def __init__(self, min_steps: Optional[int] = None, max_steps: int = -1) -> None: super().__init__() if max_steps is None: rank_zero_deprecation( "Setting `max_steps = None` is deprecated in v1.5 and will no longer be supported in v1.7." " Use `max_steps = -1` instead.") max_steps = -1 elif max_steps < -1: raise MisconfigurationException( f"`max_steps` must be a non-negative integer or -1 (infinite steps). You passed in {max_steps}." ) self.min_steps = min_steps self.max_steps = max_steps self.batch_progress = BatchProgress() self.scheduler_progress = SchedulerProgress() self.batch_loop = TrainingBatchLoop() self.val_loop = loops.EvaluationLoop(verbose=False) self._results = _ResultCollection(training=True) self._outputs: _OUTPUTS_TYPE = [] self._warning_cache = WarningCache() # caches the loaded dataloader state until dataloader objects are available self._dataloader_state_dict: Dict[str, Any] = {} self._batches_that_stepped: int = 0
def connect(self, trainer: 'pl.Trainer', *args: Any, **kwargs: Any) -> None: """Connects the loop with all necessary parts like trainer and accelerators""" super().connect(trainer, *args, **kwargs) self.batch_loop = TrainingBatchLoop() self.batch_loop.connect(trainer) self.val_loop = loops.EvaluationLoop() self.val_loop.connect(trainer)
def __init__(self, min_steps: int, max_steps: int): super().__init__() self.min_steps: int = min_steps self.max_steps: int = max_steps self.global_step: int = 0 # the total batch index across all epochs self.total_batch_idx: int = 0 # the current batch index in the loop that runs over the dataloader(s) self.iteration_count: int = 0 # the current split index when the batch gets split into chunks in truncated backprop through time self.split_idx: Optional[int] = None # the number of batches seen this run, updates immediately after batch_loop.run() self.batches_seen: int = 0 self.is_last_batch: Optional[bool] = None self.progress = TrainingEpochProgress() self.batch_loop = TrainingBatchLoop() self.val_loop = loops.EvaluationLoop() self._results = ResultCollection(training=True) self._dataloader_idx: Optional[int] = None self._warning_cache: WarningCache = WarningCache() self._epoch_output: Optional[List[List[STEP_OUTPUT]]] = None
class TrainingEpochLoop(loops.Loop[_OUTPUTS_TYPE]): """Runs over all batches in a dataloader (one epoch). Args: min_steps: The minimum number of steps (batches) to process max_steps: The maximum number of steps (batches) to process """ def __init__(self, min_steps: Optional[int] = None, max_steps: int = -1) -> None: super().__init__() if max_steps is None: rank_zero_deprecation( "Setting `max_steps = None` is deprecated in v1.5 and will no longer be supported in v1.7." " Use `max_steps = -1` instead.") max_steps = -1 elif max_steps < -1: raise MisconfigurationException( f"`max_steps` must be a non-negative integer or -1 (infinite steps). You passed in {max_steps}." ) self.min_steps = min_steps self.max_steps = max_steps self.batch_progress = BatchProgress() self.scheduler_progress = SchedulerProgress() self.batch_loop = TrainingBatchLoop() self.val_loop = loops.EvaluationLoop(verbose=False) self._results = _ResultCollection(training=True) self._outputs: _OUTPUTS_TYPE = [] self._warning_cache = WarningCache() # caches the loaded dataloader state until dataloader objects are available self._dataloader_state_dict: Dict[str, Any] = {} self._batches_that_stepped: int = 0 @property def total_batch_idx(self) -> int: """Returns the current batch index (across epochs)""" # use `ready` instead of `completed` in case this is accessed after `completed` has been increased # but before the next `ready` increase return self.batch_progress.total.ready - 1 @property def batch_idx(self) -> int: """Returns the current batch index (within this epoch)""" # use `ready` instead of `completed` in case this is accessed after `completed` has been increased # but before the next `ready` increase return self.batch_progress.current.ready - 1 @property def global_step(self) -> int: lightning_module = self.trainer.lightning_module if lightning_module is None or lightning_module.automatic_optimization: return self.batch_loop.optimizer_loop.optim_progress.optimizer_steps return self.batch_loop.manual_loop.optim_step_progress.total.completed @property def _is_training_done(self) -> bool: max_steps_reached = _is_max_limit_reached(self.global_step, self.max_steps) return max_steps_reached or self._num_ready_batches_reached() @property def _is_validation_done(self) -> bool: # when we are restarting we want to check whether the val loop has finished return not self.restarting or self.val_loop.done @property def done(self) -> bool: """Evaluates when to leave the loop.""" return (self._is_training_done and self._is_validation_done) or self.trainer.should_stop def connect( # type: ignore[override] self, batch_loop: Optional[TrainingBatchLoop] = None, val_loop: Optional["loops.EvaluationLoop"] = None, ) -> None: """Optionally connect a custom batch or validation loop to this training epoch loop.""" if batch_loop is not None: self.batch_loop = batch_loop if val_loop is not None: self.val_loop = val_loop def reset(self) -> None: """Resets the internal state of the loop for a new run.""" if self.restarting: self.batch_progress.reset_on_restart() self.scheduler_progress.reset_on_restart() self.batch_loop.optimizer_loop.optim_progress.reset_on_restart() trainer = self.trainer if not trainer.state._fault_tolerant_mode.is_enabled and trainer.num_training_batches != float( "inf"): expected_steps = math.ceil(trainer.num_training_batches / trainer.accumulate_grad_batches) if self.global_step % expected_steps != 0: rank_zero_warn( "You're resuming from a checkpoint that ended before the epoch ended. This can cause unreliable" " results if further training is done. Consider using an end-of-epoch checkpoint or enabling" " fault-tolerant training:" " https://pytorch-lightning.readthedocs.io/en/stable/advanced/fault_tolerant_training.html" ) else: self.batch_progress.reset_on_run() self.scheduler_progress.reset_on_run() self.batch_loop.optimizer_loop.optim_progress.reset_on_run() # when the epoch starts, the total val batch progress should be reset as it's supposed to count the batches # seen per epoch, this is useful for tracking when validation is run multiple times per epoch self.val_loop.epoch_loop.batch_progress.total.reset() self._outputs = [] def on_run_start( self, data_fetcher: AbstractDataFetcher ) -> None: # type: ignore[override] self._reload_dataloader_state_dict(data_fetcher) _ = iter(data_fetcher) # creates the iterator inside the fetcher # add the previous `fetched` value to properly track `is_last_batch` with no prefetching data_fetcher.fetched += self.batch_progress.current.ready def advance( self, data_fetcher: AbstractDataFetcher ) -> None: # type: ignore[override] """Runs a single training batch. Raises: StopIteration: When the epoch is canceled by the user returning -1 """ if self.restarting and self._should_check_val_fx( self.batch_idx, self.batch_progress.is_last_batch): # skip training and run validation in `on_advance_end` return # we are going to train first so the val loop does not need to restart self.val_loop.restarting = False if not isinstance(data_fetcher, DataLoaderIterDataFetcher): batch_idx = self.batch_idx + 1 batch = next(data_fetcher) else: batch_idx, batch = next(data_fetcher) self.batch_progress.is_last_batch = data_fetcher.done self.batch_progress.increment_ready() self.trainer._logger_connector.on_batch_start(batch, batch_idx) if batch is None: self._warning_cache.warn( "train_dataloader yielded None. If this was on purpose, ignore this warning..." ) batch_output = [] else: # hook self.trainer._call_callback_hooks("on_batch_start") # TODO: Update this in v1.7 (deprecation: #9816) model_fx = self.trainer.lightning_module.on_train_batch_start extra_kwargs = ({ "dataloader_idx": 0 } if callable(model_fx) and is_param_in_hook_signature( model_fx, "dataloader_idx", explicit=True) else {}) # hook self.trainer._call_callback_hooks("on_train_batch_start", batch, batch_idx, **extra_kwargs) response = self.trainer._call_lightning_module_hook( "on_train_batch_start", batch, batch_idx, **extra_kwargs) self.trainer._call_strategy_hook("on_train_batch_start", batch, batch_idx, **extra_kwargs) if response == -1: self.batch_progress.increment_processed() raise StopIteration self.batch_progress.increment_started() with self.trainer.profiler.profile("run_training_batch"): batch_output = self.batch_loop.run(batch, batch_idx) self.batch_progress.increment_processed() # update non-plateau LR schedulers # update epoch-interval ones only when we are at the end of training epoch self.update_lr_schedulers("step", update_plateau_schedulers=False) if self._num_ready_batches_reached(): self.update_lr_schedulers("epoch", update_plateau_schedulers=False) batch_end_outputs = self._prepare_outputs_training_batch_end( batch_output, lightning_module=self.trainer.lightning_module, num_optimizers=len(self.trainer.optimizers), ) # TODO: Update this in v1.7 (deprecation: #9816) model_fx = self.trainer.lightning_module.on_train_batch_end extra_kwargs = ({ "dataloader_idx": 0 } if callable(model_fx) and is_param_in_hook_signature( model_fx, "dataloader_idx", explicit=True) else {}) self.trainer._call_callback_hooks("on_train_batch_end", batch_end_outputs, batch, batch_idx, **extra_kwargs) self.trainer._call_lightning_module_hook("on_train_batch_end", batch_end_outputs, batch, batch_idx, **extra_kwargs) self.trainer._call_callback_hooks("on_batch_end") self.trainer._logger_connector.on_batch_end() self.batch_progress.increment_completed() if is_overridden("training_epoch_end", self.trainer.lightning_module): self._outputs.append(batch_output) # ----------------------------------------- # SAVE METRICS TO LOGGERS AND PROGRESS_BAR # ----------------------------------------- self.trainer._logger_connector.update_train_step_metrics() def on_advance_end(self) -> None: # ----------------------------------------- # VALIDATE IF NEEDED # ----------------------------------------- should_check_val = self._should_check_val_fx( self.batch_idx, self.batch_progress.is_last_batch) if should_check_val: self.trainer.validating = True self._run_validation() self.trainer.training = True # update plateau LR scheduler after metrics are logged self.update_lr_schedulers("step", update_plateau_schedulers=True) if not self._should_accumulate(): # this is increased once per batch disregarding multiple optimizers or tbptt on purpose for loggers self._batches_that_stepped += 1 # this will save based on the `batches_that_stepped` value self._save_loggers_on_train_batch_end() # if training finished, defer exit to the parent. this assumes there will be enough time in between # which might not be the case depending on what's in the `*_epoch_end` hooks if not self._is_training_done: # if fault tolerant is enabled and process has been notified, exit. self.trainer._exit_gracefully_on_signal() def on_run_end(self) -> _OUTPUTS_TYPE: outputs, self._outputs = self._outputs, [] return outputs def teardown(self) -> None: self._results.cpu() self.batch_loop.teardown() self.val_loop.teardown() def on_save_checkpoint(self) -> Dict: state_dict = super().on_save_checkpoint() if (self.trainer is not None and self.trainer.state._fault_tolerant_mode.is_enabled and self.trainer.train_dataloader is not None and not self._num_completed_batches_reached() # did not finish # TODO: fault-tolerance requires a minimum number of batches so probably should be > 0 and self.batch_progress.current.ready # did start ): loader: CombinedLoader = self.trainer.train_dataloader state = loader.state_dict(has_completed=self._has_completed()) if state: state_dict[ "dataloader_state_dict"] = _collect_states_on_rank_zero_over_collection( state) return state_dict def on_load_checkpoint(self, state_dict: Dict) -> None: # cache the dataloader state dict until the dataloader objects are available self._dataloader_state_dict = state_dict.get("dataloader_state_dict") def _run_validation(self) -> None: # reload dataloaders self.val_loop._reload_evaluation_dataloaders() with torch.no_grad(): self.val_loop.run() def _accumulated_batches_reached(self) -> bool: """Determine if accumulation will be finished by the end of the current batch.""" return self.batch_progress.current.ready % self.trainer.accumulate_grad_batches == 0 def _num_ready_batches_reached(self) -> bool: """Checks if we are in the last batch or if there are more batches to follow.""" epoch_finished_on_ready = self.batch_progress.current.ready == self.trainer.num_training_batches return epoch_finished_on_ready or self.batch_progress.is_last_batch def _num_completed_batches_reached(self) -> bool: epoch_finished_on_completed = self.batch_progress.current.completed == self.trainer.num_training_batches dataloader_consumed_successfully = self.batch_progress.is_last_batch and self._has_completed( ) return epoch_finished_on_completed or dataloader_consumed_successfully def _has_completed(self) -> bool: return self.batch_progress.current.ready == self.batch_progress.current.completed def _should_accumulate(self) -> bool: """Checks if the optimizer step should be performed or gradients should be accumulated for the current step.""" accumulation_done = self._accumulated_batches_reached() # Lightning steps on the final batch is_final_batch = self._num_ready_batches_reached() # but the strategy might not strategy_accumulates_on_final_batch = self.trainer.strategy.handles_gradient_accumulation or not is_final_batch return not accumulation_done and strategy_accumulates_on_final_batch @staticmethod def _prepare_outputs_training_batch_end( batch_output: _BATCH_OUTPUTS_TYPE, lightning_module: "pl.LightningModule", num_optimizers: int, ) -> Union[List[List[Dict[str, Any]]], List[Dict[str, Any]]]: """Processes the outputs from the batch loop into the format passed to the ``on_train_batch_end`` hook.""" if not batch_output: return [] # convert optimizer dicts to list if lightning_module.automatic_optimization: batch_output = apply_to_collection(batch_output, dtype=dict, function=_convert_optim_dict, num_optimizers=num_optimizers) array = np.array(batch_output, dtype=object) # TODO: remove in v1.8 if (num_optimizers > 1 and lightning_module.truncated_bptt_steps > 0 and not _v1_8_output_format(lightning_module.on_train_batch_end)): rank_zero_deprecation( "You are training with multiple optimizers AND truncated backpropagation through time enabled." " The current format of the `on_train_batch_end(outputs, ...)` is a 2d list with sizes" " (n_optimizers, tbptt_steps), however, this has been deprecated and will change in version v1.8 to" " (tbptt_steps, n_optimizers). You can update your code by adding the following parameter to your" " hook signature: `on_train_batch_end(outputs, ..., new_format=True)`." ) # (tbptt_steps, n_opt) -> (n_opt, tbptt_steps) if array.ndim == 1: array = np.expand_dims(array, 1) array = array.transpose((1, 0)) # squeeze all single-element dimensions array = array.squeeze() array = array.tolist() array = _recursive_unpad(array) return array @staticmethod def _prepare_outputs_training_epoch_end( batch_outputs: _OUTPUTS_TYPE, lightning_module: "pl.LightningModule", num_optimizers: int, ) -> Union[List[List[List[Dict[str, Any]]]], List[List[Dict[str, Any]]], List[Dict[str, Any]]]: """Processes the outputs from the batch loop into the format passed to the ``training_epoch_end`` hook.""" # `batch_outputs` (plural) is the same as `epoch_end_output` (singular) if not batch_outputs: return [] # convert optimizer dicts to list if lightning_module.automatic_optimization: batch_outputs = apply_to_collection(batch_outputs, dtype=dict, function=_convert_optim_dict, num_optimizers=num_optimizers) array = _recursive_pad(batch_outputs) # TODO: remove in v1.8 if (num_optimizers > 1 and lightning_module.truncated_bptt_steps > 0 and not _v1_8_output_format(lightning_module.on_train_epoch_end)): rank_zero_deprecation( "You are training with multiple optimizers AND truncated backpropagation through time enabled." " The current format of the `training_epoch_end(outputs)` is a 3d list with sizes" " (n_optimizers, n_batches, tbptt_steps), however, this has been deprecated and will change in version" " v1.8 to (n_batches, tbptt_steps, n_optimizers). You can update your code by adding the following" " parameter to your hook signature: `training_epoch_end(outputs, new_format=True)`." ) # (n_batches, tbptt_steps, n_opt) -> (n_opt, n_batches, tbptt_steps) if array.ndim == 2: array = np.expand_dims(array, 2) array = array.transpose((2, 0, 1)) # squeeze all single-element dimensions array = array.squeeze() array = array.tolist() array = _recursive_unpad(array) # in case we squeezed from 1-element array to a 0-dim array array = array if isinstance(array, list) else [array] # remove residual empty lists array = [ item for item in array if not isinstance(item, list) or len(item) ] return array def update_lr_schedulers(self, interval: str, update_plateau_schedulers: bool) -> None: """updates the lr schedulers based on the given interval.""" if interval == "step" and self._should_accumulate(): return active_optimizers = _get_active_optimizers( self.trainer.optimizers, self.trainer.optimizer_frequencies, self.total_batch_idx) self._update_learning_rates( interval=interval, update_plateau_schedulers=update_plateau_schedulers, opt_indices=[opt_idx for opt_idx, _ in active_optimizers], ) def _update_learning_rates( self, interval: str, update_plateau_schedulers: bool, opt_indices: Optional[List[int]] = None) -> None: """Update learning rates. Args: interval: either 'epoch' or 'step'. update_plateau_schedulers: control whether ``ReduceLROnPlateau`` or non-plateau schedulers get updated. This is used so non-plateau schedulers can be updated before running validation. Checkpoints are commonly saved during validation, however, on-plateau schedulers might monitor a validation metric so they have to be updated separately. opt_indices: indices of the optimizers to update. """ if not self.trainer.lr_scheduler_configs or not self.trainer.lightning_module.automatic_optimization: return if opt_indices is None: opt_indices = [] for config in self.trainer.lr_scheduler_configs: if config.opt_idx not in opt_indices: continue if update_plateau_schedulers ^ config.reduce_on_plateau: continue current_idx = self.batch_idx if interval == "step" else self.trainer.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 config.interval == interval and current_idx % config.frequency == 0: monitor_val = None if config.reduce_on_plateau: # If instance of ReduceLROnPlateau, we need a monitor monitor_key = config.monitor monitor_val = self._get_monitor_value(monitor_key) if monitor_val is None: if config.strict: avail_metrics = list(self.trainer.callback_metrics) 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" ) rank_zero_warn( f"ReduceLROnPlateau conditioned on metric {monitor_key}" " which is not available but strict is set to `False`." " Skipping learning rate update.", category=RuntimeWarning, ) continue self.scheduler_progress.increment_ready() # update LR self.trainer._call_lightning_module_hook( "lr_scheduler_step", config.scheduler, config.opt_idx, monitor_val, ) self.scheduler_progress.increment_completed() def _get_monitor_value(self, key: str) -> Any: # this is a separate method to aid in testing return self.trainer.callback_metrics.get(key) def _should_check_val_epoch(self): return (self.trainer.enable_validation and (self.trainer.current_epoch + 1) % self.trainer.check_val_every_n_epoch == 0) def _should_check_val_fx(self, batch_idx: int, is_last_batch: bool) -> bool: """Decide if we should run validation.""" if not self._should_check_val_epoch(): return False # val_check_batch is inf for iterable datasets with no length defined is_infinite_dataset = self.trainer.val_check_batch == float("inf") if is_last_batch and is_infinite_dataset: return True if self.trainer.should_stop: return True # TODO(@awaelchli): let training/eval loop handle logic around limit_*_batches and val_check_batch is_val_check_batch = is_last_batch if isinstance(self.trainer.limit_train_batches, int) and is_infinite_dataset: is_val_check_batch = (batch_idx + 1) % self.trainer.limit_train_batches == 0 elif self.trainer.val_check_batch != float("inf"): is_val_check_batch = (batch_idx + 1) % self.trainer.val_check_batch == 0 return is_val_check_batch def _save_loggers_on_train_batch_end(self) -> None: """Flushes loggers to disk.""" # this assumes that `batches_that_stepped` was increased before should_flush = self._batches_that_stepped % self.trainer.flush_logs_every_n_steps == 0 if should_flush or self.trainer.should_stop: for logger in self.trainer.loggers: logger.save() def _reload_dataloader_state_dict( self, data_fetcher: AbstractDataFetcher) -> None: if self._dataloader_state_dict: data_fetcher.dataloader.load_state_dict( self._dataloader_state_dict) self._dataloader_state_dict = None
class TrainingEpochLoop(loops.Loop): """ Runs over all batches in a dataloader (one epoch). """ def __init__(self, min_steps: int, max_steps: int): super().__init__() self.min_steps: int = min_steps self.max_steps: int = max_steps self.global_step: int = 0 # the total batch index across all epochs self.total_batch_idx: int = 0 # the current batch index in the loop that runs over the dataloader(s) self.iteration_count: int = 0 # the current split index when the batch gets split into chunks in truncated backprop through time self.split_idx: Optional[int] = None # the number of batches seen this run, updates immediately after batch_loop.run() self.batches_seen: int = 0 self.is_last_batch: Optional[bool] = None self.batch_loop: Optional[TrainingBatchLoop] = None self.val_loop: Optional[loops.EvaluationLoop] = None self._dataloader_idx: Optional[int] = None self._warning_cache: WarningCache = WarningCache() self._epoch_output: Optional[List[List[STEP_OUTPUT]]] = None self._results = ResultCollection(training=True) @property def results(self) -> ResultCollection: if self.trainer.training: return self._results elif self.trainer.validating: return self.val_loop.results raise RuntimeError("`FitLoop.results` property isn't defined. Accessed outside of scope") @property def batch_idx(self) -> int: """Returns the current batch index (within this epoch)""" return self.iteration_count @property def done(self) -> bool: """Returns whether the training should be stopped. The criteria are that the number of steps reached the max steps, the last batch is reached or the trainer signals to stop (e.g. by early stopping). """ max_steps_reached = self.max_steps is not None and self.global_step >= self.max_steps return max_steps_reached or self.trainer.should_stop or self._num_training_batches_reached(self.is_last_batch) def connect(self, trainer: 'pl.Trainer', *args: Any, **kwargs: Any) -> None: """Connects the loop with all necessary parts like trainer and accelerators""" super().connect(trainer, *args, **kwargs) self.batch_loop = TrainingBatchLoop() self.batch_loop.connect(trainer) self.val_loop = loops.EvaluationLoop() self.val_loop.connect(trainer) def reset(self) -> None: """Resets the internal state of the loop for a new run""" self.iteration_count = 0 self.batches_seen = 0 self.is_last_batch = False self._dataloader_idx = 0 # track epoch output self._epoch_output = [[] for _ in range(self.batch_loop.num_active_optimizers(self.total_batch_idx))] def on_run_start(self, *args: Any, **kwargs: Any) -> None: # hook self.trainer.logger_connector.on_epoch_start() self.trainer.call_hook("on_epoch_start") self.trainer.call_hook("on_train_epoch_start") def advance(self, dataloader_iter: Iterator, **kwargs: Any) -> None: """Runs a single training batch. Args: dataloader_iter: the iterator over the dataloader producing the new batch Raises: StopIteration: When the epoch is canceled by the user returning -1 """ _, (batch, is_last) = next(dataloader_iter) self.is_last_batch = is_last # ------------------------------------ # TRAINING_STEP + TRAINING_STEP_END # ------------------------------------ with self.trainer.profiler.profile("run_training_batch"): batch_output = self.batch_loop.run(batch, self.iteration_count, self._dataloader_idx) self.batches_seen += 1 # when returning -1 from train_step, we end epoch early if batch_output.signal == -1: raise StopIteration # update non-plateau LR schedulers # update epoch-interval ones only when we are at the end of training epoch self.update_lr_schedulers('step', update_plateau_schedulers=False) if self._num_training_batches_reached(is_last): self.update_lr_schedulers('epoch', update_plateau_schedulers=False) batch_end_outputs = [opt_idx_out for opt_idx_out in batch_output.training_step_output if len(opt_idx_out)] processed_batch_end_outputs = self._prepare_outputs(batch_end_outputs, batch_mode=True) # hook self.trainer.call_hook( 'on_train_batch_end', processed_batch_end_outputs, batch, self.iteration_count, self._dataloader_idx ) self.trainer.call_hook('on_batch_end') self.trainer.logger_connector.on_batch_end() # figure out what to track for epoch end self._track_epoch_end_reduce_metrics(self._epoch_output, batch_end_outputs) # ----------------------------------------- # SAVE METRICS TO LOGGERS AND PROGRESS_BAR # ----------------------------------------- self.trainer.logger_connector.update_train_step_metrics() def on_advance_end(self): """Runs validation and Checkpointing if necessary. Raises: StopIteration: if :attr:`done` evaluates to ``True`` to finish this epoch """ # ----------------------------------------- # VALIDATE IF NEEDED + CHECKPOINT CALLBACK # ----------------------------------------- should_check_val = self._should_check_val_fx(self.iteration_count, self.is_last_batch) if should_check_val: self.trainer.validating = True self._run_validation() self.trainer.training = True # ----------------------------------------- # SAVE LOGGERS (ie: Tensorboard, etc...) # ----------------------------------------- self._save_loggers_on_train_batch_end() # update plateau LR scheduler after metrics are logged self.update_lr_schedulers('step', update_plateau_schedulers=True) self.trainer.checkpoint_connector.has_trained = True self.total_batch_idx += 1 # progress global step according to grads progress self._increment_accumulated_grad_global_step() if self.done: raise StopIteration def on_run_end(self) -> List[List[STEP_OUTPUT]]: """Calls the on_epoch_end hook. Returns: The output of each training step for each optimizer Raises: MisconfigurationException: ``train_epoch_end`` does not return ``None`` """ if self.batches_seen == 0: # dataloader/iterator did not produce a batch return # inform logger the batch loop has finished self.trainer.logger_connector.epoch_end_reached() # prepare epoch output processed_outputs = self._prepare_outputs(self._epoch_output, batch_mode=False) # get the model and call model.training_epoch_end model = self.trainer.lightning_module if is_overridden('training_epoch_end', model): # run training_epoch_end # refresh the result for custom logging at the epoch level model._current_fx_name = 'training_epoch_end' # lightningmodule hook training_epoch_end_output = model.training_epoch_end(processed_outputs) if training_epoch_end_output is not None: raise MisconfigurationException( 'training_epoch_end expects a return of None. ' 'HINT: remove the return statement in training_epoch_end' ) # call train epoch end hooks self._on_train_epoch_end_hook(processed_outputs) self.trainer.call_hook('on_epoch_end') self.trainer.logger_connector.on_epoch_end() return self._epoch_output def teardown(self) -> None: """Frees memory of tracked epoch outputs.""" self.epoch_output = None def _run_validation(self): # reload dataloaders self.val_loop.reload_evaluation_dataloaders() with torch.no_grad(): self.val_loop.run() def _on_train_epoch_end_hook(self, processed_epoch_output: List[List[STEP_OUTPUT]]) -> None: """Runs ``on_train_epoch_end hook``.""" # We cannot rely on Trainer.call_hook because the signatures might be different across # lightning module and callback # As a result, we need to inspect if the module accepts `outputs` in `on_train_epoch_end` # This implementation is copied from Trainer.call_hook hook_name = "on_train_epoch_end" prev_fx_name = self.trainer.lightning_module._current_fx_name self.trainer.lightning_module._current_fx_name = hook_name # always profile hooks with self.trainer.profiler.profile(hook_name): # first call trainer hook if hasattr(self.trainer, hook_name): trainer_hook = getattr(self.trainer, hook_name) trainer_hook(processed_epoch_output) # next call hook in lightningModule model_ref = self.trainer.lightning_module if is_overridden(hook_name, model_ref): hook_fx = getattr(model_ref, hook_name) if is_param_in_hook_signature(hook_fx, "outputs"): self._warning_cache.deprecation( "The signature of `ModelHooks.on_train_epoch_end` has changed in v1.3." " `outputs` parameter has been deprecated." " Support for the old signature will be removed in v1.5", ) model_ref.on_train_epoch_end(processed_epoch_output) else: model_ref.on_train_epoch_end() # call the accelerator hook if hasattr(self.trainer.accelerator, hook_name): accelerator_hook = getattr(self.trainer.accelerator, hook_name) accelerator_hook() # restore current_fx when nested context self.trainer.lightning_module._current_fx_name = prev_fx_name def _num_training_batches_reached(self, is_last_batch: bool = False) -> bool: """Checks if we are in the last batch or if there are more batches to follow.""" # TODO: Can we combine this with training_batch_loop's arg that does a similar check? return self.batches_seen == self.trainer.num_training_batches or is_last_batch def _track_epoch_end_reduce_metrics( self, epoch_output: List[List[STEP_OUTPUT]], batch_end_outputs: STEP_OUTPUT ) -> None: """Adds the batch outputs to the epoch outputs and prepares reduction""" hook_overridden = self._should_add_batch_output_to_epoch_output() if not hook_overridden: return # track the outputs to reduce at the end of the epoch for opt_idx, opt_outputs in enumerate(batch_end_outputs): # with 1 step (no tbptt) don't use a sequence at epoch end if ( isinstance(opt_outputs, list) and len(opt_outputs) == 1 and not isinstance(opt_outputs[0], ResultCollection) ): opt_outputs = opt_outputs[0] epoch_output[opt_idx].append(opt_outputs) def _should_add_batch_output_to_epoch_output(self) -> bool: """ We add to the epoch outputs if 1. The model defines training_epoch_end OR 2. The model overrides on_train_epoch_end which has `outputs` in the signature """ # TODO: in v1.5 this only needs to check if training_epoch_end is overridden lightning_module = self.trainer.lightning_module if is_overridden("training_epoch_end", lightning_module): return True if is_overridden("on_train_epoch_end", lightning_module): model_hook_fx = getattr(lightning_module, "on_train_epoch_end") if is_param_in_hook_signature(model_hook_fx, "outputs"): return True return False @staticmethod def _prepare_outputs( outputs: List[List[List['ResultCollection']]], batch_mode: bool, ) -> Union[List[List[List[Dict]]], List[List[Dict]], List[Dict], Dict]: """ Extract required information from batch or epoch end results. Args: outputs: A 3-dimensional list of ``ResultCollection`` objects with dimensions: ``[optimizer outs][batch outs][tbptt steps]``. batch_mode: If True, ignore the batch output dimension. Returns: The cleaned outputs with ``ResultCollection`` objects converted to dictionaries. All list dimensions of size one will be collapsed. """ processed_outputs = [] for opt_outputs in outputs: # handle an edge case where an optimizer output is the empty list if len(opt_outputs) == 0: continue processed_batch_outputs = [] if batch_mode: opt_outputs = [opt_outputs] for batch_outputs in opt_outputs: processed_tbptt_outputs = [] if isinstance(batch_outputs, ResultCollection): batch_outputs = [batch_outputs] for tbptt_output in batch_outputs: out = tbptt_output.extra if tbptt_output.minimize is not None: out['loss'] = tbptt_output.minimize.detach() processed_tbptt_outputs.append(out) # if there was only one tbptt step then we can collapse that dimension if len(processed_tbptt_outputs) == 1: processed_tbptt_outputs = processed_tbptt_outputs[0] processed_batch_outputs.append(processed_tbptt_outputs) # batch_outputs should be just one dict (or a list of dicts if using tbptt) per optimizer if batch_mode: processed_batch_outputs = processed_batch_outputs[0] processed_outputs.append(processed_batch_outputs) # if there is only one optimiser then we collapse that dimension if len(processed_outputs) == 1: processed_outputs = processed_outputs[0] return processed_outputs def update_lr_schedulers(self, interval: str, update_plateau_schedulers: bool) -> None: """updates the lr schedulers based on the given interval""" if interval == "step" and self.batch_loop.should_accumulate(): return self.trainer.optimizer_connector.update_learning_rates( interval=interval, update_plateau_schedulers=update_plateau_schedulers, opt_indices=[opt_idx for opt_idx, _ in self.batch_loop.get_active_optimizers(self.total_batch_idx)], ) def _increment_accumulated_grad_global_step(self) -> None: """increments global step""" num_accumulated_batches_reached = self.batch_loop._accumulated_batches_reached() num_training_batches_reached = self._num_training_batches_reached() # progress global step according to grads progress if num_accumulated_batches_reached or num_training_batches_reached: self.global_step = self.trainer.accelerator.update_global_step( self.total_batch_idx, self.trainer.global_step ) def _should_check_val_fx(self, batch_idx: int, is_last_batch: bool) -> bool: """ Decide if we should run validation. """ if not self.trainer.enable_validation: return False is_val_check_epoch = (self.trainer.current_epoch + 1) % self.trainer.check_val_every_n_epoch == 0 if not is_val_check_epoch: return False # val_check_batch is inf for iterable datasets with no length defined is_infinite_dataset = self.trainer.val_check_batch == float('inf') if is_last_batch and is_infinite_dataset: return True if self.trainer.should_stop: return True # TODO(@awaelchli): let training/eval loop handle logic around limit_*_batches and val_check_batch is_val_check_batch = is_last_batch if isinstance(self.trainer.limit_train_batches, int) and is_infinite_dataset: is_val_check_batch = (batch_idx + 1) % self.trainer.limit_train_batches == 0 elif self.trainer.val_check_batch != float('inf'): is_val_check_batch = (batch_idx + 1) % self.trainer.val_check_batch == 0 return is_val_check_batch def _save_loggers_on_train_batch_end(self) -> None: """Flushes loggers to disk""" # when loggers should save to disk should_flush_logs = self.trainer.logger_connector.should_flush_logs if should_flush_logs and self.trainer.is_global_zero and self.trainer.logger is not None: self.trainer.logger.save()
class TrainingEpochLoop(loops.Loop[_OUTPUTS_TYPE]): """Runs over all batches in a dataloader (one epoch). Args: min_steps: The minimum number of steps (batches) to process max_steps: The maximum number of steps (batches) to process """ def __init__(self, min_steps: Optional[int] = None, max_steps: int = -1) -> None: super().__init__() if max_steps is None: rank_zero_deprecation( "Setting `max_steps = None` is deprecated in v1.5 and will no longer be supported in v1.7." " Use `max_steps = -1` instead.") max_steps = -1 elif max_steps < -1: raise MisconfigurationException( f"`max_steps` must be a non-negative integer or -1 (infinite steps). You passed in {max_steps}." ) self.min_steps = min_steps self.max_steps = max_steps self.global_step: int = 0 self.batch_progress = BatchProgress() self.scheduler_progress = SchedulerProgress() self.batch_loop = TrainingBatchLoop() self.val_loop = loops.EvaluationLoop(verbose=False) self._results = _ResultCollection(training=True) self._outputs: _OUTPUTS_TYPE = [] self._warning_cache = WarningCache() self._dataloader_iter: Optional[Iterator] = None # caches the loaded dataloader state until dataloader objects are available self._dataloader_state_dict: Dict[str, Any] = {} @property def total_batch_idx(self) -> int: """Returns the current batch index (across epochs)""" # use `ready` instead of `completed` in case this is accessed after `completed` has been increased # but before the next `ready` increase return self.batch_progress.total.ready - 1 @property def batch_idx(self) -> int: """Returns the current batch index (within this epoch)""" # use `ready` instead of `completed` in case this is accessed after `completed` has been increased # but before the next `ready` increase return self.batch_progress.current.ready - 1 @property def _is_training_done(self) -> bool: max_steps_reached = _is_max_limit_reached(self.global_step, self.max_steps) return max_steps_reached or self._num_ready_batches_reached() @property def _is_validation_done(self) -> bool: # when we are restarting we want to check whether the val loop has finished return not self.restarting or self.val_loop.done @property def done(self) -> bool: """Returns whether the training should be stopped. The criteria are that the number of steps reached the max steps, the last batch is reached or the trainer signals to stop (e.g. by early stopping). """ return (self._is_training_done and self._is_validation_done) or self.trainer.should_stop def connect( # type: ignore[override] self, batch_loop: Optional[TrainingBatchLoop] = None, val_loop: Optional["loops.EvaluationLoop"] = None, ) -> None: """Optionally connect a custom batch or validation loop to this training epoch loop.""" if batch_loop is not None: self.batch_loop = batch_loop if val_loop is not None: self.val_loop = val_loop def reset(self) -> None: """Resets the internal state of the loop for a new run.""" if self.restarting: self.batch_progress.reset_on_restart() self.scheduler_progress.reset_on_restart() self.batch_loop.optimizer_loop.optim_progress.reset_on_restart() else: self.batch_progress.reset_on_run() self.scheduler_progress.reset_on_run() self.batch_loop.optimizer_loop.optim_progress.reset_on_run() self._outputs = [] def on_run_start( self, data_fetcher: AbstractDataFetcher ) -> None: # type: ignore[override] # hook self.trainer.logger_connector.on_epoch_start() self.trainer._call_callback_hooks("on_epoch_start") self.trainer._call_lightning_module_hook("on_epoch_start") self.trainer._call_callback_hooks("on_train_epoch_start") self.trainer._call_lightning_module_hook("on_train_epoch_start") self.trainer.fit_loop.epoch_progress.increment_started() self._reload_dataloader_state_dict(data_fetcher) self._dataloader_iter = _update_dataloader_iter( data_fetcher, self.batch_idx + 1) def advance( self, data_fetcher: AbstractDataFetcher ) -> None: # type: ignore[override] """Runs a single training batch. Raises: StopIteration: When the epoch is canceled by the user returning -1 """ if self.restarting and self._should_check_val_fx( self.batch_idx, self.batch_progress.is_last_batch): # skip training and run validation in `on_advance_end` return assert self._dataloader_iter is not None batch_idx, (batch, self.batch_progress.is_last_batch) = next( self._dataloader_iter) if not data_fetcher.store_on_device: batch = self.trainer._call_strategy_hook("batch_to_device", batch) self.batch_progress.increment_ready() self.trainer.logger_connector.on_batch_start(batch, batch_idx) if batch is None: self._warning_cache.warn( "train_dataloader yielded None. If this was on purpose, ignore this warning..." ) batch_output = [] else: # hook self.trainer._call_callback_hooks("on_batch_start") # TODO: Update this in v1.7 (deprecation: #9816) model_fx = self.trainer.lightning_module.on_train_batch_start extra_kwargs = ({ "dataloader_idx": 0 } if callable(model_fx) and is_param_in_hook_signature( model_fx, "dataloader_idx", explicit=True) else {}) # hook self.trainer._call_callback_hooks("on_train_batch_start", batch, batch_idx, **extra_kwargs) response = self.trainer._call_lightning_module_hook( "on_train_batch_start", batch, batch_idx, **extra_kwargs) self.trainer._call_strategy_hook("on_train_batch_start", batch, batch_idx, **extra_kwargs) if response == -1: self.batch_progress.increment_processed() raise StopIteration self.batch_progress.increment_started() with self.trainer.profiler.profile("run_training_batch"): batch_output = self.batch_loop.run(batch, batch_idx) self.batch_progress.increment_processed() # update non-plateau LR schedulers # update epoch-interval ones only when we are at the end of training epoch self.update_lr_schedulers("step", update_plateau_schedulers=False) if self._num_ready_batches_reached(): self.update_lr_schedulers("epoch", update_plateau_schedulers=False) batch_end_outputs = self._prepare_outputs_training_batch_end( batch_output, automatic=self.trainer.lightning_module.trainer.lightning_module. automatic_optimization, num_optimizers=len(self.trainer.optimizers), ) # TODO: Update this in v1.7 (deprecation: #9816) model_fx = self.trainer.lightning_module.on_train_batch_end extra_kwargs = ({ "dataloader_idx": 0 } if callable(model_fx) and is_param_in_hook_signature( model_fx, "dataloader_idx", explicit=True) else {}) self.trainer._call_callback_hooks("on_train_batch_end", batch_end_outputs, batch, batch_idx, **extra_kwargs) self.trainer._call_lightning_module_hook("on_train_batch_end", batch_end_outputs, batch, batch_idx, **extra_kwargs) self.trainer._call_callback_hooks("on_batch_end") self.trainer.logger_connector.on_batch_end() self.batch_progress.increment_completed() if is_overridden("training_epoch_end", self.trainer.lightning_module): self._outputs.append(batch_output) # ----------------------------------------- # SAVE METRICS TO LOGGERS AND PROGRESS_BAR # ----------------------------------------- self.trainer.logger_connector.update_train_step_metrics() def on_advance_end(self) -> None: """Runs validation and Checkpointing if necessary. Raises: StopIteration: if :attr:`done` evaluates to ``True`` to finish this epoch """ # ----------------------------------------- # VALIDATE IF NEEDED + CHECKPOINT CALLBACK # ----------------------------------------- should_check_val = self._should_check_val_fx( self.batch_idx, self.batch_progress.is_last_batch) if should_check_val: self.trainer.validating = True self._run_validation() self.trainer.training = True # ----------------------------------------- # SAVE LOGGERS (ie: Tensorboard, etc...) # ----------------------------------------- self._save_loggers_on_train_batch_end() # update plateau LR scheduler after metrics are logged self.update_lr_schedulers("step", update_plateau_schedulers=True) if not self._should_accumulate(): # progress global step according to grads progress self.global_step += 1 # if training finished, try to exit in `on_run_end` instead as we should have enough time # TODO: @tchaton verify this assumption is True. if not self._is_training_done: # if fault tolerant is enabled and process has been notified, exit. self.trainer._exit_gracefully_on_signal() def on_run_end(self) -> None: """Calls the on_epoch_end hook. Returns: The output of each training step for each optimizer Raises: MisconfigurationException: ``train_epoch_end`` does not return ``None`` """ # inform logger the batch loop has finished self.trainer.logger_connector.epoch_end_reached() # get the model and call model.training_epoch_end model = self.trainer.lightning_module if is_overridden("training_epoch_end", model) and self._outputs: epoch_end_outputs = self._prepare_outputs_training_epoch_end( self._outputs, automatic=model.automatic_optimization, num_optimizers=len(self.trainer.optimizers), ) # run lightning module hook training_epoch_end # refresh the result for custom logging at the epoch level epoch_end_outputs = self.trainer._call_lightning_module_hook( "training_epoch_end", epoch_end_outputs) if epoch_end_outputs is not None: raise MisconfigurationException( "`training_epoch_end` expects a return of None. " "HINT: remove the return statement in `training_epoch_end`." ) # free memory self._outputs = [] self.trainer.fit_loop.epoch_progress.increment_processed() # call train epoch end hooks self.trainer._call_callback_hooks("on_train_epoch_end") self.trainer._call_lightning_module_hook("on_train_epoch_end") self.trainer._call_callback_hooks("on_epoch_end") self.trainer._call_lightning_module_hook("on_epoch_end") self.trainer.logger_connector.on_epoch_end() if self._num_ready_batches_reached(): self.update_lr_schedulers("epoch", update_plateau_schedulers=True) # if fault tolerant is enabled and process has been notified, exit. self.trainer._exit_gracefully_on_signal() def teardown(self) -> None: self._results.cpu() self.batch_loop.teardown() self.val_loop.teardown() def on_save_checkpoint(self) -> Dict: state_dict = super().on_save_checkpoint() if (self.trainer.train_dataloader is None or self._num_completed_batches_reached() # did not finish # TODO: fault-tolerance requires a minimum number of batches so probably should be > 0 or self.batch_progress.current.ready == 0 # did not start ): return state_dict state_dict[ "dataloader_state_dict"] = _collect_states_on_rank_zero_over_collection( self.trainer.train_dataloader.state_dict( has_completed=self._has_completed())) return state_dict def on_load_checkpoint(self, state_dict: Dict) -> None: # cache the dataloader state dict until the dataloader objects are available self._dataloader_state_dict = state_dict.get("dataloader_state_dict") def _run_validation(self) -> None: # reload dataloaders self.val_loop._reload_evaluation_dataloaders() with torch.no_grad(): self.val_loop.run() def _accumulated_batches_reached(self) -> bool: """Determine if accumulation will be finished by the end of the current batch.""" return self.batch_progress.current.ready % self.trainer.accumulate_grad_batches == 0 def _num_ready_batches_reached(self) -> bool: """Checks if we are in the last batch or if there are more batches to follow.""" epoch_finished_on_ready = self.batch_progress.current.ready == self.trainer.num_training_batches return epoch_finished_on_ready or self.batch_progress.is_last_batch def _num_completed_batches_reached(self) -> bool: epoch_finished_on_completed = self.batch_progress.current.completed == self.trainer.num_training_batches dataloader_consumed_successfully = self.batch_progress.is_last_batch and self._has_completed( ) return epoch_finished_on_completed or dataloader_consumed_successfully def _has_completed(self) -> bool: return self.batch_progress.current.ready == self.batch_progress.current.completed def _should_accumulate(self) -> bool: """Checks if the optimizer step should be performed or gradients should be accumulated for the current step.""" accumulation_done = self._accumulated_batches_reached() # Lightning steps on the final batch is_final_batch = self._num_ready_batches_reached() # but the TTP might not ttp_accumulates_on_final_batch = self.trainer.strategy.handles_gradient_accumulation or not is_final_batch return not accumulation_done and ttp_accumulates_on_final_batch @staticmethod def _prepare_outputs_training_batch_end( batch_output: _BATCH_OUTPUTS_TYPE, automatic: bool, num_optimizers: int, ) -> Union[List[List[Dict[str, Any]]], List[Dict[str, Any]]]: """Processes the outputs from the batch loop into the format passed to the ``training_batch_end`` hook. ``(tbptt_steps, n_opt) -> (n_opt, tbptt_steps)``. The optimizer dimension might have been squeezed. """ if not batch_output: return [] # convert optimizer dicts to list if automatic: batch_output = apply_to_collection(batch_output, dtype=dict, function=_convert_optim_dict, num_optimizers=num_optimizers) array = np.array(batch_output, dtype=object) if array.ndim == 1: array = np.expand_dims(array, 1) array = array.transpose((1, 0)) array = array.squeeze() array = array.tolist() array = _recursive_unpad(array) return array @staticmethod def _prepare_outputs_training_epoch_end( batch_outputs: _OUTPUTS_TYPE, automatic: bool, num_optimizers: int, ) -> Union[List[List[List[Dict[str, Any]]]], List[List[Dict[str, Any]]], List[Dict[str, Any]]]: """Processes the outputs from the batch loop into the format passed to the ``training_epoch_end`` hook. ``(n_batches, tbptt_steps, n_opt) -> (n_opt, n_batches, tbptt_steps)``. All single-element dimensions might have been squeezed. This processing is necessary because the format of the inputs to the ``training_epoch_end`` hook does not match the loop structure and because empty dimensions are squeezed. This could break with loop customization. """ # `batch_outputs` (plural) is the same as `epoch_end_output` (singular) if not batch_outputs: return [] # convert optimizer dicts to list if automatic: batch_outputs = apply_to_collection(batch_outputs, dtype=dict, function=_convert_optim_dict, num_optimizers=num_optimizers) array = _recursive_pad(batch_outputs) if array.ndim == 2: array = np.expand_dims(array, 2) array = array.transpose((2, 0, 1)) array = array.squeeze() array = array.tolist() array = _recursive_unpad(array) # in case we squeezed from 1-element array to a 0-dim array array = array if isinstance(array, list) else [array] # remove residual empty lists array = [ item for item in array if not isinstance(item, list) or len(item) ] return array def update_lr_schedulers(self, interval: str, update_plateau_schedulers: bool) -> None: """updates the lr schedulers based on the given interval.""" if interval == "step" and self._should_accumulate(): return active_optimizers = _get_active_optimizers( self.trainer.optimizers, self.trainer.optimizer_frequencies, self.total_batch_idx) self._update_learning_rates( interval=interval, update_plateau_schedulers=update_plateau_schedulers, opt_indices=[opt_idx for opt_idx, _ in active_optimizers], ) def _update_learning_rates( self, interval: str, update_plateau_schedulers: bool, opt_indices: Optional[List[int]] = None) -> None: """Update learning rates. Args: interval: either 'epoch' or 'step'. update_plateau_schedulers: control whether ``ReduceLROnPlateau`` or non-plateau schedulers get updated. This is used so non-plateau schedulers can be updated before running validation. Checkpoints are commonly saved during validation, however, on-plateau schedulers might monitor a validation metric so they have to be updated separately. opt_indices: indices of the optimizers to update. """ if not self.trainer.lr_schedulers or not self.trainer.lightning_module.automatic_optimization: return if opt_indices is None: opt_indices = [] for lr_scheduler in self.trainer.lr_schedulers: if isinstance(lr_scheduler["opt_idx"], int) and lr_scheduler["opt_idx"] not in opt_indices: continue if update_plateau_schedulers ^ lr_scheduler["reduce_on_plateau"]: continue current_idx = self.batch_idx if interval == "step" else self.trainer.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: monitor_val = None if lr_scheduler["reduce_on_plateau"]: # If instance of ReduceLROnPlateau, we need a monitor monitor_key = lr_scheduler["monitor"] monitor_val = self._get_monitor_value(monitor_key) if monitor_val is None: if lr_scheduler.get("strict", True): avail_metrics = list(self.trainer.callback_metrics) 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" ) rank_zero_warn( f"ReduceLROnPlateau conditioned on metric {monitor_key}" " which is not available but strict is set to `False`." " Skipping learning rate update.", category=RuntimeWarning, ) continue self.scheduler_progress.increment_ready() # update LR if lr_scheduler["reduce_on_plateau"]: lr_scheduler["scheduler"].step(monitor_val) else: lr_scheduler["scheduler"].step() self.scheduler_progress.increment_completed() def _get_monitor_value(self, key: str) -> Any: # this is a separate method to aid in testing return self.trainer.callback_metrics.get(key) def _should_check_val_fx(self, batch_idx: int, is_last_batch: bool) -> bool: """Decide if we should run validation.""" if not self.trainer.enable_validation: return False is_val_check_epoch = (self.trainer.current_epoch + 1) % self.trainer.check_val_every_n_epoch == 0 if not is_val_check_epoch: return False # val_check_batch is inf for iterable datasets with no length defined is_infinite_dataset = self.trainer.val_check_batch == float("inf") if is_last_batch and is_infinite_dataset: return True if self.trainer.should_stop: return True # TODO(@awaelchli): let training/eval loop handle logic around limit_*_batches and val_check_batch is_val_check_batch = is_last_batch if isinstance(self.trainer.limit_train_batches, int) and is_infinite_dataset: is_val_check_batch = (batch_idx + 1) % self.trainer.limit_train_batches == 0 elif self.trainer.val_check_batch != float("inf"): is_val_check_batch = (batch_idx + 1) % self.trainer.val_check_batch == 0 return is_val_check_batch def _save_loggers_on_train_batch_end(self) -> None: """Flushes loggers to disk.""" # when loggers should save to disk should_flush_logs = self.trainer.logger_connector.should_flush_logs if should_flush_logs and self.trainer.is_global_zero and self.trainer.logger is not None: self.trainer.logger.save() def _reload_dataloader_state_dict( self, data_fetcher: AbstractDataFetcher) -> None: if self._dataloader_state_dict: data_fetcher.dataloader.load_state_dict( self._dataloader_state_dict) self._dataloader_state_dict = None