truncated_history = [hist[:10] for hist in history] truncated_candidate = candidate[:10] instance, _ = build_input_from_segments(topic, truncated_history, emotions, actions, truncated_candidate, candidate_emotion, candidate_act, tokenizer, lm_labels) for input_name, input_array in instance.items(): datasets[dataset_name][input_name].append(input_array) datasets[dataset_name]["sc_labels"].append(num_candidates - 1) datasets[dataset_name]["n_candidates"] = num_candidates logger.info("Pad inputs and convert to Tensor") tensor_datasets = {"train": [], "valid": []} for dataset_name, dataset in datasets.items(): dataset = pad_dataset(dataset, padding=tokenizer.convert_tokens_to_ids(SPECIAL_TOKENS[-1])) for input_name in MODEL_INPUTS: tensor = torch.tensor(dataset[input_name]) if input_name != "sc_labels": tensor = tensor.view((-1, datasets[dataset_name]["n_candidates"]) + tensor.shape[1:]) tensor_datasets[dataset_name].append(tensor) logger.info("Build train and validation dataloaders") train_dataset, valid_dataset = TensorDataset(*tensor_datasets["train"]), TensorDataset(*tensor_datasets["valid"]) train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset) if config.distributed else None valid_sampler = torch.utils.data.distributed.DistributedSampler(valid_dataset) if config.distributed else None train_loader = DataLoader(train_dataset, sampler=train_sampler, batch_size=config.train_batch_size, shuffle=False) valid_loader = DataLoader(valid_dataset, sampler=valid_sampler, batch_size=config.valid_batch_size, shuffle=False) logger.info("Train dataset (Batch, Candidates, Seq length): {}".format(train_dataset.tensors[0].shape)) logger.info("Valid dataset (Batch, Candidates, Seq length): {}".format(valid_dataset.tensors[0].shape)) return train_loader, valid_loader, train_sampler, valid_sampler def train(): config_file = "configs/train_multihead_config.json" config = Config.from_json_file(config_file) ec_coef = 1 sc_coef = 1 # logging is set to INFO (resp. WARN) for main (resp. auxiliary) process. logger.info => log main process only, logger.warning => log all processes logging.basicConfig(level=logging.INFO if config.local_rank in [-1, 0] else logging.WARN) logger.warning("Running process %d", config.local_rank) # This is a logger.warning: it will be printed by all distributed processes logger.info("Arguments: %s", pformat(config)) # Initialize distributed training if needed config.distributed = (config.local_rank != -1) if config.distributed: torch.cuda.set_device(config.local_rank) config.device = torch.device("cuda", config.local_rank) torch.distributed.init_process_group(backend='nccl', init_method='env://') logger.info("Prepare tokenizer, pretrained model and optimizer - add special tokens for fine-tuning") tokenizer_class = OpenAIGPTTokenizer tokenizer = tokenizer_class.from_pretrained(config.model_checkpoint) model_class = OpenAIGPTMultiHeadModel model = model_class.from_pretrained(config.model_checkpoint) tokenizer.set_special_tokens(SPECIAL_TOKENS) model.set_num_special_tokens(len(SPECIAL_TOKENS)) model.to(config.device) optimizer = OpenAIAdam(model.parameters(), lr=config.lr) # Prepare model for FP16 and distributed training if needed (order is important, distributed should be the last) if config.fp16: from apex import amp # Apex is only required if we use fp16 training model, optimizer = amp.initialize(model, optimizer, opt_level=config.fp16) if config.distributed: model = DistributedDataParallel(model, device_ids=[config.local_rank], output_device=config.local_rank) logger.info("Prepare datasets") train_loader, val_loader, train_sampler, valid_sampler = get_data_loaders(config, tokenizer) # Training function and trainer def update(engine, batch): model.train() # input_ids, mc_token_ids, lm_labels, mc_labels, token_type_ids, token_emotion_ids, token_action_ids = tuple(input_tensor.to(config.device) for input_tensor in batch) input_ids, ec_token_ids, sc_token_ids, lm_labels, ec_labels, sc_labels, token_type_ids, token_emotion_ids, token_action_ids = tuple( input_tensor.to(config.device) for input_tensor in batch) lm_loss, emotion_loss, sentence_loss = model(input_ids, ec_token_ids, sc_token_ids, lm_labels, ec_labels, sc_labels, token_type_ids, token_emotion_ids, token_action_ids) loss = (lm_loss * config.lm_coef + emotion_loss * ec_coef + sentence_loss * sc_coef) / config.gradient_accumulation_steps if config.fp16: with amp.scale_loss(loss, optimizer) as scaled_loss: scaled_loss.backward() torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), config.max_norm) else: loss.backward() torch.nn.utils.clip_grad_norm_(model.parameters(), config.max_norm) if engine.state.iteration % config.gradient_accumulation_steps == 0: optimizer.step() optimizer.zero_grad() return loss.item() trainer = Engine(update) # Evaluation function and evaluator (evaluator output is the input of the metrics) def inference(engine, batch): model.eval() with torch.no_grad(): batch = tuple(input_tensor.to(config.device) for input_tensor in batch) input_ids, ec_token_ids, sc_token_ids, lm_labels, ec_labels, \ sc_labels, token_type_ids, token_emotion_ids, token_action_ids = batch # logger.info(tokenizer.decode(input_ids[0, -1, :].tolist())) model_outputs = model(input_ids, ec_token_ids, sc_token_ids, token_type_ids=token_type_ids, token_emotion_ids=token_emotion_ids, token_action_ids=token_action_ids) lm_logits, mc_logits = model_outputs[0], model_outputs[2] # So we can also use GPT2 outputs lm_logits_flat_shifted = lm_logits[..., :-1, :].contiguous().view(-1, lm_logits.size(-1)) lm_labels_flat_shifted = lm_labels[..., 1:].contiguous().view(-1) return (lm_logits_flat_shifted, mc_logits), (lm_labels_flat_shifted, sc_labels) evaluator = Engine(inference) # Attach evaluation to trainer: we evaluate when we start the training and at the end of each epoch trainer.add_event_handler(Events.EPOCH_COMPLETED, lambda _: evaluator.run(val_loader)) if config.n_epochs < 1: trainer.add_event_handler(Events.COMPLETED, lambda _: evaluator.run(val_loader)) if config.eval_before_start: trainer.add_event_handler(Events.STARTED, lambda _: evaluator.run(val_loader)) # Make sure distributed data samplers split the dataset nicely between the distributed processes if config.distributed: trainer.add_event_handler(Events.EPOCH_STARTED, lambda engine: train_sampler.set_epoch(engine.state.epoch)) evaluator.add_event_handler(Events.EPOCH_STARTED, lambda engine: valid_sampler.set_epoch(engine.state.epoch)) # Linearly decrease the learning rate from lr to zero scheduler = PiecewiseLinear(optimizer, "lr", [(0, config.lr), (config.n_epochs * len(train_loader), 0.0)]) trainer.add_event_handler(Events.ITERATION_STARTED, scheduler) # Prepare metrics - note how we compute distributed metrics RunningAverage(output_transform=lambda x: x).attach(trainer, "loss") metrics = {"nll": Loss(torch.nn.CrossEntropyLoss(ignore_index=-1), output_transform=lambda x: (x[0][0], x[1][0])), "accuracy": Accuracy(output_transform=lambda x: (x[0][1], x[1][1]))} metrics.update({"average_nll": MetricsLambda(average_distributed_scalar, metrics["nll"], config), "average_accuracy": MetricsLambda(average_distributed_scalar, metrics["accuracy"], config)}) metrics["average_ppl"] = MetricsLambda(math.exp, metrics["average_nll"]) for name, metric in metrics.items(): metric.attach(evaluator, name) # On the main process: add progress bar, tensorboard, checkpoints and save model, configuration and tokenizer before we start to train if config.local_rank in [-1, 0]: pbar = ProgressBar(persist=True) pbar.attach(trainer, metric_names=["loss"]) evaluator.add_event_handler(Events.COMPLETED, lambda _: pbar.log_message("Validation: %s" % pformat(evaluator.state.metrics))) tb_logger = TensorboardLogger(log_dir=config.log_dir) tb_logger.attach(trainer, log_handler=OutputHandler(tag="training", metric_names=["loss"]), event_name=Events.ITERATION_COMPLETED) tb_logger.attach(trainer, log_handler=OptimizerParamsHandler(optimizer), event_name=Events.ITERATION_STARTED) tb_logger.attach(evaluator, log_handler=OutputHandler(tag="validation", metric_names=list(metrics.keys()), another_engine=trainer), event_name=Events.EPOCH_COMPLETED) checkpoint_handler = ModelCheckpoint(tb_logger.writer.log_dir, 'checkpoint', save_interval=1, n_saved=3) trainer.add_event_handler(Events.EPOCH_COMPLETED, checkpoint_handler, { 'mymodel': getattr(model, 'module', model)}) # "getattr" take care of distributed encapsulation torch.save(config, tb_logger.writer.log_dir + '/model_training_args.bin') getattr(model, 'module', model).config.to_json_file(os.path.join(tb_logger.writer.log_dir, CONFIG_NAME)) tokenizer.save_vocabulary(tb_logger.writer.log_dir) # Run the training trainer.run(train_loader, max_epochs=config.n_epochs) # On the main process: close tensorboard logger and rename the last checkpoint (for easy re-loading with OpenAIGPTModel.from_pretrained method) if config.local_rank in [-1, 0] and config.n_epochs > 0: os.rename(checkpoint_handler._saved[-1][1][-1], os.path.join(tb_logger.writer.log_dir, WEIGHTS_NAME)) # TODO: PR in ignite to have better access to saved file paths (cleaner) tb_logger.close() if __name__ == "__main__": train()

py

EmpTransfo

EmpTransfo-master/eval_emotion_recognition.py

# Copyright (c) 2019-present, HuggingFace Inc. # All rights reserved. This source code is licensed under the BSD-style license found in the LICENSE file in the root directory of this source tree. import logging from pprint import pformat from collections import defaultdict from itertools import chain import torch from torch.nn.parallel import DistributedDataParallel from torch.utils.data import DataLoader, TensorDataset from config import Config from pytorch_pretrained_bert import (OpenAIAdam, OpenAIGPTDoubleHeadLMEmotionRecognitionModel, OpenAIGPTTokenizer, GPT2DoubleHeadsModel, GPT2Tokenizer, WEIGHTS_NAME, CONFIG_NAME, BertModel, BertTokenizer) from utils import get_dataset, get_dataset_for_daily_dialog SPECIAL_TOKENS = ["<bos>", "<eos>", "<speaker1>", "<speaker2>", "<no_emotion>", "<happiness>", "<surprise>", "<sadness>", "<disgust>", "<anger>", "<fear>", "<directive>", "<inform>", "<commissive>", "<question>", "<pad>"] MODEL_INPUTS = ["input_ids", "mc_token_ids", "lm_labels", "mc_labels", "token_type_ids", "token_emotion_ids"] PADDED_INPUTS = ["input_ids", "lm_labels", "token_type_ids", "token_emotion_ids"] logger = logging.getLogger(__file__) def average_distributed_scalar(scalar, config): """ Average a scalar over the nodes if we are in distributed training. We use this for distributed evaluation. """ if config.local_rank == -1: return scalar scalar_t = torch.tensor(scalar, dtype=torch.float, device=config.device) / torch.distributed.get_world_size() torch.distributed.all_reduce(scalar_t, op=torch.distributed.ReduceOp.SUM) return scalar_t.item() def pad_dataset(dataset, padding=0): """ Pad the dataset. This could be optimized by defining a Dataset class and padd only batches but this is simpler. """ max_l = max(len(x) for x in dataset["input_ids"]) for name in PADDED_INPUTS: dataset[name] = [x + [padding if name != "lm_labels" else -1] * (max_l - len(x)) for x in dataset[name]] return dataset def get_emotion_label(tokenizer, candidate_emotion): _, _, _, _, no_emotion_id, happiness_id, surprise_id, sadness_id, disgust_id, anger_id, fear_id, _, _, _, _, _ = tokenizer.convert_tokens_to_ids(SPECIAL_TOKENS) if candidate_emotion == happiness_id: return 0 elif candidate_emotion == surprise_id: return 1 elif candidate_emotion == sadness_id: return 2 elif candidate_emotion == disgust_id: return 3 elif candidate_emotion == anger_id: return 4 elif candidate_emotion == fear_id: return 5 elif candidate_emotion == no_emotion_id: return 6 def build_input_from_segments(history, emotions, reply, true_emotion, tokenizer, with_eos=True): """ Build a sequence of input from 3 segments: persona, history and last reply """ bos, eos, speaker1, speaker2 = tokenizer.convert_tokens_to_ids(SPECIAL_TOKENS[:4]) #tokenizer.convert_tokens_to_ids(SPECIAL_TOKENS[-1]) instance = {} # sequence = [[bos] + history[0] + list(chain(*history[1:]))] + [reply + ([eos] if with_eos else [])] #seq = [personas, history, reply] concatenate all persona sentences sequence = [[bos] + history[0]] + history[1:] + [reply + ([eos] if with_eos else [])] sequence = [[speaker2 if (len(sequence)-i) % 2 else speaker1] + s for i, s in enumerate(sequence)] instance["input_ids"] = list(chain(*sequence)) instance["token_type_ids"] = [speaker2 if i % 2 else speaker1 for i, s in enumerate(sequence) for _ in s] # the last for is for repeating the speaker1 and speaker2 for all tokens #instance["token_emotion_ids"] = [emotions[i] for i, s in enumerate(sequence[:-1]) for _ in s] + [true_emotion] * len(sequence[-1]) instance["token_emotion_ids"] = [emotions[i] for i, s in enumerate(sequence[:-1]) for _ in s] instance["mc_token_ids"] = len(instance["input_ids"]) - 1 instance["mc_labels"] = get_emotion_label(tokenizer, true_emotion) instance["lm_labels"] = ([-1] * sum(len(s) for s in sequence[:-1])) + [-1] + sequence[-1][1:] #all -1 except for reply, reply is just the ids return instance, sequence def get_data_loaders(config, tokenizer): """ Prepare the dataset for training and evaluation """ personachat = get_dataset_for_daily_dialog(tokenizer, config.dataset_path, config.dataset_cache, SPECIAL_TOKENS) #personachat["train"] = personachat["train"][:100] #personachat["valid"] = personachat["valid"][:10] logger.info("Build inputs and labels") datasets = {"train": defaultdict(list), "valid": defaultdict(list)} c = 0 for dataset_name, dataset in personachat.items(): num_candidates = 2#len(dataset[0]["utterances"][0]["candidates"]) if config.num_candidates > 0 and dataset_name == 'train': num_candidates = min(config.num_candidates, num_candidates) for dialog in dataset: for utterance in dialog["utterances"]: history = utterance["history"][-(2 * config.max_history + 1):] emotions = utterance["emotion"][-(2 * config.max_history + 1):] reply = utterance["candidates"][-1] true_emotion = utterance['candidates_emotions'][-1] if true_emotion == tokenizer.convert_tokens_to_ids(SPECIAL_TOKENS)[4]: continue instance, _ = build_input_from_segments(history, emotions, reply, true_emotion, tokenizer) if len(instance["input_ids"]) > 310: truncated_history = [hist[:10] for hist in history] truncated_candidate = reply[:10] true_emotion = utterance['candidates_emotions'][-1] instance, _ = build_input_from_segments(truncated_history, emotions, truncated_candidate, true_emotion, tokenizer) c+=1 for input_name, input_array in instance.items(): datasets[dataset_name][input_name].append(input_array) #datasets[dataset_name]["mc_labels"].append(num_candidates - 1) datasets[dataset_name]["n_candidates"] = num_candidates print(c) logger.info("Pad inputs and convert to Tensor") tensor_datasets = {"train": [], "valid": []} for dataset_name, dataset in datasets.items(): dataset = pad_dataset(dataset, padding=tokenizer.convert_tokens_to_ids(SPECIAL_TOKENS[-1])) for input_name in MODEL_INPUTS: tensor = torch.tensor(dataset[input_name]) #if input_name != "mc_labels": # tensor = tensor.view((-1, datasets[dataset_name]["n_candidates"]) + tensor.shape[1:]) tensor_datasets[dataset_name].append(tensor) logger.info("Build train and validation dataloaders") train_dataset, valid_dataset = TensorDataset(*tensor_datasets["train"]), TensorDataset(*tensor_datasets["valid"]) train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset) if config.distributed else None valid_sampler = torch.utils.data.distributed.DistributedSampler(valid_dataset) if config.distributed else None train_loader = DataLoader(train_dataset, sampler=train_sampler, batch_size=config.train_batch_size, shuffle=False) valid_loader = DataLoader(valid_dataset, sampler=valid_sampler, batch_size=config.valid_batch_size, shuffle=False) logger.info("Train dataset (Batch, Candidates, Seq length): {}".format(train_dataset.tensors[0].shape)) logger.info("Valid dataset (Batch, Candidates, Seq length): {}".format(valid_dataset.tensors[0].shape)) return train_loader, valid_loader, train_sampler, valid_sampler def train(): config_file = "configs/train_full_pipeline_config.json" config = Config.from_json_file(config_file) # logging is set to INFO (resp. WARN) for main (resp. auxiliary) process. logger.info => log main process only, logger.warning => log all processes logging.basicConfig(level=logging.INFO if config.local_rank in [-1, 0] else logging.WARN) logger.warning("Running process %d", config.local_rank) # This is a logger.warning: it will be printed by all distributed processes logger.info("Arguments: %s", pformat(config)) # Initialize distributed training if needed config.distributed = (config.local_rank != -1) if config.distributed: torch.cuda.set_device(config.local_rank) config.device = torch.device("cuda", config.local_rank) torch.distributed.init_process_group(backend='nccl', init_method='env://') logger.info("Prepare tokenizer, pretrained model and optimizer - add special tokens for fine-tuning") tokenizer_class = GPT2Tokenizer if "gpt2" in config.model_checkpoint else OpenAIGPTTokenizer tokenizer = tokenizer_class.from_pretrained(config.model_checkpoint) model_class = GPT2DoubleHeadsModel if "gpt2" in config.model_checkpoint else OpenAIGPTDoubleHeadLMEmotionRecognitionModel model = model_class.from_pretrained(config.model_checkpoint) tokenizer.set_special_tokens(SPECIAL_TOKENS) model.set_num_special_tokens(len(SPECIAL_TOKENS)) model.to(config.device) optimizer = OpenAIAdam(model.parameters(), lr=config.lr) # Prepare model for FP16 and distributed training if needed (order is important, distributed should be the last) if config.fp16: from apex import amp # Apex is only required if we use fp16 training model, optimizer = amp.initialize(model, optimizer, opt_level=config.fp16) if config.distributed: model = DistributedDataParallel(model, device_ids=[config.local_rank], output_device=config.local_rank) logger.info("Prepare datasets") train_loader, val_loader, train_sampler, valid_sampler = get_data_loaders(config, tokenizer) # Evaluation function and evaluator (evaluator output is the input of the metrics) model.eval() num_correct = 0 num_all = len(val_loader) for batch in val_loader: with torch.no_grad(): batch = tuple(input_tensor.to(config.device) for input_tensor in batch) input_ids, mc_token_ids, lm_labels, mc_labels, token_type_ids, token_emotion_ids = batch model_outputs = model(input_ids, mc_token_ids, token_type_ids=token_type_ids, token_emotion_ids=token_emotion_ids) lm_logits, mc_logits = model_outputs[0], model_outputs[1] # So we can also use GPT2 outputs indices = torch.argmax(mc_logits, dim=1) correct = torch.eq(indices, mc_labels).view(-1) num_correct += torch.sum(correct).item() print(num_correct / num_all) if __name__ == "__main__": train()

py