""" The lightning training loop handles everything except the actual computations of your model. To decide what will happen in your training loop, define the `training_step` function. Below are all the things lightning automates for you in the training loop. Accumulated gradients --------------------- Accumulated gradients runs K small batches of size N before doing a backwards pass. The effect is a large effective batch size of size KxN. .. code-block:: python # DEFAULT (ie: no accumulated grads) trainer = Trainer(accumulate_grad_batches=1) Force training for min or max epochs ------------------------------------ It can be useful to force training for a minimum number of epochs or limit to a max number .. code-block:: python # DEFAULT trainer = Trainer(min_epochs=1, max_epochs=1000) Force disable early stop ------------------------ To disable early stopping pass None to the early_stop_callback .. code-block:: python # DEFAULT trainer = Trainer(early_stop_callback=None) Gradient Clipping ----------------- Gradient clipping may be enabled to avoid exploding gradients. Specifically, this will `clip the gradient norm computed over all model parameters `together `_. .. code-block:: python # DEFAULT (ie: don't clip) trainer = Trainer(gradient_clip_val=0) # clip gradients with norm above 0.5 trainer = Trainer(gradient_clip_val=0.5) Inspect gradient norms ---------------------- Looking at grad norms can help you figure out where training might be going wrong. .. code-block:: python # DEFAULT (-1 doesn't track norms) trainer = Trainer(track_grad_norm=-1) # track the LP norm (P=2 here) trainer = Trainer(track_grad_norm=2) Set how much of the training set to check ----------------------------------------- If you don't want to check 100% of the training set (for debugging or if it's huge), set this flag. train_percent_check will be overwritten by overfit_pct if `overfit_pct > 0` .. code-block:: python # DEFAULT trainer = Trainer(train_percent_check=1.0) # check 10% only trainer = Trainer(train_percent_check=0.1) Packed sequences as inputs -------------------------- When using PackedSequence, do 2 things: 1. return either a padded tensor in dataset or a list of variable length tensors in the dataloader collate_fn (example above shows the list implementation). 2. Pack the sequence in forward or training and validation steps depending on use case. .. code-block:: python # For use in dataloader def collate_fn(batch): x = [item[0] for item in batch] y = [item[1] for item in batch] return x, y # In module def training_step(self, batch, batch_idx): x = rnn.pack_sequence(batch[0], enforce_sorted=False) y = rnn.pack_sequence(batch[1], enforce_sorted=False) Truncated Backpropagation Through Time -------------------------------------- There are times when multiple backwards passes are needed for each batch. For example, it may save memory to use Truncated Backpropagation Through Time when training RNNs. When this flag is enabled each batch is split into sequences of size truncated_bptt_steps and passed to training_step(...) separately. A default splitting function is provided, however, you can override it for more flexibility. See `tbptt_split_batch`. .. code-block:: python # DEFAULT (single backwards pass per batch) trainer = Trainer(truncated_bptt_steps=None) # (split batch into sequences of size 2) trainer = Trainer(truncated_bptt_steps=2) NaN detection and intervention ------------------------------ When the `terminate_on_nan` flag is enabled, after every forward pass during training, Lightning will check that 1. the loss you return in `training_step` is finite (not NaN and not +/-inf) 2. the model parameters have finite values. Lightning will terminate the training loop with an error message if NaN or infinite values are detected. If this happens, you should investigate numerically unstable operations in your model. .. code-block:: python # DEFAULT (won't perform the NaN check) trainer = Trainer(terminate_on_nan=False) # (NaN check each batch and terminate on NaN or infinite values) trainer = Trainer(terminate_on_nan=True) """ import atexit import signal from abc import ABC, abstractmethod from typing import Callable from typing import Union, List import numpy as np import torch from torch.utils.data import DataLoader from pytorch_lightning import _logger as log from pytorch_lightning.callbacks.base import Callback from pytorch_lightning.core.lightning import LightningModule from pytorch_lightning.loggers import LightningLoggerBase from pytorch_lightning.trainer.supporters import TensorRunningAccum from pytorch_lightning.utilities import rank_zero_warn from pytorch_lightning.utilities.exceptions import MisconfigurationException import subprocess try: from apex import amp except ImportError: APEX_AVAILABLE = False else: APEX_AVAILABLE = True try: import torch_xla.distributed.parallel_loader as xla_pl import torch_xla.core.xla_model as xm except ImportError: XLA_AVAILABLE = False else: XLA_AVAILABLE = True try: import horovod.torch as hvd except ImportError: HOROVOD_AVAILABLE = False else: HOROVOD_AVAILABLE = True # constant which signals should be catched for graceful trainer shutdown SIGNAL_TERMINATE = ('SIGTERM', 'SIGSEGV', 'SIGINT') class TrainerTrainLoopMixin(ABC): # this is just a summary on variables used in this abstract class, # the proper values/initialisation should be done in child class max_epochs: int min_epochs: int on_gpu: bool use_ddp: bool use_dp: bool use_ddp2: bool use_horovod: bool single_gpu: bool use_tpu: bool data_parallel_device_ids: ... check_val_every_n_epoch: ... num_training_batches: int val_check_batch: ... num_val_batches: int disable_validation: bool fast_dev_run: ... accumulation_scheduler: ... lr_schedulers: ... enable_early_stop: ... early_stop_callback: ... callback_metrics: ... logger: Union[LightningLoggerBase, bool] global_step: int testing: bool log_save_interval: float proc_rank: int row_log_interval: float truncated_bptt_steps: ... optimizers: ... optimizer_frequencies: ... accumulate_grad_batches: int track_grad_norm: ... model: LightningModule interrupted: bool running_loss: ... progress_bar_dict: ... reduce_lr_on_plateau_scheduler: ... profiler: ... batch_idx: int precision: ... train_dataloader: DataLoader reload_dataloaders_every_epoch: bool max_steps: int min_steps: int total_batch_idx: int checkpoint_callback: ... terminate_on_nan: bool tpu_id: int # Callback system callbacks: List[Callback] on_train_start: Callable on_train_end: Callable on_batch_start: Callable on_batch_end: Callable on_epoch_start: Callable on_epoch_end: Callable on_validation_end: Callable @abstractmethod def get_model(self): """Warning: this is just empty shell for code implemented in other class.""" @abstractmethod def is_function_implemented(self, *args): """Warning: this is just empty shell for code implemented in other class.""" @abstractmethod def run_evaluation(self, *args): """Warning: this is just empty shell for code implemented in other class.""" @abstractmethod def transfer_batch_to_gpu(self, *args): """Warning: this is just empty shell for code implemented in other class.""" @abstractmethod def transfer_batch_to_tpu(self, *args): """Warning: this is just empty shell for code implemented in other class.""" @abstractmethod def clip_gradients(self): """Warning: this is just empty shell for code implemented in other class.""" @abstractmethod def detect_nan_tensors(self, *args): """Warning: this is just empty shell for code implemented in other class.""" @abstractmethod def is_overridden(self, *args): """Warning: this is just empty shell for code implemented in other class.""" @abstractmethod def add_progress_bar_metrics(self, *args): """Warning: this is just empty shell for code implemented in other class.""" @abstractmethod def log_metrics(self, *args): """Warning: this is just empty shell for code implemented in other class.""" @abstractmethod def process_output(self, *args): """Warning: this is just empty shell for code implemented in other class.""" @abstractmethod def reset_train_dataloader(self, *args): """Warning: this is just empty shell for code implemented in other class.""" @abstractmethod def reset_val_dataloader(self, model): """Warning: this is just empty shell for code implemented in other class.""" @abstractmethod def has_arg(self, *args): """Warning: this is just empty shell for code implemented in other class.""" def train(self): # add signal handlers for process kills # def _signal_kill_handler(*args): # return TrainerTrainLoopMixin.run_training_teardown(self) # # orig_signal_handlers = {} # for sig_name in SIGNAL_TERMINATE: # orig_signal_handlers[sig_name] = signal.signal(getattr(signal, sig_name), # _signal_kill_handler) # get model model = self.get_model() # load data # if reload_dataloaders_every_epoch, this is moved to the epoch loop if not self.reload_dataloaders_every_epoch: self.reset_train_dataloader(model) self.reset_val_dataloader(model) # Train start events with self.profiler.profile('on_train_start'): # callbacks self.on_train_start() # initialize early stop callback if self.early_stop_callback is not None: self.early_stop_callback.on_train_start(self, self.get_model()) # model hooks model.on_train_start() try: # run all epochs for epoch in range(self.current_epoch, self.max_epochs): # reset train dataloader if self.reload_dataloaders_every_epoch: self.reset_train_dataloader(model) # set seed for distributed sampler (enables shuffling for each epoch) if (self.use_ddp or self.use_horovod) \ and hasattr(self.train_dataloader, 'sampler') \ and hasattr(self.train_dataloader.sampler, 'set_epoch'): self.train_dataloader.sampler.set_epoch(epoch) # update training progress in trainer and model model.current_epoch = epoch self.current_epoch = epoch # changing gradient according accumulation_scheduler self.accumulation_scheduler.on_epoch_start(self, self.get_model()) # stores accumulated grad fractions per batch self.batch_loss_value = TensorRunningAccum( window_length=self.accumulate_grad_batches ) # ----------------- # RUN TNG EPOCH # ----------------- self.run_training_epoch() if self.max_steps and self.max_steps == self.global_step: self.run_training_teardown() return # update LR schedulers self.update_learning_rates(interval='epoch') # early stopping met_min_epochs = epoch >= self.min_epochs - 1 met_min_steps = self.global_step >= self.min_steps if self.min_steps else True # TODO wrap this logic into the callback if self.enable_early_stop: if (met_min_epochs and met_min_steps) or self.fast_dev_run: should_stop = self.early_stop_callback.on_validation_end(self, self.get_model()) # stop training stop = should_stop and met_min_epochs if stop: self.run_training_teardown() return self.run_training_teardown() except KeyboardInterrupt: rank_zero_warn('Detected KeyboardInterrupt, attempting graceful shutdown...') # user could press ctrl+c many times... only shutdown once if not self.interrupted: self.interrupted = True for proc in self.interactive_ddp_procs: subprocess.Popen.kill(proc) self.run_training_teardown() def run_training_epoch(self): # get model model = self.get_model() # Epoch start events with self.profiler.profile('on_epoch_start'): # callbacks self.on_epoch_start() # model hooks if self.is_function_implemented('on_epoch_start'): model.on_epoch_start() # track local dataloader so TPU can wrap each epoch train_dataloader = self.train_dataloader # on TPU we have to wrap it under the ParallelLoader if self.use_tpu: device = xm.xla_device(self.tpu_id) train_dataloader = xla_pl.ParallelLoader(train_dataloader, [device]) train_dataloader = train_dataloader.per_device_loader(device) # bookkeeping outputs = [] # run epoch for batch_idx, (batch, is_last_batch) in self.profiler.profile_iterable( enumerate(_with_is_last(train_dataloader)), "get_train_batch" ): # stop epoch if we limited the number of training batches if batch_idx >= self.num_training_batches: break self.batch_idx = batch_idx model.global_step = self.global_step # --------------- # RUN TRAIN STEP # --------------- _outputs = self.run_training_batch(batch, batch_idx) batch_result, grad_norm_dic, batch_step_metrics, batch_output = _outputs # only track outputs when user implements training_epoch_end # otherwise we will build up unnecessary memory if self.is_overridden('training_epoch_end', model=self.get_model()): outputs.append(batch_output) # when returning -1 from train_step, we end epoch early early_stop_epoch = batch_result == -1 # TODO: consolidate all actions that need to take place only after # self.accumulate_grad_batches steps (optimizer step, lr update, global step increment) if (self.batch_idx + 1) % self.accumulate_grad_batches == 0: # update lr self.update_learning_rates(interval='step') # --------------- # RUN VAL STEP # --------------- is_val_check_batch = (batch_idx + 1) % self.val_check_batch == 0 can_check_epoch = (self.current_epoch + 1) % self.check_val_every_n_epoch == 0 can_check_val = not self.disable_validation and can_check_epoch should_check_val = is_val_check_batch or early_stop_epoch should_check_val = should_check_val or (is_last_batch and self.val_check_batch == float('inf')) should_check_val = can_check_val and should_check_val # --------------- # CHECKPOINTING, EARLY STOPPING # --------------- # fast_dev_run always forces val checking after train batch if self.fast_dev_run or should_check_val: self.run_evaluation(test_mode=self.testing) self.call_checkpoint_callback() # when logs should be saved should_save_log = (batch_idx + 1) % self.log_save_interval == 0 or early_stop_epoch if should_save_log or self.fast_dev_run: 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 or self.fast_dev_run: # logs user requested information to logger self.log_metrics(batch_step_metrics, grad_norm_dic) # progress global step according to grads progress if (self.batch_idx + 1) % self.accumulate_grad_batches == 0: self.global_step += 1 self.total_batch_idx += 1 # max steps reached, end training if self.max_steps is not None and self.max_steps == self.global_step: break # end epoch early # stop when the flag is changed or we've gone past the amount # requested in the batches if early_stop_epoch or self.fast_dev_run: break if self.use_horovod: hvd.join(hvd.local_rank() if self.on_gpu else -1) # process epoch outputs model = self.get_model() if self.is_overridden('training_epoch_end', model=model): epoch_output = model.training_epoch_end(outputs) _processed_outputs = self.process_output(epoch_output) log_epoch_metrics = _processed_outputs[2] callback_epoch_metrics = _processed_outputs[3] self.log_metrics(log_epoch_metrics, {}) self.callback_metrics.update(callback_epoch_metrics) self.add_progress_bar_metrics(_processed_outputs[1]) # when no val loop is present or fast-dev-run still need to call checkpoints if not self.is_overridden('validation_step') and not (self.fast_dev_run or should_check_val): self.call_checkpoint_callback() # Epoch end events with self.profiler.profile('on_epoch_end'): # callbacks self.on_epoch_end() # model hooks if self.is_function_implemented('on_epoch_end'): 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, {}, {} # Batch start events with self.profiler.profile('on_batch_start'): # callbacks self.on_batch_start() # hooks if self.is_function_implemented('on_batch_start'): response = self.get_model().on_batch_start(batch) if response == -1: return -1, grad_norm_dic, {}, {} splits = [batch] if self.truncated_bptt_steps is not None: model_ref = self.get_model() with self.profiler.profile('tbptt_split_batch'): splits = model_ref.tbptt_split_batch(batch, self.truncated_bptt_steps) self.hiddens = None for split_idx, split_batch in enumerate(splits): self.split_idx = split_idx for opt_idx, optimizer in self._get_optimizers_iterable(): # make sure only the gradients of the current optimizer's parameters 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 with self.profiler.profile('model_forward'): if self.use_amp and self.use_native_amp: with torch.cuda.amp.autocast(): output_dict = self.training_forward(split_batch, batch_idx, opt_idx, self.hiddens) else: output_dict = self.training_forward(split_batch, batch_idx, opt_idx, self.hiddens) # format and reduce outputs accordingly processed_output = self.process_output(output_dict, train=True) closure_loss, progress_bar_metrics, log_metrics, callback_metrics, self.hiddens = processed_output # accumulate loss # (if accumulate_grad_batches = 1 no effect) closure_loss = closure_loss / self.accumulate_grad_batches # backward pass model_ref = self.get_model() with self.profiler.profile('model_backward'): model_ref.backward(self, closure_loss, optimizer, opt_idx) # track metrics for callbacks all_callback_metrics.append(callback_metrics) # track progress bar metrics self.add_progress_bar_metrics(progress_bar_metrics) all_log_metrics.append(log_metrics) if self.use_horovod: # Synchronize Horovod to ensure gradient manipulations (e.g., loss scaling) are valid optimizer.synchronize() # insert after step hook if self.is_function_implemented('on_after_backward'): model_ref = self.get_model() with self.profiler.profile('on_after_backward'): model_ref.on_after_backward() return closure_loss, callback_metrics # calculate loss loss, batch_output = optimizer_closure() # check if loss or model weights are nan if self.terminate_on_nan: self.detect_nan_tensors(loss) # track total loss for logging (avoid mem leaks) self.batch_loss_value.append(loss) # 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 float(self.track_grad_norm) > 0: model = self.get_model() grad_norm_dic = model.grad_norm( self.track_grad_norm) # clip gradients if self.use_amp and self.use_native_amp: self.scaler.unscale_(optimizer) self.clip_gradients() # calls .step(), .zero_grad() # override function to modify this behavior model = self.get_model() with self.profiler.profile('optimizer_step'): model.optimizer_step(self.current_epoch, batch_idx, optimizer, opt_idx, lambda: optimizer_closure()[0]) # calculate running loss for display self.running_loss.append(self.batch_loss_value.mean()) # reset for next set of accumulated grads self.batch_loss_value.reset() # Batch end events with self.profiler.profile('on_batch_end'): # callbacks self.on_batch_end() # model hooks if self.is_function_implemented('on_batch_end'): self.get_model().on_batch_end() # 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, batch_output def _get_optimizers_iterable(self): if not self.optimizer_frequencies: # call training_step once per optimizer return list(enumerate(self.optimizers)) optimizer_freq_cumsum = np.cumsum(self.optimizer_frequencies) optimizers_loop_length = optimizer_freq_cumsum[-1] current_place_in_loop = self.total_batch_idx % optimizers_loop_length # find optimzier index by looking for the first {item > current_place} in the cumsum list opt_idx = np.argmax(optimizer_freq_cumsum > current_place_in_loop) return [(opt_idx, self.optimizers[opt_idx])] # @atexit.register def run_training_teardown(self): if hasattr(self, '_teardown_already_run') and self._teardown_already_run: return # Train end events with self.profiler.profile('on_train_end'): # callbacks self.on_train_end() # model hooks if self.is_function_implemented('on_train_end'): self.get_model().on_train_end() if self.logger is not None: self.logger.finalize("success") # summarize profile results self.profiler.describe() self._teardown_already_run = True 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] if len(self.optimizers) > 1: if self.has_arg('training_step', 'optimizer_idx'): args.append(opt_idx) else: num_opts = len(self.optimizers) raise ValueError( f'Your LightningModule defines {num_opts} optimizers but ' f'training_step is missing the "optimizer_idx" argument.' ) # pass hiddens if using tbptt if self.truncated_bptt_steps is not None: args.append(hiddens) # distributed forward if self.use_ddp or self.use_ddp2 or self.use_dp: output = self.model(*args) # Horovod elif self.use_horovod and self.on_gpu: batch = self.transfer_batch_to_gpu(batch, hvd.local_rank()) args[0] = batch output = self.model.training_step(*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] # Don't copy the batch since there is a single gpu that the batch could # be referenced from and if there are multiple optimizers the batch will # wind up copying it to the same device repeatedly. batch = self.transfer_batch_to_gpu(batch, gpu_id) args[0] = batch output = self.model.training_step(*args) # TPU support elif self.use_tpu: batch = self.transfer_batch_to_tpu(batch, self.tpu_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_step_end # do something will all the dp outputs (like softmax) if self.is_overridden('training_step_end'): model_ref = self.get_model() with self.profiler.profile('training_step_end'): output = model_ref.training_step_end(output) # allow any mode to define training_end # TODO: remove in 1.0.0 if self.is_overridden('training_end'): model_ref = self.get_model() with self.profiler.profile('training_end'): output = model_ref.training_end(output) rank_zero_warn('`training_end` was deprecated in 0.7.0 and will be removed 1.0.0.' ' Use training_epoch_end instead', DeprecationWarning) return output def update_learning_rates(self, interval: str): """Update learning rates. Args: interval: either 'epoch' or 'step'. """ if not self.lr_schedulers: return for lr_scheduler in self.lr_schedulers: current_idx = self.batch_idx if interval == 'step' else self.current_epoch current_idx += 1 # account for both batch and epoch starts from 0 # Take step if call to update_learning_rates matches the interval key and # the current step modulo the schedulers frequency is zero if lr_scheduler['interval'] == interval and current_idx % lr_scheduler['frequency'] == 0: # If instance of ReduceLROnPlateau, we need to pass validation loss if lr_scheduler['reduce_on_plateau']: monitor_key = lr_scheduler['monitor'] monitor_val = self.callback_metrics.get(monitor_key) if monitor_val is None: avail_metrics = ','.join(list(self.callback_metrics.keys())) raise MisconfigurationException( f'ReduceLROnPlateau conditioned on metric {monitor_key}' f' which is not available. Available metrics are: {avail_metrics}.' ' Condition can be set using `monitor` key in lr scheduler dict' ) lr_scheduler['scheduler'].step(monitor_val) else: lr_scheduler['scheduler'].step() def call_checkpoint_callback(self): if self.checkpoint_callback is not None: self.checkpoint_callback.on_validation_end(self, self.get_model()) def _with_is_last(iterable): """Pass through values from the given iterable with an added boolean indicating if this is the last item. See `https://stackoverflow.com/a/1630350 `_""" it = iter(iterable) last = next(it) for val in it: # yield last and has next yield last, False last = val # yield last, no longer has next yield last, True