clean repo

.gitattributes

@@ -21,3 +21,5 @@ wit_data_dir/test/test.tsv filter=lfs diff=lfs merge=lfs -text
 train.json filter=lfs diff=lfs merge=lfs -text
 val.json filter=lfs diff=lfs merge=lfs -text
 outputs/ckpt_5/flax_model.msgpack filter=lfs diff=lfs merge=lfs -text
+C:/Users/33611/Desktop/hub/vit-gpt2/coco_data/train.json filter=lfs diff=lfs merge=lfs -text
+C:/Users/33611/Desktop/hub/vit-gpt2/coco_data/val.json filter=lfs diff=lfs merge=lfs -text

outputs-wit/.gitattributes

@@ -1 +0,0 @@
-events.out.tfevents.1626423408.t1v-n-cab111a8-w-0.820839.3.v2 filter=lfs diff=lfs merge=lfs -text

outputs-wit/ckpt_7/.gitattributes

@@ -1,2 +0,0 @@
-flax_model.msgpack filter=lfs diff=lfs merge=lfs -text
-config.json filter=lfs diff=lfs merge=lfs -text

outputs-wit/ckpt_7/config.json

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:162da9f3eab4459a15a6ae25e50ddaa379938891fb788ae8a395c7de0977e81d
-size 4229

outputs-wit/ckpt_7/flax_model.msgpack

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:64476c6a5afee2acfefaa36f1eb14a97d13999e787cdfb2fc361e7df4acbf562
-size 1012706583

outputs-wit/events.out.tfevents.1626423408.t1v-n-cab111a8-w-0.820839.3.v2

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:cc8d3f9c17dc59b0c63eb82d776844477760da16269693db8d13985d95516206
-size 476565

outputs-wit/summary.txt

@@ -1,21 +0,0 @@
-Epoch... (1/20 | Loss: 2.377821683883667, Learning Rate: 9.501103704678826e-06)
-Epoch... (1/20 | Eval Loss: 2.2912707328796387 | Eval rouge1: 17.1532 | Eval rouge2: 2.1991 | Eval rougeL: 12.1665 | Eval rougeLsum: 13.7971 | Eval gen_len: 58.3398 |)
-Predict Loss: 2.3000617027282715 | Predict rouge1: 17.3228 | Predict rouge2: 2.1974 | Predict rougeL: 12.2257 | Predict rougeLsum: 13.863 | Predict gen_len: 58.0887 |)
-Epoch... (2/20 | Loss: 2.318763017654419, Learning Rate: 9.001103535410948e-06)
-Epoch... (2/20 | Eval Loss: 2.2495603561401367 | Eval rouge1: 13.6938 | Eval rouge2: 0.963 | Eval rougeL: 10.0782 | Eval rougeLsum: 10.8405 | Eval gen_len: 58.3382 |)
-Predict Loss: 2.2592482566833496 | Predict rouge1: 13.7749 | Predict rouge2: 0.9371 | Predict rougeL: 10.0138 | Predict rougeLsum: 10.8695 | Predict gen_len: 58.1836 |)
-Epoch... (3/20 | Loss: 2.3419060707092285, Learning Rate: 8.501104275637772e-06)
-Epoch... (3/20 | Eval Loss: 2.22269344329834 | Eval rouge1: 12.0579 | Eval rouge2: 0.7251 | Eval rougeL: 9.092 | Eval rougeLsum: 9.3802 | Eval gen_len: 60.7578 |)
-Predict Loss: 2.233069896697998 | Predict rouge1: 12.5721 | Predict rouge2: 0.8881 | Predict rougeL: 9.4823 | Predict rougeLsum: 9.7638 | Predict gen_len: 60.5006 |)
-Epoch... (4/20 | Loss: 2.2800769805908203, Learning Rate: 8.001104106369894e-06)
-Epoch... (4/20 | Eval Loss: 2.2039794921875 | Eval rouge1: 14.2541 | Eval rouge2: 0.7585 | Eval rougeL: 10.3604 | Eval rougeLsum: 11.1679 | Eval gen_len: 60.3655 |)
-Predict Loss: 2.214798927307129 | Predict rouge1: 14.4009 | Predict rouge2: 0.8344 | Predict rougeL: 10.3895 | Predict rougeLsum: 11.2357 | Predict gen_len: 60.2483 |)
-Epoch... (5/20 | Loss: 2.220062494277954, Learning Rate: 7.501103482354665e-06)
-Epoch... (5/20 | Eval Loss: 2.1913952827453613 | Eval rouge1: 14.1698 | Eval rouge2: 0.8184 | Eval rougeL: 10.2918 | Eval rougeLsum: 11.245 | Eval gen_len: 60.1311 |)
-Predict Loss: 2.202223300933838 | Predict rouge1: 14.4567 | Predict rouge2: 0.9169 | Predict rougeL: 10.5117 | Predict rougeLsum: 11.3823 | Predict gen_len: 59.875 |)
-Epoch... (6/20 | Loss: 2.105752944946289, Learning Rate: 7.001103767834138e-06)
-Epoch... (6/20 | Eval Loss: 2.1800718307495117 | Eval rouge1: 14.6613 | Eval rouge2: 0.924 | Eval rougeL: 10.5021 | Eval rougeLsum: 11.672 | Eval gen_len: 61.7065 |)
-Predict Loss: 2.1911959648132324 | Predict rouge1: 14.972 | Predict rouge2: 0.9993 | Predict rougeL: 10.7166 | Predict rougeLsum: 11.843 | Predict gen_len: 61.8092 |)
-Epoch... (7/20 | Loss: 2.1191587448120117, Learning Rate: 6.50110359856626e-06)
-Epoch... (7/20 | Eval Loss: 2.1725244522094727 | Eval rouge1: 12.9676 | Eval rouge2: 1.1282 | Eval rougeL: 9.5649 | Eval rougeLsum: 10.702 | Eval gen_len: 59.4275 |)
-Predict Loss: 2.1837007999420166 | Predict rouge1: 13.161 | Predict rouge2: 1.1852 | Predict rougeL: 9.7344 | Predict rougeLsum: 10.9045 | Predict gen_len: 59.3945 |)

run_summarization.py

@@ -1,832 +0,0 @@
-#!/usr/bin/env python
-# coding=utf-8
-# Copyright 2021 The HuggingFace Team All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""
-Fine-tuning the library models for summarization.
-"""
-# You can also adapt this script on your own sequence to sequence task. Pointers for this are left as comments.
-
-import sys, os
-
-current_path = os.path.dirname(os.path.abspath(__file__))
-sys.path.append(current_path)
-
-import logging
-import os
-import sys
-import time
-from dataclasses import dataclass, field
-from functools import partial
-from pathlib import Path
-from typing import Callable, Optional
-
-import datasets
-import nltk  # Here to have a nice missing dependency error message early on
-import numpy as np
-from datasets import Dataset, load_dataset, load_metric
-from tqdm import tqdm
-
-import jax
-import jax.numpy as jnp
-import optax
-import transformers
-from filelock import FileLock
-from flax import jax_utils, traverse_util
-from flax.jax_utils import unreplicate
-from flax.training import train_state
-from flax.training.common_utils import get_metrics, onehot, shard, shard_prng_key
-from transformers import (
-    CONFIG_MAPPING,
-    FLAX_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING,
-    AutoConfig,
-    AutoTokenizer,
-    FlaxAutoModelForSeq2SeqLM,
-    HfArgumentParser,
-    TrainingArguments,
-    is_tensorboard_available,
-)
-from transformers.file_utils import is_offline_mode
-
-from transformers import ViTFeatureExtractor, GPT2Tokenizer, GPT2Config
-from vit_gpt2.modeling_flax_vit_gpt2_lm import FlaxViTGPT2LMForConditionalGeneration
-
-logger = logging.getLogger(__name__)
-
-try:
-    nltk.data.find("tokenizers/punkt")
-except (LookupError, OSError):
-    if is_offline_mode():
-        raise LookupError(
-            "Offline mode: run this script without TRANSFORMERS_OFFLINE first to download nltk data files"
-        )
-    with FileLock(".lock") as lock:
-        nltk.download("punkt", quiet=True)
-
-
-MODEL_CONFIG_CLASSES = list(FLAX_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING.keys())
-MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
-
-
-@dataclass
-class ModelArguments:
-    """
-    Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
-    """
-
-    model_name_or_path: Optional[str] = field(
-        default=None,
-        metadata={
-            "help": "The model checkpoint for weights initialization."
-            "Don't set if you want to train a model from scratch."
-        },
-    )
-    model_type: Optional[str] = field(
-        default=None,
-        metadata={"help": "If training from scratch, pass a model type from the list: " + ", ".join(MODEL_TYPES)},
-    )
-    config_name: Optional[str] = field(
-        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
-    )
-    tokenizer_name: Optional[str] = field(
-        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
-    )
-    cache_dir: Optional[str] = field(
-        default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
-    )
-    use_fast_tokenizer: bool = field(
-        default=True,
-        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
-    )
-    dtype: Optional[str] = field(
-        default="float32",
-        metadata={
-            "help": "Floating-point format in which the model weights should be initialized and trained. Choose one of `[float32, float16, bfloat16]`."
-        },
-    )
-
-
-@dataclass
-class DataTrainingArguments:
-    """
-    Arguments pertaining to what data we are going to input our model for training and eval.
-    """
-
-    dataset_name: Optional[str] = field(
-        default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
-    )
-    dataset_config_name: Optional[str] = field(
-        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
-    )
-    text_column: Optional[str] = field(
-        default=None,
-        metadata={"help": "The name of the column in the datasets containing the full texts (for summarization)."},
-    )
-    summary_column: Optional[str] = field(
-        default=None,
-        metadata={"help": "The name of the column in the datasets containing the summaries (for summarization)."},
-    )
-    train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
-    validation_file: Optional[str] = field(
-        default=None,
-        metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
-    )
-    max_source_length: Optional[int] = field(
-        default=1024,
-        metadata={
-            "help": "The maximum total input sequence length after tokenization. Sequences longer "
-            "than this will be truncated, sequences shorter will be padded."
-        },
-    )
-    max_target_length: Optional[int] = field(
-        default=128,
-        metadata={
-            "help": "The maximum total sequence length for target text after tokenization. Sequences longer "
-            "than this will be truncated, sequences shorter will be padded."
-        },
-    )
-    val_max_target_length: Optional[int] = field(
-        default=None,
-        metadata={
-            "help": "The maximum total sequence length for validation target text after tokenization. Sequences longer "
-            "than this will be truncated, sequences shorter will be padded. Will default to `max_target_length`."
-            "This argument is also used to override the `max_length` param of `model.generate`, which is used "
-            "during evaluation."
-        },
-    )
-    max_train_samples: Optional[int] = field(
-        default=None,
-        metadata={
-            "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
-            "value if set."
-        },
-    )
-    max_eval_samples: Optional[int] = field(
-        default=None,
-        metadata={
-            "help": "For debugging purposes or quicker training, truncate the number of evaluation examples to this "
-            "value if set."
-        },
-    )
-    max_predict_samples: Optional[int] = field(
-        default=None,
-        metadata={
-            "help": "For debugging purposes or quicker training, truncate the number of prediction examples to this "
-            "value if set."
-        },
-    )
-    preprocessing_num_workers: Optional[int] = field(
-        default=None,
-        metadata={"help": "The number of processes to use for the preprocessing."},
-    )
-    source_prefix: Optional[str] = field(
-        default=None, metadata={"help": "A prefix to add before every source text (useful for T5 models)."}
-    )
-    predict_with_generate: bool = field(
-        default=False, metadata={"help": "Whether to use generate to calculate generative metrics (ROUGE, BLEU)."}
-    )
-    num_beams: Optional[int] = field(
-        default=None,
-        metadata={
-            "help": "Number of beams to use for evaluation. This argument will be passed to `model.generate`, "
-            "which is used during evaluation."
-        },
-    )
-    overwrite_cache: bool = field(
-        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
-    )
-
-    def __post_init__(self):
-        if self.dataset_name is None and self.train_file is None and self.validation_file is None:
-            raise ValueError("Need either a dataset name or a training/validation file.")
-        else:
-            if self.train_file is not None:
-                extension = self.train_file.split(".")[-1]
-                assert extension in ["csv", "json"], "`train_file` should be a csv or a json file."
-            if self.validation_file is not None:
-                extension = self.validation_file.split(".")[-1]
-                assert extension in ["csv", "json"], "`validation_file` should be a csv or a json file."
-        if self.val_max_target_length is None:
-            self.val_max_target_length = self.max_target_length
-
-
-summarization_name_mapping = {
-    "amazon_reviews_multi": ("review_body", "review_title"),
-    "big_patent": ("description", "abstract"),
-    "cnn_dailymail": ("article", "highlights"),
-    "orange_sum": ("text", "summary"),
-    "pn_summary": ("article", "summary"),
-    "psc": ("extract_text", "summary_text"),
-    "samsum": ("dialogue", "summary"),
-    "thaisum": ("body", "summary"),
-    "xglue": ("news_body", "news_title"),
-    "xsum": ("document", "summary"),
-    "wiki_summary": ("article", "highlights"),
-}
-
-
-class TrainState(train_state.TrainState):
-    dropout_rng: jnp.ndarray
-
-    def replicate(self):
-        return jax_utils.replicate(self).replace(dropout_rng=shard_prng_key(self.dropout_rng))
-
-
-def data_loader(rng: jax.random.PRNGKey, dataset: Dataset, batch_size: int, shuffle: bool = False):
-    """
-    Returns batches of size `batch_size` from truncated `dataset`, sharded over all local devices.
-    Shuffle batches if `shuffle` is `True`.
-    """
-    steps_per_epoch = len(dataset) // batch_size
-
-    if shuffle:
-        batch_idx = jax.random.permutation(rng, len(dataset))
-    else:
-        batch_idx = jnp.arange(len(dataset))
-
-    batch_idx = batch_idx[: steps_per_epoch * batch_size]  # Skip incomplete batch.
-    batch_idx = batch_idx.reshape((steps_per_epoch, batch_size))
-
-    for idx in batch_idx:
-        batch = dataset[idx]
-        batch = {k: jnp.array(v) for k, v in batch.items()}
-
-        batch = shard(batch)
-
-        yield batch
-
-
-def write_metric(summary_writer, train_metrics, eval_metrics, train_time, step):
-    summary_writer.scalar("train_time", train_time, step)
-
-    train_metrics = get_metrics(train_metrics)
-    for key, vals in train_metrics.items():
-        tag = f"train_{key}"
-        for i, val in enumerate(vals):
-            summary_writer.scalar(tag, val, step - len(vals) + i + 1)
-
-    for metric_name, value in eval_metrics.items():
-        summary_writer.scalar(f"eval_{metric_name}", value, step)
-
-
-def create_learning_rate_fn(
-    train_ds_size: int, train_batch_size: int, num_train_epochs: int, num_warmup_steps: int, learning_rate: float
-) -> Callable[[int], jnp.array]:
-    """Returns a linear warmup, linear_decay learning rate function."""
-    steps_per_epoch = train_ds_size // train_batch_size
-    num_train_steps = steps_per_epoch * num_train_epochs
-    warmup_fn = optax.linear_schedule(init_value=0.0, end_value=learning_rate, transition_steps=num_warmup_steps)
-    decay_fn = optax.linear_schedule(
-        init_value=learning_rate, end_value=0, transition_steps=num_train_steps - num_warmup_steps
-    )
-    schedule_fn = optax.join_schedules(schedules=[warmup_fn, decay_fn], boundaries=[num_warmup_steps])
-    return schedule_fn
-
-
-def main():
-    # See all possible arguments in src/transformers/training_args.py
-    # or by passing the --help flag to this script.
-    # We now keep distinct sets of args, for a cleaner separation of concerns.
-
-    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
-    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
-        # If we pass only one argument to the script and it's the path to a json file,
-        # let's parse it to get our arguments.
-        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
-    else:
-        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
-
-    if (
-        os.path.exists(training_args.output_dir)
-        and os.listdir(training_args.output_dir)
-        and training_args.do_train
-        and not training_args.overwrite_output_dir
-    ):
-        raise ValueError(
-            f"Output directory ({training_args.output_dir}) already exists and is not empty."
-            "Use --overwrite_output_dir to overcome."
-        )
-
-    # Make one log on every process with the configuration for debugging.
-    logging.basicConfig(
-        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
-        datefmt="%m/%d/%Y %H:%M:%S",
-        level=logging.INFO,
-    )
-    # Setup logging, we only want one process per machine to log things on the screen.
-    logger.setLevel(logging.INFO if jax.process_index() == 0 else logging.ERROR)
-    if jax.process_index() == 0:
-        datasets.utils.logging.set_verbosity_warning()
-        transformers.utils.logging.set_verbosity_info()
-    else:
-        datasets.utils.logging.set_verbosity_error()
-        transformers.utils.logging.set_verbosity_error()
-
-    # Set the verbosity to info of the Transformers logger (on main process only):
-    logger.info(f"Training/evaluation parameters {training_args}")
-
-    # Get the datasets: you can either provide your own CSV/JSON training and evaluation files (see below)
-    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
-    # (the dataset will be downloaded automatically from the datasets Hub).
-    #
-    # For CSV/JSON files this script will use the first column for the full texts and the second column for the
-    # summaries (unless you specify column names for this with the `text_column` and `summary_column` arguments).
-    #
-    if data_args.dataset_name is not None:
-        # Downloading and loading a dataset from the hub.
-        dataset = load_dataset(
-            data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir, keep_in_memory=False, data_dir='./wit_data_dir/'
-        )
-    else:
-        data_files = {}
-        if data_args.train_file is not None:
-            data_files["train"] = data_args.train_file
-            extension = data_args.train_file.split(".")[-1]
-        if data_args.validation_file is not None:
-            data_files["validation"] = data_args.validation_file
-            extension = data_args.validation_file.split(".")[-1]
-        if data_args.test_file is not None:
-            data_files["test"] = data_args.test_file
-            extension = data_args.test_file.split(".")[-1]
-        dataset = load_dataset(extension, data_files=data_files, cache_dir=model_args.cache_dir)
-
-    vit_name_path = 'google/vit-base-patch16-224-in21k'
-    gpt2_name_path = 'asi/gpt-fr-cased-small'
-    
-    gpt2_config = GPT2Config.from_pretrained(gpt2_name_path)
-    gpt2_config.add_cross_attention = True
-    
-    
-    vit_gpt2_name_path = ''
-
-    feature_extractor = ViTFeatureExtractor.from_pretrained(vit_name_path)
-
-    tokenizer = GPT2Tokenizer.from_pretrained(gpt2_name_path)
-
-    if not vit_gpt2_name_path:
-        assert vit_name_path
-        assert gpt2_name_path
-        vit_gpt2_model = FlaxViTGPT2LMForConditionalGeneration.from_vit_gpt2_pretrained(
-            vit_name_path, gpt2_name_path
-        )
-    else:
-        vit_gpt2_model = FlaxViTGPT2LMForConditionalGeneration.from_pretrained(
-            vit_gpt2_name_path
-        )
-
-    model = vit_gpt2_model
-    model.config.is_encoder_decoder = True
-    model.config.decoder_start_token_id = gpt2_config.bos_token_id
-    model.config.bos_token_id = gpt2_config.bos_token_id
-    model.config.eos_token_id = gpt2_config.eos_token_id
-    model.config.pad_token_id = gpt2_config.pad_token_id
-
-    # Preprocessing the datasets.
-    # We need to tokenize inputs and targets.
-    if training_args.do_train:
-        column_names = dataset["train"].column_names
-    elif training_args.do_eval:
-        column_names = dataset["validation"].column_names
-    elif training_args.do_predict:
-        column_names = dataset["test"].column_names
-    else:
-        logger.info("There is nothing to do. Please pass `do_train`, `do_eval` and/or `do_predict`.")
-        return
-
-    image_file_column = 'image_file'
-    caption_column = 'caption'
-    pixels_file_column = 'pixels_file'
-
-    # Temporarily set max_target_length for training.
-    max_target_length = data_args.max_target_length
-
-    # In Flax, for seq2seq models we need to pass `decoder_input_ids`
-    # as the Flax models don't accept `labels`, we need to prepare the decoder_input_ids here
-    # for that dynamically import the `shift_tokens_right` function from the model file
-    model_module = __import__(vit_gpt2_model.__module__, fromlist=["shift_tokens_right"])
-    shift_tokens_right_fn = getattr(model_module, "shift_tokens_right")
-
-    # Setting padding="max_length" as we need fixed length inputs for jitted functions
-    def preprocess_function(examples):
-    
-        pixels_file = examples[pixels_file_column]
-        if not pixels_file:
-            assert examples[image_file_column]
-            _pixel_values = []
-            for y in examples[image_file_column]:
-                with Image.open(y) as image:
-                    encoder_inputs = feature_extractor(images=image, return_tensors="np")
-                    x = encoder_inputs.pixel_values
-                    _pixel_values.append(x)
-            pixel_values = np.concatenate(_pixel_values)
-        else:
-            pixel_values = np.concatenate([np.load(x) for x in pixels_file])
-
-        targets = examples[caption_column]
-
-        # Add eos_token!!
-        targets = [x + ' ' + tokenizer.eos_token for x in targets]
-
-        model_inputs = {}
-        model_inputs['pixel_values'] = pixel_values
-
-        # Setup the tokenizer for targets
-        with tokenizer.as_target_tokenizer():
-            labels = tokenizer(
-                targets, max_length=max_target_length, padding="max_length", truncation=True, return_tensors="np"
-            )
-
-        model_inputs["labels"] = labels["input_ids"]
-
-        #print(labels["input_ids"])
-        #print(gpt2_config.pad_token_id)
-        #rint(gpt2_config.bos_token_id)
-
-        decoder_input_ids = shift_tokens_right_fn(
-            jnp.array(labels["input_ids"]), gpt2_config.pad_token_id, gpt2_config.bos_token_id
-        )
-        model_inputs["input_ids"] = np.asarray(decoder_input_ids)
-
-        # We need decoder_attention_mask so we can ignore pad tokens from loss
-        model_inputs["attention_mask"] = labels["attention_mask"]
-
-        return model_inputs
-
-    if training_args.do_train:
-        if "train" not in dataset:
-            raise ValueError("--do_train requires a train dataset")
-        train_dataset = dataset["train"]
-        if data_args.max_train_samples is not None:
-            train_dataset = train_dataset.select(range(data_args.max_train_samples))
-            
-        train_dataset = train_dataset.map(
-            preprocess_function,
-            batched=True,
-            num_proc=data_args.preprocessing_num_workers,
-            remove_columns=column_names,
-            load_from_cache_file=not data_args.overwrite_cache,
-            desc="Running tokenizer on train dataset",
-        )
-
-    if training_args.do_eval:
-        max_target_length = data_args.val_max_target_length
-        if "validation" not in dataset:
-            raise ValueError("--do_eval requires a validation dataset")
-        eval_dataset = dataset["validation"]
-        if data_args.max_eval_samples is not None:
-            eval_dataset = eval_dataset.select(range(data_args.max_eval_samples))
-        eval_dataset = eval_dataset.map(
-            preprocess_function,
-            batched=True,
-            num_proc=data_args.preprocessing_num_workers,
-            remove_columns=column_names,
-            load_from_cache_file=not data_args.overwrite_cache,
-            desc="Running tokenizer on validation dataset",
-        )
-
-    if training_args.do_predict:
-        max_target_length = data_args.val_max_target_length
-        if "test" not in dataset:
-            raise ValueError("--do_predict requires a test dataset")
-        predict_dataset = dataset["test"]
-        if data_args.max_predict_samples is not None:
-            predict_dataset = predict_dataset.select(range(data_args.max_predict_samples))
-        predict_dataset = predict_dataset.map(
-            preprocess_function,
-            batched=True,
-            num_proc=data_args.preprocessing_num_workers,
-            remove_columns=column_names,
-            load_from_cache_file=not data_args.overwrite_cache,
-            desc="Running tokenizer on prediction dataset",
-        )
-
-    # Metric
-    metric = load_metric("rouge")
-
-    def postprocess_text(preds, labels):
-        preds = [pred.strip() for pred in preds]
-        labels = [label.strip() for label in labels]
-
-        # rougeLSum expects newline after each sentence
-        preds = ["\n".join(nltk.sent_tokenize(pred)) for pred in preds]
-        labels = ["\n".join(nltk.sent_tokenize(label)) for label in labels]
-
-        return preds, labels
-
-    def compute_metrics(preds, labels):
-        decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
-        decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
-
-        # Some simple post-processing
-        decoded_preds, decoded_labels = postprocess_text(decoded_preds, decoded_labels)
-
-        result = metric.compute(predictions=decoded_preds, references=decoded_labels, use_stemmer=True)
-        # Extract a few results from ROUGE
-        result = {key: value.mid.fmeasure * 100 for key, value in result.items()}
-
-        prediction_lens = [np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds]
-        result["gen_len"] = np.mean(prediction_lens)
-        result = {k: round(v, 4) for k, v in result.items()}
-        return result
-
-    # Enable tensorboard only on the master node
-    has_tensorboard = is_tensorboard_available()
-    if has_tensorboard and jax.process_index() == 0:
-        try:
-            from flax.metrics.tensorboard import SummaryWriter
-
-            summary_writer = SummaryWriter(log_dir=Path(training_args.output_dir))
-        except ImportError as ie:
-            has_tensorboard = False
-            logger.warning(
-                f"Unable to display metrics through TensorBoard because some package are not installed: {ie}"
-            )
-    else:
-        logger.warning(
-            "Unable to display metrics through TensorBoard because the package is not installed: "
-            "Please run pip install tensorboard to enable."
-        )
-
-    # Initialize our training
-    rng = jax.random.PRNGKey(training_args.seed)
-    rng, dropout_rng = jax.random.split(rng)
-
-    # Store some constant
-    num_epochs = int(training_args.num_train_epochs)
-    train_batch_size = int(training_args.per_device_train_batch_size) * jax.device_count()
-    eval_batch_size = int(training_args.per_device_eval_batch_size) * jax.device_count()
-    steps_per_epoch = len(train_dataset) // train_batch_size
-    total_train_steps = steps_per_epoch * num_epochs
-
-    # Create learning rate schedule
-    linear_decay_lr_schedule_fn = create_learning_rate_fn(
-        len(train_dataset),
-        train_batch_size,
-        training_args.num_train_epochs,
-        training_args.warmup_steps,
-        training_args.learning_rate,
-    )
-
-    # We use Optax's "masking" functionality to not apply weight decay
-    # to bias and LayerNorm scale parameters. decay_mask_fn returns a
-    # mask boolean with the same structure as the parameters.
-    # The mask is True for parameters that should be decayed.
-    # Note that this mask is specifically adapted for FlaxBart.
-    # For FlaxT5, one should correct the layer norm parameter naming
-    # accordingly - see `run_t5_mlm_flax.py` e.g.
-    def decay_mask_fn(params):
-        flat_params = traverse_util.flatten_dict(params)
-        layer_norm_params = [
-            (name, "scale") for name in ["self_attn_layer_norm", "layernorm_embedding", "final_layer_norm"]
-        ]
-        flat_mask = {path: (path[-1] != "bias" and path[-2:] not in layer_norm_params) for path in flat_params}
-        return traverse_util.unflatten_dict(flat_mask)
-
-    # create adam optimizer
-    adamw = optax.adamw(
-        learning_rate=linear_decay_lr_schedule_fn,
-        b1=training_args.adam_beta1,
-        b2=training_args.adam_beta2,
-        eps=training_args.adam_epsilon,
-        weight_decay=training_args.weight_decay,
-        mask=decay_mask_fn,
-    )
-
-    # Setup train state
-    state = TrainState.create(apply_fn=vit_gpt2_model.__call__, params=vit_gpt2_model.params, tx=adamw, dropout_rng=dropout_rng)
-
-    # label smoothed cross entropy
-    def loss_fn(logits, labels, padding_mask, label_smoothing_factor=0.0):
-        """
-        The label smoothing implementation is adapted from Flax's official example:
-        https://github.com/google/flax/blob/87a211135c6a377c8f29048a1cac3840e38b9da4/examples/wmt/train.py#L104
-        """
-        vocab_size = logits.shape[-1]
-        confidence = 1.0 - label_smoothing_factor
-        low_confidence = (1.0 - confidence) / (vocab_size - 1)
-        normalizing_constant = -(
-            confidence * jnp.log(confidence) + (vocab_size - 1) * low_confidence * jnp.log(low_confidence + 1e-20)
-        )
-        soft_labels = onehot(labels, vocab_size, on_value=confidence, off_value=low_confidence)
-
-        loss = optax.softmax_cross_entropy(logits, soft_labels)
-        loss = loss - normalizing_constant
-
-        # ignore padded tokens from loss
-        loss = loss * padding_mask
-        loss = loss.sum() / padding_mask.sum()
-        return loss
-
-    # Define gradient update step fn
-    def train_step(state, batch, label_smoothing_factor=0.0):
-        dropout_rng, new_dropout_rng = jax.random.split(state.dropout_rng)
-
-        def compute_loss(params):
-            labels = batch.pop("labels")
-            logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0]
-            loss = loss_fn(logits, labels, batch["attention_mask"], label_smoothing_factor)
-            return loss
-
-        grad_fn = jax.value_and_grad(compute_loss)
-        loss, grad = grad_fn(state.params)
-        grad = jax.lax.pmean(grad, "batch")
-
-        new_state = state.apply_gradients(grads=grad, dropout_rng=new_dropout_rng)
-
-        metrics = {"loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step)}
-        metrics = jax.lax.pmean(metrics, axis_name="batch")
-
-        return new_state, metrics
-
-    # Define eval fn
-    def eval_step(params, batch, label_smoothing_factor=0.0):
-        labels = batch.pop("labels")
-        logits = model(**batch, params=params, train=False)[0]
-        loss = loss_fn(logits, labels, batch["attention_mask"], label_smoothing_factor)
-
-        # summarize metrics
-        metrics = {"loss": loss}
-        metrics = jax.lax.pmean(metrics, axis_name="batch")
-        return metrics
-
-    # Define generation function
-    max_length = (
-        data_args.val_max_target_length if data_args.val_max_target_length is not None else model.config.max_length
-    )
-    num_beams = data_args.num_beams if data_args.num_beams is not None else model.config.num_beams
-    gen_kwargs = {"max_length": max_length, "num_beams": num_beams}
-
-    def generate_step(params, batch):
-        model.params = params
-        # output_ids = model.generate(batch["pixel_values"], **gen_kwargs)
-
-        #encoder_outputs = model.encode(pixel_values=batch['pixel_values'])
-        #output_ids = model.generate(batch["input_ids"], attention_mask=batch["attention_mask"], encoder_outputs=encoder_outputs, **gen_kwargs)
-
-        # encoder_outputs = model.encode(pixel_values=batch['pixel_values'], params=params, train=False)
-        output_ids = model.generate(batch['pixel_values'], **gen_kwargs)
-
-
-        return output_ids.sequences
-
-    # Create parallel version of the train and eval step
-    p_train_step = jax.pmap(
-        partial(train_step, label_smoothing_factor=training_args.label_smoothing_factor), "batch", donate_argnums=(0,)
-    )
-    p_eval_step = jax.pmap(partial(eval_step, label_smoothing_factor=training_args.label_smoothing_factor), "batch")
-    p_generate_step = jax.pmap(generate_step, "batch")
-
-    # Replicate the train state on each device
-    state = state.replicate()
-
-    logger.info("***** Running training *****")
-    logger.info(f"  Num examples = {len(train_dataset)}")
-    logger.info(f"  Num Epochs = {num_epochs}")
-    logger.info(f"  Instantaneous batch size per device = {training_args.per_device_train_batch_size}")
-    logger.info(f"  Total train batch size (w. parallel & distributed) = {train_batch_size}")
-    logger.info(f"  Total optimization steps = {total_train_steps}")
-
-    train_time = 0
-    epochs = tqdm(range(num_epochs), desc=f"Epoch ... (1/{num_epochs})", position=0)
-    for epoch in epochs:
-        # ======================== Training ================================
-        train_start = time.time()
-
-        # Create sampling rng
-        rng, input_rng = jax.random.split(rng)
-        train_metrics = []
-
-        # Generate an epoch by shuffling sampling indices from the train dataset
-        train_loader = data_loader(input_rng, train_dataset, train_batch_size, shuffle=True)
-        steps_per_epoch = len(train_dataset) // train_batch_size
-        # train
-        for _ in tqdm(range(steps_per_epoch), desc="Training...", position=1, leave=False):
-            batch = next(train_loader)
-            state, train_metric = p_train_step(state, batch)
-            train_metrics.append(train_metric)
-
-        train_time += time.time() - train_start
-
-        train_metric = unreplicate(train_metric)
-
-        desc = f"Epoch... ({epoch + 1}/{num_epochs} | Loss: {train_metric['loss']}, Learning Rate: {train_metric['learning_rate']})"
-        epochs.write(desc)
-        epochs.desc = desc
-        logger.info(desc)
-        with open(os.path.join(training_args.output_dir, f'report.txt'), 'a', encoding='UTF-8') as fp:
-            fp.write(desc + '\n')
-
-
-        # ======================== Evaluating ==============================
-        eval_metrics = []
-        eval_preds = []
-        eval_labels = []
-
-        eval_loader = data_loader(input_rng, eval_dataset, eval_batch_size)
-        eval_steps = len(eval_dataset) // eval_batch_size
-        for _ in tqdm(range(eval_steps), desc="Evaluating...", position=2, leave=False):
-            # Model forward
-            batch = next(eval_loader)
-            labels = batch["labels"]
-
-            metrics = p_eval_step(state.params, batch)
-            eval_metrics.append(metrics)
-
-            # generation
-            if data_args.predict_with_generate:
-                generated_ids = p_generate_step(state.params, batch)
-                eval_preds.extend(jax.device_get(generated_ids.reshape(-1, gen_kwargs["max_length"])))
-                eval_labels.extend(jax.device_get(labels.reshape(-1, labels.shape[-1])))
-
-        # normalize eval metrics
-        eval_metrics = get_metrics(eval_metrics)
-        eval_metrics = jax.tree_map(jnp.mean, eval_metrics)
-
-        # compute ROUGE metrics
-        rouge_desc = ""
-        if data_args.predict_with_generate:
-            rouge_metrics = compute_metrics(eval_preds, eval_labels)
-            eval_metrics.update(rouge_metrics)
-            rouge_desc = " ".join([f"Eval {key}: {value} |" for key, value in rouge_metrics.items()])
-
-        # Print metrics and update progress bar
-        desc = f"Epoch... ({epoch + 1}/{num_epochs} | Eval Loss: {eval_metrics['loss']} | {rouge_desc})"
-        epochs.write(desc)
-        epochs.desc = desc
-        logger.info(desc)
-        with open(os.path.join(training_args.output_dir, f'report.txt'), 'a', encoding='UTF-8') as fp:
-            fp.write(desc + '\n')
-
-
-        # Save metrics
-        if has_tensorboard and jax.process_index() == 0:
-            cur_step = epoch * (len(train_dataset) // train_batch_size)
-            write_metric(summary_writer, train_metrics, eval_metrics, train_time, cur_step)
-
-        # ======================== Prediction loop ==============================
-        if training_args.do_predict:
-            logger.info("*** Predict ***")
-    
-            pred_metrics = []
-            pred_generations = []
-            pred_labels = []
-    
-            pred_loader = data_loader(input_rng, predict_dataset, eval_batch_size)
-            pred_steps = len(predict_dataset) // eval_batch_size
-            for _ in tqdm(range(pred_steps), desc="Predicting...", position=2, leave=False):
-                # Model forward
-                batch = next(pred_loader)
-                labels = batch["labels"]
-    
-                metrics = p_eval_step(state.params, batch)
-                pred_metrics.append(metrics)
-    
-                # generation
-                if data_args.predict_with_generate:
-                    generated_ids = p_generate_step(state.params, batch)
-                    pred_generations.extend(jax.device_get(generated_ids.reshape(-1, gen_kwargs["max_length"])))
-                    pred_labels.extend(jax.device_get(labels.reshape(-1, labels.shape[-1])))
-    
-            # normalize prediction metrics
-            pred_metrics = get_metrics(pred_metrics)
-            pred_metrics = jax.tree_map(jnp.mean, pred_metrics)
-    
-            # compute ROUGE metrics
-            rouge_desc = ""
-            if data_args.predict_with_generate:
-                rouge_metrics = compute_metrics(pred_generations, pred_labels)
-                pred_metrics.update(rouge_metrics)
-                rouge_desc = " ".join([f"Predict {key}: {value} |" for key, value in rouge_metrics.items()])
-    
-            # Print metrics
-            desc = f"Predict Loss: {pred_metrics['loss']} | {rouge_desc})"
-            epochs.write(desc)
-            epochs.desc = desc
-            logger.info(desc)
-            with open(os.path.join(training_args.output_dir, f'report.txt'), 'a', encoding='UTF-8') as fp:
-                fp.write(desc + '\n')
-    
-
-        # save checkpoint after each epoch and push checkpoint to the hub
-        if jax.process_index() == 0:
-            params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
-            model.save_pretrained(
-                os.path.join(training_args.output_dir, f'ckpt_{epoch+1}'),
-                params=params,
-                push_to_hub=training_args.push_to_hub,
-                commit_message=f"Saving weights and logs of epoch {epoch+1}",
-            )
-
-if __name__ == "__main__":
-    main()

test_wit_dataset_script.py

@@ -1,23 +0,0 @@
-import csv
-import json
-import os
-
-import datasets
-import pandas as pd
-import numpy as np
-
-ds = datasets.load_dataset('./wit_dataset_script.py', data_dir='./wit_data_dir/')
-test_ds = ds['test']
-
-
-def transform(example):
-
-    example['pixel_values'] = np.load(example['pixels_file'])
-    return example
-
-
-test_ds = test_ds.map(transform)
-
-for x in test_ds:
-    print(x)
-    break

wit_data_dir/dev/dev.tsv

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:ef1ecdcd132885a8f29c8707fad649431c6ff3d9bbd295d56b8520e7046c0eb7
-size 1418232

wit_data_dir/test/test.tsv

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:f0517292749005808b1d1d75343c76b8b16c3ed74fde030f7af8b611ad7b4d5d
-size 1406997

wit_data_dir/train/train.tsv

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:267de5cc6965e95f44795e78019ad9d0dfc648bee83a9fdb9cf9a92e8e4ac9d3
-size 45417661

wit_dataset_script.py

@@ -1,145 +0,0 @@
-import csv
-import json
-import os
-
-import datasets
-import pandas as pd
-import numpy as np
-
-
-# TODO: Add BibTeX citation
-# Find for instance the citation on arxiv or on the dataset repo/website
-_CITATION = """\
-@InProceedings{huggingface:dataset,
-title = {A great new dataset},
-author={huggingface, Inc.
-},
-year={2020}
-}
-"""
-
-# TODO: Add description of the dataset here
-# You can copy an official description
-_DESCRIPTION = """\
-This new dataset is designed to solve this great NLP task and is crafted with a lot of care.
-"""
-
-# TODO: Add a link to an official homepage for the dataset here
-_HOMEPAGE = ""
-
-# TODO: Add the licence for the dataset here if you can find it
-_LICENSE = ""
-
-# TODO: Add link to the official dataset URLs here
-# The HuggingFace dataset library don't host the datasets but only point to the original files
-# This can be an arbitrary nested dict/list of URLs (see below in `_split_generators` method)
-_URLs = {
-}
-
-
-# TODO: Name of the dataset usually match the script name with CamelCase instead of snake_case
-class WITDataset(datasets.GeneratorBasedBuilder):
-    """TODO: Short description of my dataset."""
-
-    VERSION = datasets.Version("1.1.0")
-
-    DEFAULT_CONFIG_NAME = "en"
-
-    def _info(self):
-        # TODO: This method specifies the datasets.DatasetInfo object which contains informations and typings for the dataset
-
-        features = datasets.Features(
-            {
-                "id": datasets.Value("int64"),
-                "lang": datasets.Value("string"),
-                "caption": datasets.Value("string"),
-                "context": datasets.Value("string"),
-                "image_url": datasets.Value("string"),
-                "page_url": datasets.Value("string"),
-                "image_file": datasets.Value("string"),
-                "pixels_file": datasets.Value("string")
-                # These are the features of your dataset like images, labels ...
-            }
-        )
-
-        return datasets.DatasetInfo(
-            # This is the description that will appear on the datasets page.
-            description=_DESCRIPTION,
-            # This defines the different columns of the dataset and their types
-            features=features,  # Here we define them above because they are different between the two configurations
-            # If there's a common (input, target) tuple from the features,
-            # specify them here. They'll be used if as_supervised=True in
-            # builder.as_dataset.
-            supervised_keys=None,
-            # Homepage of the dataset for documentation
-            homepage=_HOMEPAGE,
-            # License for the dataset if available
-            license=_LICENSE,
-            # Citation for the dataset
-            citation=_CITATION,
-        )
-
-    def _split_generators(self, dl_manager):
-        """Returns SplitGenerators."""
-        # TODO: This method is tasked with downloading/extracting the data and defining the splits depending on the configuration
-        # If several configurations are possible (listed in BUILDER_CONFIGS), the configuration selected by the user is in self.config.name
-
-        data_dir = self.config.data_dir
-
-        return [
-            datasets.SplitGenerator(
-                name=datasets.Split.TRAIN,
-                # These kwargs will be passed to _generate_examples
-                gen_kwargs={
-                    "data_dir": os.path.join(data_dir, "train"),
-                    "split": "train",
-                },
-            ),
-            datasets.SplitGenerator(
-                name=datasets.Split.TEST,
-                # These kwargs will be passed to _generate_examples
-                gen_kwargs={
-                    "data_dir": os.path.join(data_dir, "test"),
-                    "split": "test"
-                },
-            ),
-            datasets.SplitGenerator(
-                name=datasets.Split.VALIDATION,
-                # These kwargs will be passed to _generate_examples
-                gen_kwargs={
-                    "data_dir": os.path.join(data_dir, "dev"),
-                    "split": "dev",
-                },
-            ),
-        ]
-
-    def _generate_examples(
-        self, data_dir, split  # method parameters are unpacked from `gen_kwargs` as given in `_split_generators`
-    ):
-        """ Yields examples as (key, example) tuples. """
-        # This method handles input defined in _split_generators to yield (key, example) tuples from the dataset.
-        # The `key` is here for legacy reason (tfds) and is not important in itself.
-
-        df = pd.read_csv(os.path.join(data_dir, f'{split}.tsv'), sep='\t')
-
-        for id_, row in df.iterrows():
-
-            _id = row[0]
-
-            # null caption and context
-            if type(row[4]) != str or type(row[5]) != str:
-                continue
-
-            image_file = os.path.join(data_dir, 'images', f'{_id}.jpg')
-            pixels_file = os.path.join(data_dir, 'numpy', f'{_id}.npy')
-
-            yield id_, {
-                "id": row[0],
-                "lang": row[1],
-                "caption": row[4],
-                "context": row[5],
-                "image_url": row[2],
-                "page_url": row[3],
-                "image_file": image_file,
-                "pixels_file": pixels_file
-            }