import matplotlib from torch.nn import DataParallel from torch.nn.parallel import DistributedDataParallel matplotlib.use('Agg') import glob import itertools import subprocess import threading import traceback from pytorch_lightning.callbacks import GradientAccumulationScheduler from pytorch_lightning.callbacks import ModelCheckpoint from functools import wraps from torch.cuda._utils import _get_device_index import numpy as np import torch.optim import torch.utils.data import copy import logging import os import re import sys import torch import torch.distributed as dist import torch.multiprocessing as mp import tqdm from torch.optim.optimizer import Optimizer def get_a_var(obj): # pragma: no cover if isinstance(obj, torch.Tensor): return obj if isinstance(obj, list) or isinstance(obj, tuple): for result in map(get_a_var, obj): if isinstance(result, torch.Tensor): return result if isinstance(obj, dict): for result in map(get_a_var, obj.items()): if isinstance(result, torch.Tensor): return result return None def data_loader(fn): """ Decorator to make any fx with this use the lazy property :param fn: :return: """ wraps(fn) attr_name = '_lazy_' + fn.__name__ def _get_data_loader(self): try: value = getattr(self, attr_name) except AttributeError: try: value = fn(self) # Lazy evaluation, done only once. if ( value is not None and not isinstance(value, list) and fn.__name__ in ['test_dataloader', 'val_dataloader'] ): value = [value] except AttributeError as e: # Guard against AttributeError suppression. (Issue #142) traceback.print_exc() error = f'{fn.__name__}: An AttributeError was encountered: ' + str(e) raise RuntimeError(error) from e setattr(self, attr_name, value) # Memoize evaluation. return value return _get_data_loader def parallel_apply(modules, inputs, kwargs_tup=None, devices=None): # pragma: no cover r"""Applies each `module` in :attr:`modules` in parallel on arguments contained in :attr:`inputs` (positional) and :attr:`kwargs_tup` (keyword) on each of :attr:`devices`. Args: modules (Module): modules to be parallelized inputs (tensor): inputs to the modules devices (list of int or torch.device): CUDA devices :attr:`modules`, :attr:`inputs`, :attr:`kwargs_tup` (if given), and :attr:`devices` (if given) should all have same length. Moreover, each element of :attr:`inputs` can either be a single object as the only argument to a module, or a collection of positional arguments. """ assert len(modules) == len(inputs) if kwargs_tup is not None: assert len(modules) == len(kwargs_tup) else: kwargs_tup = ({},) * len(modules) if devices is not None: assert len(modules) == len(devices) else: devices = [None] * len(modules) devices = list(map(lambda x: _get_device_index(x, True), devices)) lock = threading.Lock() results = {} grad_enabled = torch.is_grad_enabled() def _worker(i, module, input, kwargs, device=None): torch.set_grad_enabled(grad_enabled) if device is None: device = get_a_var(input).get_device() try: with torch.cuda.device(device): # this also avoids accidental slicing of `input` if it is a Tensor if not isinstance(input, (list, tuple)): input = (input,) # --------------- # CHANGE if module.training: output = module.training_step(*input, **kwargs) elif module.testing: output = module.test_step(*input, **kwargs) else: output = module.validation_step(*input, **kwargs) # --------------- with lock: results[i] = output except Exception as e: with lock: results[i] = e # make sure each module knows what training state it's in... # fixes weird bug where copies are out of sync root_m = modules[0] for m in modules[1:]: m.training = root_m.training m.testing = root_m.testing if len(modules) > 1: threads = [threading.Thread(target=_worker, args=(i, module, input, kwargs, device)) for i, (module, input, kwargs, device) in enumerate(zip(modules, inputs, kwargs_tup, devices))] for thread in threads: thread.start() for thread in threads: thread.join() else: _worker(0, modules[0], inputs[0], kwargs_tup[0], devices[0]) outputs = [] for i in range(len(inputs)): output = results[i] if isinstance(output, Exception): raise output outputs.append(output) return outputs def _find_tensors(obj): # pragma: no cover r""" Recursively find all tensors contained in the specified object. """ if isinstance(obj, torch.Tensor): return [obj] if isinstance(obj, (list, tuple)): return itertools.chain(*map(_find_tensors, obj)) if isinstance(obj, dict): return itertools.chain(*map(_find_tensors, obj.values())) return [] class DDP(DistributedDataParallel): """ Override the forward call in lightning so it goes to training and validation step respectively """ def parallel_apply(self, replicas, inputs, kwargs): return parallel_apply(replicas, inputs, kwargs, self.device_ids[:len(replicas)]) def forward(self, *inputs, **kwargs): # pragma: no cover self._sync_params() if self.device_ids: inputs, kwargs = self.scatter(inputs, kwargs, self.device_ids) if len(self.device_ids) == 1: # -------------- # LIGHTNING MOD # -------------- # normal # output = self.module(*inputs[0], **kwargs[0]) # lightning if self.module.training: output = self.module.training_step(*inputs[0], **kwargs[0]) elif self.module.testing: output = self.module.test_step(*inputs[0], **kwargs[0]) else: output = self.module.validation_step(*inputs[0], **kwargs[0]) else: outputs = self.parallel_apply(self._module_copies[:len(inputs)], inputs, kwargs) output = self.gather(outputs, self.output_device) else: # normal output = self.module(*inputs, **kwargs) if torch.is_grad_enabled(): # We'll return the output object verbatim since it is a freeform # object. We need to find any tensors in this object, though, # because we need to figure out which parameters were used during # this forward pass, to ensure we short circuit reduction for any # unused parameters. Only if `find_unused_parameters` is set. if self.find_unused_parameters: self.reducer.prepare_for_backward(list(_find_tensors(output))) else: self.reducer.prepare_for_backward([]) return output class DP(DataParallel): """ Override the forward call in lightning so it goes to training and validation step respectively """ def forward(self, *inputs, **kwargs): if not self.device_ids: return self.module(*inputs, **kwargs) for t in itertools.chain(self.module.parameters(), self.module.buffers()): if t.device != self.src_device_obj: raise RuntimeError("module must have its parameters and buffers " "on device {} (device_ids[0]) but found one of " "them on device: {}".format(self.src_device_obj, t.device)) inputs, kwargs = self.scatter(inputs, kwargs, self.device_ids) if len(self.device_ids) == 1: # lightning if self.module.training: return self.module.training_step(*inputs[0], **kwargs[0]) elif self.module.testing: return self.module.test_step(*inputs[0], **kwargs[0]) else: return self.module.validation_step(*inputs[0], **kwargs[0]) replicas = self.replicate(self.module, self.device_ids[:len(inputs)]) outputs = self.parallel_apply(replicas, inputs, kwargs) return self.gather(outputs, self.output_device) def parallel_apply(self, replicas, inputs, kwargs): return parallel_apply(replicas, inputs, kwargs, self.device_ids[:len(replicas)]) class GradientAccumulationScheduler: def __init__(self, scheduling: dict): if scheduling == {}: # empty dict error raise TypeError("Empty dict cannot be interpreted correct") for key in scheduling.keys(): if not isinstance(key, int) or not isinstance(scheduling[key], int): raise TypeError("All epoches and accumulation factor must be integers") minimal_epoch = min(scheduling.keys()) if minimal_epoch < 1: msg = f"Epochs indexing from 1, epoch {minimal_epoch} cannot be interpreted correct" raise IndexError(msg) elif minimal_epoch != 1: # if user didnt define first epoch accumulation factor scheduling.update({1: 1}) self.scheduling = scheduling self.epochs = sorted(scheduling.keys()) def on_epoch_begin(self, epoch, trainer): epoch += 1 # indexing epochs from 1 for i in reversed(range(len(self.epochs))): if epoch >= self.epochs[i]: trainer.accumulate_grad_batches = self.scheduling.get(self.epochs[i]) break class LatestModelCheckpoint(ModelCheckpoint): def __init__(self, filepath, monitor='val_loss', verbose=0, num_ckpt_keep=5, save_weights_only=False, mode='auto', period=1, prefix='model', save_best=True): super(ModelCheckpoint, self).__init__() self.monitor = monitor self.verbose = verbose self.filepath = filepath os.makedirs(filepath, exist_ok=True) self.num_ckpt_keep = num_ckpt_keep self.save_best = save_best self.save_weights_only = save_weights_only self.period = period self.epochs_since_last_check = 0 self.prefix = prefix self.best_k_models = {} # {filename: monitor} self.kth_best_model = '' self.save_top_k = 1 self.task = None if mode == 'min': self.monitor_op = np.less self.best = np.Inf self.mode = 'min' elif mode == 'max': self.monitor_op = np.greater self.best = -np.Inf self.mode = 'max' else: if 'acc' in self.monitor or self.monitor.startswith('fmeasure'): self.monitor_op = np.greater self.best = -np.Inf self.mode = 'max' else: self.monitor_op = np.less self.best = np.Inf self.mode = 'min' if os.path.exists(f'{self.filepath}/best_valid.npy'): self.best = np.load(f'{self.filepath}/best_valid.npy')[0] def get_all_ckpts(self): return sorted(glob.glob(f'{self.filepath}/{self.prefix}_ckpt_steps_*.ckpt'), key=lambda x: -int(re.findall('.*steps\_(\d+)\.ckpt', x)[0])) def on_epoch_end(self, epoch, logs=None): logs = logs or {} self.epochs_since_last_check += 1 best_filepath = f'{self.filepath}/{self.prefix}_ckpt_best.pt' if self.epochs_since_last_check >= self.period: self.epochs_since_last_check = 0 filepath = f'{self.filepath}/{self.prefix}_ckpt_steps_{self.task.global_step}.ckpt' if self.verbose > 0: logging.info(f'Epoch {epoch:05d}@{self.task.global_step}: saving model to {filepath}') self._save_model(filepath) for old_ckpt in self.get_all_ckpts()[self.num_ckpt_keep:]: # TODO: test filesystem calls os.remove(old_ckpt) # subprocess.check_call(f'del "{old_ckpt}"', shell=True) if self.verbose > 0: logging.info(f'Delete ckpt: {os.path.basename(old_ckpt)}') current = logs.get(self.monitor) if current is not None and self.save_best: if self.monitor_op(current, self.best): self.best = current if self.verbose > 0: logging.info( f'Epoch {epoch:05d}@{self.task.global_step}: {self.monitor} reached' f' {current:0.5f} (best {self.best:0.5f}), saving model to' f' {best_filepath} as top 1') self._save_model(best_filepath) np.save(f'{self.filepath}/best_valid.npy', [self.best]) def _save_model(self,path): return self.save_function(path) class BaseTrainer: def __init__( self, logger=True, checkpoint_callback=True, default_save_path=None, gradient_clip_val=0, process_position=0, gpus=-1, log_gpu_memory=None, show_progress_bar=True, track_grad_norm=-1, check_val_every_n_epoch=1, accumulate_grad_batches=1, max_updates=1000, min_epochs=1, val_check_interval=1.0, log_save_interval=100, row_log_interval=10, print_nan_grads=False, weights_summary='full', num_sanity_val_steps=5, resume_from_checkpoint=None, ): self.log_gpu_memory = log_gpu_memory self.gradient_clip_val = gradient_clip_val self.check_val_every_n_epoch = check_val_every_n_epoch self.track_grad_norm = track_grad_norm self.on_gpu = True if (gpus and torch.cuda.is_available()) else False self.process_position = process_position self.weights_summary = weights_summary self.max_updates = max_updates self.min_epochs = min_epochs self.num_sanity_val_steps = num_sanity_val_steps self.print_nan_grads = print_nan_grads self.resume_from_checkpoint = resume_from_checkpoint self.default_save_path = default_save_path # training bookeeping self.total_batch_idx = 0 self.running_loss = [] self.avg_loss = 0 self.batch_idx = 0 self.tqdm_metrics = {} self.callback_metrics = {} self.num_val_batches = 0 self.num_training_batches = 0 self.num_test_batches = 0 self.get_train_dataloader = None self.get_test_dataloaders = None self.get_val_dataloaders = None self.is_iterable_train_dataloader = False # training state self.model = None self.testing = False self.disable_validation = False self.lr_schedulers = [] self.optimizers = None self.global_step = 0 self.current_epoch = 0 self.total_batches = 0 # configure checkpoint callback self.checkpoint_callback = checkpoint_callback self.checkpoint_callback.save_function = self.save_checkpoint self.weights_save_path = self.checkpoint_callback.filepath # accumulated grads self.configure_accumulated_gradients(accumulate_grad_batches) # allow int, string and gpu list self.data_parallel_device_ids = [ int(x) for x in os.environ.get("CUDA_VISIBLE_DEVICES", "").split(",") if x != ''] if len(self.data_parallel_device_ids) == 0: self.root_gpu = None self.on_gpu = False else: self.root_gpu = self.data_parallel_device_ids[0] self.on_gpu = True # distributed backend choice self.use_ddp = False self.use_dp = False self.single_gpu = False self.distributed_backend = 'ddp' if self.num_gpus > 0 else 'dp' self.set_distributed_mode(self.distributed_backend) self.proc_rank = 0 self.world_size = 1 self.node_rank = 0 # can't init progress bar here because starting a new process # means the progress_bar won't survive pickling self.show_progress_bar = show_progress_bar # logging self.log_save_interval = log_save_interval self.val_check_interval = val_check_interval self.logger = logger self.logger.rank = 0 self.row_log_interval = row_log_interval @property def num_gpus(self): gpus = self.data_parallel_device_ids if gpus is None: return 0 else: return len(gpus) @property def data_parallel(self): return self.use_dp or self.use_ddp def get_model(self): is_dp_module = isinstance(self.model, (DDP, DP)) model = self.model.module if is_dp_module else self.model return model # ----------------------------- # MODEL TRAINING # ----------------------------- def fit(self, model): if self.use_ddp: mp.spawn(self.ddp_train, nprocs=self.num_gpus, args=(model,)) else: model.model = model.build_model() if not self.testing: self.optimizers, self.lr_schedulers = self.init_optimizers(model.configure_optimizers()) if self.use_dp: model.cuda(self.root_gpu) model = DP(model, device_ids=self.data_parallel_device_ids) elif self.single_gpu: model.cuda(self.root_gpu) self.run_pretrain_routine(model) return 1 def init_optimizers(self, optimizers): # single optimizer if isinstance(optimizers, Optimizer): return [optimizers], [] # two lists elif len(optimizers) == 2 and isinstance(optimizers[0], list): optimizers, lr_schedulers = optimizers return optimizers, lr_schedulers # single list or tuple elif isinstance(optimizers, list) or isinstance(optimizers, tuple): return optimizers, [] def run_pretrain_routine(self, model): """Sanity check a few things before starting actual training. :param model: """ ref_model = model if self.data_parallel: ref_model = model.module # give model convenience properties ref_model.trainer = self # set local properties on the model self.copy_trainer_model_properties(ref_model) # link up experiment object if self.logger is not None: ref_model.logger = self.logger self.logger.save() if self.use_ddp: dist.barrier() # set up checkpoint callback # self.configure_checkpoint_callback() # transfer data loaders from model self.get_dataloaders(ref_model) # track model now. # if cluster resets state, the model will update with the saved weights self.model = model # restore training and model before hpc call self.restore_weights(model) # when testing requested only run test and return if self.testing: self.run_evaluation(test=True) return # check if we should run validation during training self.disable_validation = self.num_val_batches == 0 # run tiny validation (if validation defined) # to make sure program won't crash during val ref_model.on_sanity_check_start() ref_model.on_train_start() if not self.disable_validation and self.num_sanity_val_steps > 0: # init progress bars for validation sanity check pbar = tqdm.tqdm(desc='Validation sanity check', total=self.num_sanity_val_steps * len(self.get_val_dataloaders()), leave=False, position=2 * self.process_position, disable=not self.show_progress_bar, dynamic_ncols=True, unit='batch') self.main_progress_bar = pbar # dummy validation progress bar self.val_progress_bar = tqdm.tqdm(disable=True) self.evaluate(model, self.get_val_dataloaders(), self.num_sanity_val_steps, self.testing) # close progress bars self.main_progress_bar.close() self.val_progress_bar.close() # init progress bar pbar = tqdm.tqdm(leave=True, position=2 * self.process_position, disable=not self.show_progress_bar, dynamic_ncols=True, unit='batch', file=sys.stdout) self.main_progress_bar = pbar # clear cache before training if self.on_gpu: torch.cuda.empty_cache() # CORE TRAINING LOOP self.train() def test(self, model): self.testing = True self.fit(model) @property def training_tqdm_dict(self): tqdm_dict = { 'step': '{}'.format(self.global_step), } tqdm_dict.update(self.tqdm_metrics) return tqdm_dict # -------------------- # restore ckpt # -------------------- def restore_weights(self, model): """ To restore weights we have two cases. First, attempt to restore hpc weights. If successful, don't restore other weights. Otherwise, try to restore actual weights :param model: :return: """ # clear cache before restore if self.on_gpu: torch.cuda.empty_cache() if self.resume_from_checkpoint is not None: self.restore(self.resume_from_checkpoint, on_gpu=self.on_gpu) else: # restore weights if same exp version self.restore_state_if_checkpoint_exists(model) # wait for all models to restore weights if self.use_ddp: # wait for all processes to catch up dist.barrier() # clear cache after restore if self.on_gpu: torch.cuda.empty_cache() def restore_state_if_checkpoint_exists(self, model): did_restore = False # do nothing if there's not dir or callback no_ckpt_callback = (self.checkpoint_callback is None) or (not self.checkpoint_callback) if no_ckpt_callback or not os.path.exists(self.checkpoint_callback.filepath): return did_restore # restore trainer state and model if there is a weight for this experiment last_steps = -1 last_ckpt_name = None # find last epoch checkpoints = os.listdir(self.checkpoint_callback.filepath) for name in checkpoints: if '.ckpt' in name and not name.endswith('part'): if 'steps_' in name: steps = name.split('steps_')[1] steps = int(re.sub('[^0-9]', '', steps)) if steps > last_steps: last_steps = steps last_ckpt_name = name # restore last checkpoint if last_ckpt_name is not None: last_ckpt_path = os.path.join(self.checkpoint_callback.filepath, last_ckpt_name) self.restore(last_ckpt_path, self.on_gpu) logging.info(f'model and trainer restored from checkpoint: {last_ckpt_path}') did_restore = True return did_restore def restore(self, checkpoint_path, on_gpu): checkpoint = torch.load(checkpoint_path, map_location='cpu') # load model state model = self.get_model() # load the state_dict on the model automatically model.load_state_dict(checkpoint['state_dict'], strict=False) if on_gpu: model.cuda(self.root_gpu) # load training state (affects trainer only) self.restore_training_state(checkpoint) model.global_step = self.global_step del checkpoint try: if dist.is_initialized() and dist.get_rank() > 0: return except Exception as e: print(e) return def restore_training_state(self, checkpoint): """ Restore trainer state. Model will get its change to update :param checkpoint: :return: """ if self.checkpoint_callback is not None and self.checkpoint_callback is not False: # return allowing checkpoints with meta information (global_step, etc) self.checkpoint_callback.best = checkpoint['checkpoint_callback_best'] self.global_step = checkpoint['global_step'] self.current_epoch = checkpoint['epoch'] if self.testing: return # restore the optimizers optimizer_states = checkpoint['optimizer_states'] for optimizer, opt_state in zip(self.optimizers, optimizer_states): if optimizer is None: return optimizer.load_state_dict(opt_state) # move optimizer to GPU 1 weight at a time # avoids OOM if self.root_gpu is not None: for state in optimizer.state.values(): for k, v in state.items(): if isinstance(v, torch.Tensor): state[k] = v.cuda(self.root_gpu) # restore the lr schedulers lr_schedulers = checkpoint['lr_schedulers'] for scheduler, lrs_state in zip(self.lr_schedulers, lr_schedulers): scheduler.load_state_dict(lrs_state) # -------------------- # MODEL SAVE CHECKPOINT # -------------------- def _atomic_save(self, checkpoint, filepath): """Saves a checkpoint atomically, avoiding the creation of incomplete checkpoints. This will create a temporary checkpoint with a suffix of ``.part``, then copy it to the final location once saving is finished. Args: checkpoint (object): The object to save. Built to be used with the ``dump_checkpoint`` method, but can deal with anything which ``torch.save`` accepts. filepath (str|pathlib.Path): The path to which the checkpoint will be saved. This points to the file that the checkpoint will be stored in. """ tmp_path = str(filepath) + ".part" torch.save(checkpoint, tmp_path) os.replace(tmp_path, filepath) def save_checkpoint(self, filepath): checkpoint = self.dump_checkpoint() self._atomic_save(checkpoint, filepath) def dump_checkpoint(self): checkpoint = { 'epoch': self.current_epoch, 'global_step': self.global_step } if self.checkpoint_callback is not None and self.checkpoint_callback is not False: checkpoint['checkpoint_callback_best'] = self.checkpoint_callback.best # save optimizers optimizer_states = [] for i, optimizer in enumerate(self.optimizers): if optimizer is not None: optimizer_states.append(optimizer.state_dict()) checkpoint['optimizer_states'] = optimizer_states # save lr schedulers lr_schedulers = [] for i, scheduler in enumerate(self.lr_schedulers): lr_schedulers.append(scheduler.state_dict()) checkpoint['lr_schedulers'] = lr_schedulers # add the hparams and state_dict from the model model = self.get_model() checkpoint['state_dict'] = model.state_dict() # give the model a chance to add a few things model.on_save_checkpoint(checkpoint) return checkpoint def copy_trainer_model_properties(self, model): if isinstance(model, DP): ref_model = model.module elif isinstance(model, DDP): ref_model = model.module else: ref_model = model for m in [model, ref_model]: m.trainer = self m.on_gpu = self.on_gpu m.use_dp = self.use_dp m.use_ddp = self.use_ddp m.testing = self.testing m.single_gpu = self.single_gpu def transfer_batch_to_gpu(self, batch, gpu_id): # base case: object can be directly moved using `cuda` or `to` if callable(getattr(batch, 'cuda', None)): return batch.cuda(gpu_id, non_blocking=True) elif callable(getattr(batch, 'to', None)): return batch.to(torch.device('cuda', gpu_id), non_blocking=True) # when list elif isinstance(batch, list): for i, x in enumerate(batch): batch[i] = self.transfer_batch_to_gpu(x, gpu_id) return batch # when tuple elif isinstance(batch, tuple): batch = list(batch) for i, x in enumerate(batch): batch[i] = self.transfer_batch_to_gpu(x, gpu_id) return tuple(batch) # when dict elif isinstance(batch, dict): for k, v in batch.items(): batch[k] = self.transfer_batch_to_gpu(v, gpu_id) return batch # nothing matches, return the value as is without transform return batch def set_distributed_mode(self, distributed_backend): # skip for CPU if self.num_gpus == 0: return # single GPU case # in single gpu case we allow ddp so we can train on multiple # nodes, 1 gpu per node elif self.num_gpus == 1: self.single_gpu = True self.use_dp = False self.use_ddp = False self.root_gpu = 0 self.data_parallel_device_ids = [0] else: if distributed_backend is not None: self.use_dp = distributed_backend == 'dp' self.use_ddp = distributed_backend == 'ddp' elif distributed_backend is None: self.use_dp = True self.use_ddp = False logging.info(f'gpu available: {torch.cuda.is_available()}, used: {self.on_gpu}') def ddp_train(self, gpu_idx, model): """ Entry point into a DP thread :param gpu_idx: :param model: :param cluster_obj: :return: """ # otherwise default to node rank 0 self.node_rank = 0 # show progressbar only on progress_rank 0 self.show_progress_bar = self.show_progress_bar and self.node_rank == 0 and gpu_idx == 0 # determine which process we are and world size if self.use_ddp: self.proc_rank = self.node_rank * self.num_gpus + gpu_idx self.world_size = self.num_gpus # let the exp know the rank to avoid overwriting logs if self.logger is not None: self.logger.rank = self.proc_rank # set up server using proc 0's ip address # try to init for 20 times at max in case ports are taken # where to store ip_table model.trainer = self model.init_ddp_connection(self.proc_rank, self.world_size) # CHOOSE OPTIMIZER # allow for lr schedulers as well model.model = model.build_model() if not self.testing: self.optimizers, self.lr_schedulers = self.init_optimizers(model.configure_optimizers()) # MODEL # copy model to each gpu if self.distributed_backend == 'ddp': torch.cuda.set_device(gpu_idx) model.cuda(gpu_idx) # set model properties before going into wrapper self.copy_trainer_model_properties(model) # override root GPU self.root_gpu = gpu_idx if self.distributed_backend == 'ddp': device_ids = [gpu_idx] else: device_ids = None # allow user to configure ddp model = model.configure_ddp(model, device_ids) # continue training routine self.run_pretrain_routine(model) def resolve_root_node_address(self, root_node): if '[' in root_node: name = root_node.split('[')[0] number = root_node.split(',')[0] if '-' in number: number = number.split('-')[0] number = re.sub('[^0-9]', '', number) root_node = name + number return root_node def log_metrics(self, metrics, grad_norm_dic, step=None): """Logs the metric dict passed in. :param metrics: :param grad_norm_dic: """ # added metrics by Lightning for convenience metrics['epoch'] = self.current_epoch # add norms metrics.update(grad_norm_dic) # turn all tensors to scalars scalar_metrics = self.metrics_to_scalars(metrics) step = step if step is not None else self.global_step # log actual metrics if self.proc_rank == 0 and self.logger is not None: self.logger.log_metrics(scalar_metrics, step=step) self.logger.save() def add_tqdm_metrics(self, metrics): for k, v in metrics.items(): if type(v) is torch.Tensor: v = v.item() self.tqdm_metrics[k] = v def metrics_to_scalars(self, metrics): new_metrics = {} for k, v in metrics.items(): if isinstance(v, torch.Tensor): v = v.item() if type(v) is dict: v = self.metrics_to_scalars(v) new_metrics[k] = v return new_metrics def process_output(self, output, train=False): """Reduces output according to the training mode. Separates loss from logging and tqdm metrics :param output: :return: """ # --------------- # EXTRACT CALLBACK KEYS # --------------- # all keys not progress_bar or log are candidates for callbacks callback_metrics = {} for k, v in output.items(): if k not in ['progress_bar', 'log', 'hiddens']: callback_metrics[k] = v if train and self.use_dp: num_gpus = self.num_gpus callback_metrics = self.reduce_distributed_output(callback_metrics, num_gpus) for k, v in callback_metrics.items(): if isinstance(v, torch.Tensor): callback_metrics[k] = v.item() # --------------- # EXTRACT PROGRESS BAR KEYS # --------------- try: progress_output = output['progress_bar'] # reduce progress metrics for tqdm when using dp if train and self.use_dp: num_gpus = self.num_gpus progress_output = self.reduce_distributed_output(progress_output, num_gpus) progress_bar_metrics = progress_output except Exception: progress_bar_metrics = {} # --------------- # EXTRACT LOGGING KEYS # --------------- # extract metrics to log to experiment try: log_output = output['log'] # reduce progress metrics for tqdm when using dp if train and self.use_dp: num_gpus = self.num_gpus log_output = self.reduce_distributed_output(log_output, num_gpus) log_metrics = log_output except Exception: log_metrics = {} # --------------- # EXTRACT LOSS # --------------- # if output dict doesn't have the keyword loss # then assume the output=loss if scalar loss = None if train: try: loss = output['loss'] except Exception: if type(output) is torch.Tensor: loss = output else: raise RuntimeError( 'No `loss` value in the dictionary returned from `model.training_step()`.' ) # when using dp need to reduce the loss if self.use_dp: loss = self.reduce_distributed_output(loss, self.num_gpus) # --------------- # EXTRACT HIDDEN # --------------- hiddens = output.get('hiddens') # use every metric passed in as a candidate for callback callback_metrics.update(progress_bar_metrics) callback_metrics.update(log_metrics) # convert tensors to numpy for k, v in callback_metrics.items(): if isinstance(v, torch.Tensor): callback_metrics[k] = v.item() return loss, progress_bar_metrics, log_metrics, callback_metrics, hiddens def reduce_distributed_output(self, output, num_gpus): if num_gpus <= 1: return output # when using DP, we get one output per gpu # average outputs and return if type(output) is torch.Tensor: return output.mean() for k, v in output.items(): # recurse on nested dics if isinstance(output[k], dict): output[k] = self.reduce_distributed_output(output[k], num_gpus) # do nothing when there's a scalar elif isinstance(output[k], torch.Tensor) and output[k].dim() == 0: pass # reduce only metrics that have the same number of gpus elif output[k].size(0) == num_gpus: reduced = torch.mean(output[k]) output[k] = reduced return output def clip_gradients(self): if self.gradient_clip_val > 0: model = self.get_model() torch.nn.utils.clip_grad_norm_(model.parameters(), self.gradient_clip_val) def print_nan_gradients(self): model = self.get_model() for param in model.parameters(): if (param.grad is not None) and torch.isnan(param.grad.float()).any(): logging.info(param, param.grad) def configure_accumulated_gradients(self, accumulate_grad_batches): self.accumulate_grad_batches = None if isinstance(accumulate_grad_batches, dict): self.accumulation_scheduler = GradientAccumulationScheduler(accumulate_grad_batches) elif isinstance(accumulate_grad_batches, int): schedule = {1: accumulate_grad_batches} self.accumulation_scheduler = GradientAccumulationScheduler(schedule) else: raise TypeError("Gradient accumulation supports only int and dict types") def get_dataloaders(self, model): if not self.testing: self.init_train_dataloader(model) self.init_val_dataloader(model) else: self.init_test_dataloader(model) if self.use_ddp: dist.barrier() if not self.testing: self.get_train_dataloader() self.get_val_dataloaders() else: self.get_test_dataloaders() def init_train_dataloader(self, model): self.fisrt_epoch = True self.get_train_dataloader = model.train_dataloader if isinstance(self.get_train_dataloader(), torch.utils.data.DataLoader): self.num_training_batches = len(self.get_train_dataloader()) self.num_training_batches = int(self.num_training_batches) else: self.num_training_batches = float('inf') self.is_iterable_train_dataloader = True if isinstance(self.val_check_interval, int): self.val_check_batch = self.val_check_interval else: self._percent_range_check('val_check_interval') self.val_check_batch = int(self.num_training_batches * self.val_check_interval) self.val_check_batch = max(1, self.val_check_batch) def init_val_dataloader(self, model): self.get_val_dataloaders = model.val_dataloader self.num_val_batches = 0 if self.get_val_dataloaders() is not None: if isinstance(self.get_val_dataloaders()[0], torch.utils.data.DataLoader): self.num_val_batches = sum(len(dataloader) for dataloader in self.get_val_dataloaders()) self.num_val_batches = int(self.num_val_batches) else: self.num_val_batches = float('inf') def init_test_dataloader(self, model): self.get_test_dataloaders = model.test_dataloader if self.get_test_dataloaders() is not None: if isinstance(self.get_test_dataloaders()[0], torch.utils.data.DataLoader): self.num_test_batches = sum(len(dataloader) for dataloader in self.get_test_dataloaders()) self.num_test_batches = int(self.num_test_batches) else: self.num_test_batches = float('inf') def evaluate(self, model, dataloaders, max_batches, test=False): """Run evaluation code. :param model: PT model :param dataloaders: list of PT dataloaders :param max_batches: Scalar :param test: boolean :return: """ # enable eval mode model.zero_grad() model.eval() # copy properties for forward overrides self.copy_trainer_model_properties(model) # disable gradients to save memory torch.set_grad_enabled(False) if test: self.get_model().test_start() # bookkeeping outputs = [] # run training for dataloader_idx, dataloader in enumerate(dataloaders): dl_outputs = [] for batch_idx, batch in enumerate(dataloader): if batch is None: # pragma: no cover continue # stop short when on fast_dev_run (sets max_batch=1) if batch_idx >= max_batches: break # ----------------- # RUN EVALUATION STEP # ----------------- output = self.evaluation_forward(model, batch, batch_idx, dataloader_idx, test) # track outputs for collation dl_outputs.append(output) # batch done if test: self.test_progress_bar.update(1) else: self.val_progress_bar.update(1) outputs.append(dl_outputs) # with a single dataloader don't pass an array if len(dataloaders) == 1: outputs = outputs[0] # give model a chance to do something with the outputs (and method defined) model = self.get_model() if test: eval_results_ = model.test_end(outputs) else: eval_results_ = model.validation_end(outputs) eval_results = eval_results_ # enable train mode again model.train() # enable gradients to save memory torch.set_grad_enabled(True) return eval_results def run_evaluation(self, test=False): # when testing make sure user defined a test step model = self.get_model() model.on_pre_performance_check() # select dataloaders if test: dataloaders = self.get_test_dataloaders() max_batches = self.num_test_batches else: # val dataloaders = self.get_val_dataloaders() max_batches = self.num_val_batches # init validation or test progress bar # main progress bar will already be closed when testing so initial position is free position = 2 * self.process_position + (not test) desc = 'Testing' if test else 'Validating' pbar = tqdm.tqdm(desc=desc, total=max_batches, leave=test, position=position, disable=not self.show_progress_bar, dynamic_ncols=True, unit='batch', file=sys.stdout) setattr(self, f'{"test" if test else "val"}_progress_bar', pbar) # run evaluation eval_results = self.evaluate(self.model, dataloaders, max_batches, test) if eval_results is not None: _, prog_bar_metrics, log_metrics, callback_metrics, _ = self.process_output( eval_results) # add metrics to prog bar self.add_tqdm_metrics(prog_bar_metrics) # log metrics self.log_metrics(log_metrics, {}) # track metrics for callbacks self.callback_metrics.update(callback_metrics) # hook model.on_post_performance_check() # add model specific metrics tqdm_metrics = self.training_tqdm_dict if not test: self.main_progress_bar.set_postfix(**tqdm_metrics) # close progress bar if test: self.test_progress_bar.close() else: self.val_progress_bar.close() # model checkpointing if self.proc_rank == 0 and self.checkpoint_callback is not None and not test: self.checkpoint_callback.on_epoch_end(epoch=self.current_epoch, logs=self.callback_metrics) def evaluation_forward(self, model, batch, batch_idx, dataloader_idx, test=False): # make dataloader_idx arg in validation_step optional args = [batch, batch_idx] # print(batch) if test and len(self.get_test_dataloaders()) > 1: args.append(dataloader_idx) elif not test and len(self.get_val_dataloaders()) > 1: args.append(dataloader_idx) # handle DP, DDP forward if self.use_ddp or self.use_dp: output = model(*args) return output # single GPU if self.single_gpu: # for single GPU put inputs on gpu manually root_gpu = 0 if isinstance(self.data_parallel_device_ids, list): root_gpu = self.data_parallel_device_ids[0] batch = self.transfer_batch_to_gpu(batch, root_gpu) args[0] = batch # CPU if test: output = model.test_step(*args) else: output = model.validation_step(*args) return output def train(self): model = self.get_model() # run all epochs for epoch in range(self.current_epoch, 1000000): # set seed for distributed sampler (enables shuffling for each epoch) if self.use_ddp and hasattr(self.get_train_dataloader().sampler, 'set_epoch'): self.get_train_dataloader().sampler.set_epoch(epoch) # get model model = self.get_model() # update training progress in trainer and model model.current_epoch = epoch self.current_epoch = epoch total_val_batches = 0 if not self.disable_validation: # val can be checked multiple times in epoch is_val_epoch = (self.current_epoch + 1) % self.check_val_every_n_epoch == 0 val_checks_per_epoch = self.num_training_batches // self.val_check_batch val_checks_per_epoch = val_checks_per_epoch if is_val_epoch else 0 total_val_batches = self.num_val_batches * val_checks_per_epoch # total batches includes multiple val checks self.total_batches = self.num_training_batches + total_val_batches self.batch_loss_value = 0 # accumulated grads if self.is_iterable_train_dataloader: # for iterable train loader, the progress bar never ends num_iterations = None else: num_iterations = self.total_batches # reset progress bar # .reset() doesn't work on disabled progress bar so we should check desc = f'Epoch {epoch + 1}' if not self.is_iterable_train_dataloader else '' self.main_progress_bar.set_description(desc) # changing gradient according accumulation_scheduler self.accumulation_scheduler.on_epoch_begin(epoch, self) # ----------------- # RUN TNG EPOCH # ----------------- self.run_training_epoch() # update LR schedulers if self.lr_schedulers is not None: for lr_scheduler in self.lr_schedulers: lr_scheduler.step(epoch=self.current_epoch) self.main_progress_bar.close() model.on_train_end() if self.logger is not None: self.logger.finalize("success") def run_training_epoch(self): # before epoch hook if self.is_function_implemented('on_epoch_start'): model = self.get_model() model.on_epoch_start() # run epoch for batch_idx, batch in enumerate(self.get_train_dataloader()): # stop epoch if we limited the number of training batches if batch_idx >= self.num_training_batches: break self.batch_idx = batch_idx model = self.get_model() model.global_step = self.global_step # --------------- # RUN TRAIN STEP # --------------- output = self.run_training_batch(batch, batch_idx) batch_result, grad_norm_dic, batch_step_metrics = output # when returning -1 from train_step, we end epoch early early_stop_epoch = batch_result == -1 # --------------- # RUN VAL STEP # --------------- should_check_val = ( not self.disable_validation and self.global_step % self.val_check_batch == 0 and not self.fisrt_epoch) self.fisrt_epoch = False if should_check_val: self.run_evaluation(test=self.testing) # when logs should be saved should_save_log = (batch_idx + 1) % self.log_save_interval == 0 or early_stop_epoch if should_save_log: if self.proc_rank == 0 and self.logger is not None: self.logger.save() # when metrics should be logged should_log_metrics = batch_idx % self.row_log_interval == 0 or early_stop_epoch if should_log_metrics: # logs user requested information to logger self.log_metrics(batch_step_metrics, grad_norm_dic) self.global_step += 1 self.total_batch_idx += 1 # end epoch early # stop when the flag is changed or we've gone past the amount # requested in the batches if early_stop_epoch: break if self.global_step > self.max_updates: print("| Training end..") exit() # epoch end hook if self.is_function_implemented('on_epoch_end'): model = self.get_model() model.on_epoch_end() def run_training_batch(self, batch, batch_idx): # track grad norms grad_norm_dic = {} # track all metrics for callbacks all_callback_metrics = [] # track metrics to log all_log_metrics = [] if batch is None: return 0, grad_norm_dic, {} # hook if self.is_function_implemented('on_batch_start'): model_ref = self.get_model() response = model_ref.on_batch_start(batch) if response == -1: return -1, grad_norm_dic, {} splits = [batch] self.hiddens = None for split_idx, split_batch in enumerate(splits): self.split_idx = split_idx # call training_step once per optimizer for opt_idx, optimizer in enumerate(self.optimizers): if optimizer is None: continue # make sure only the gradients of the current optimizer's paramaters are calculated # in the training step to prevent dangling gradients in multiple-optimizer setup. if len(self.optimizers) > 1: for param in self.get_model().parameters(): param.requires_grad = False for group in optimizer.param_groups: for param in group['params']: param.requires_grad = True # wrap the forward step in a closure so second order methods work def optimizer_closure(): # forward pass output = self.training_forward( split_batch, batch_idx, opt_idx, self.hiddens) closure_loss = output[0] progress_bar_metrics = output[1] log_metrics = output[2] callback_metrics = output[3] self.hiddens = output[4] if closure_loss is None: return None # accumulate loss # (if accumulate_grad_batches = 1 no effect) closure_loss = closure_loss / self.accumulate_grad_batches # backward pass model_ref = self.get_model() if closure_loss.requires_grad: model_ref.backward(closure_loss, optimizer) # track metrics for callbacks all_callback_metrics.append(callback_metrics) # track progress bar metrics self.add_tqdm_metrics(progress_bar_metrics) all_log_metrics.append(log_metrics) # insert after step hook if self.is_function_implemented('on_after_backward'): model_ref = self.get_model() model_ref.on_after_backward() return closure_loss # calculate loss loss = optimizer_closure() if loss is None: continue # nan grads if self.print_nan_grads: self.print_nan_gradients() # track total loss for logging (avoid mem leaks) self.batch_loss_value += loss.item() # gradient update with accumulated gradients if (self.batch_idx + 1) % self.accumulate_grad_batches == 0: # track gradient norms when requested if batch_idx % self.row_log_interval == 0: if self.track_grad_norm > 0: model = self.get_model() grad_norm_dic = model.grad_norm( self.track_grad_norm) # clip gradients self.clip_gradients() # calls .step(), .zero_grad() # override function to modify this behavior model = self.get_model() model.optimizer_step(self.current_epoch, batch_idx, optimizer, opt_idx) # calculate running loss for display self.running_loss.append(self.batch_loss_value) self.batch_loss_value = 0 self.avg_loss = np.mean(self.running_loss[-100:]) # activate batch end hook if self.is_function_implemented('on_batch_end'): model = self.get_model() model.on_batch_end() # update progress bar self.main_progress_bar.update(1) self.main_progress_bar.set_postfix(**self.training_tqdm_dict) # collapse all metrics into one dict all_log_metrics = {k: v for d in all_log_metrics for k, v in d.items()} # track all metrics for callbacks self.callback_metrics.update({k: v for d in all_callback_metrics for k, v in d.items()}) return 0, grad_norm_dic, all_log_metrics def training_forward(self, batch, batch_idx, opt_idx, hiddens): """ Handle forward for each training case (distributed, single gpu, etc...) :param batch: :param batch_idx: :return: """ # --------------- # FORWARD # --------------- # enable not needing to add opt_idx to training_step args = [batch, batch_idx, opt_idx] # distributed forward if self.use_ddp or self.use_dp: output = self.model(*args) # single GPU forward elif self.single_gpu: gpu_id = 0 if isinstance(self.data_parallel_device_ids, list): gpu_id = self.data_parallel_device_ids[0] batch = self.transfer_batch_to_gpu(copy.copy(batch), gpu_id) args[0] = batch output = self.model.training_step(*args) # CPU forward else: output = self.model.training_step(*args) # allow any mode to define training_end model_ref = self.get_model() output_ = model_ref.training_end(output) if output_ is not None: output = output_ # format and reduce outputs accordingly output = self.process_output(output, train=True) return output # --------------- # Utils # --------------- def is_function_implemented(self, f_name): model = self.get_model() f_op = getattr(model, f_name, None) return callable(f_op) def _percent_range_check(self, name): value = getattr(self, name) msg = f"`{name}` must lie in the range [0.0, 1.0], but got {value:.3f}." if name == "val_check_interval": msg += " If you want to disable validation set `val_percent_check` to 0.0 instead." if not 0. <= value <= 1.: raise ValueError(msg)