class TrainLoop:
def __init__(
self,
trainer,
max_epochs: Optional[int],
min_epochs: Optional[int],
max_steps: Optional[int],
min_steps: Optional[int],
num_sanity_val_steps: int,
):
self.trainer = trainer
self.accumulated_loss = None
self.warning_cache = WarningCache()
self.running_loss = TensorRunningAccum(window_length=20)
self._skip_backward = False
self._optimizer_freq_cumsum = None
self._hiddens = None
self.global_step = 0
self.current_epoch = 0
self.trainer.should_stop = False
# the total batch index across all epochs
self.total_batch_idx = 0
# the current batch index in the loop that runs over the dataloader(s)
self.batch_idx = 0
# the current split index when the batch gets split into chunks in truncated backprop through time
self.split_idx = None
self.trainer.num_training_batches = 0
self.trainer.train_dataloader = None
# If neither max_epochs or max_steps is set, then use existing default of max_epochs = 1000
self.max_epochs = 1000 if (max_epochs is None and max_steps is None) else max_epochs
# If neither min_epochs or min_steps is set, then use existing default of min_epochs = 1
self.min_epochs = 1 if (min_epochs is None and min_steps is None) else min_epochs
self.max_steps = max_steps
self.min_steps = min_steps
if num_sanity_val_steps == -1:
self.trainer.num_sanity_val_steps = float("inf")
else:
self.trainer.num_sanity_val_steps = num_sanity_val_steps
self.results = ResultCollection(training=True)
@property
def num_active_optimizers(self) -> int:
return len(self.get_active_optimizers())
@property
def optimizer_freq_cumsum(self):
if self._optimizer_freq_cumsum is None:
self._optimizer_freq_cumsum = np.cumsum(self.trainer.optimizer_frequencies)
return self._optimizer_freq_cumsum
def should_skip_training(self) -> bool:
should_by_max_steps = self.max_steps is not None and self.global_step >= self.max_steps
should_by_epoch = self.max_epochs is not None and self.current_epoch >= self.max_epochs
return should_by_max_steps or should_by_epoch or self.trainer.num_training_batches == 0
def on_train_start(self):
self.results.to(device=self.trainer.lightning_module.device)
self.trainer.call_hook("on_train_start")
def on_train_end(self):
# trigger checkpoint check. need to temporarily decrease the global step to avoid saving duplicates
# when a checkpoint was saved at the last step
self.global_step -= 1
self.check_checkpoint_callback(should_update=True, is_last=True)
self.global_step += 1
# hook
self.trainer.call_hook("on_train_end")
# todo: TPU 8 cores hangs in flush with TensorBoard. Might do for all loggers.
# It might be related to xla tensors blocked when moving the cpu
# kill loggers
if self.trainer.logger is not None:
self.trainer.logger.finalize("success")
# summarize profile results
self.trainer.profiler.describe()
# give accelerators a chance to finish
self.trainer.accelerator.on_train_end()
# reset bookkeeping
self.trainer.state.stage = None
def check_checkpoint_callback(self, should_update, is_last=False):
# TODO bake this logic into the ModelCheckpoint callback
if should_update and self.trainer.checkpoint_connector.has_trained:
callbacks = self.trainer.checkpoint_callbacks
if is_last and any(cb.save_last and cb.verbose for cb in callbacks):
rank_zero_info("Saving latest checkpoint...")
model = self.trainer.lightning_module
for cb in callbacks:
cb.on_validation_end(self.trainer, model)
def on_train_epoch_start(self, epoch):
# update training progress in trainer
self.current_epoch = epoch
model = self.trainer.lightning_module
# reset train dataloader
if epoch != 0 and self.trainer.reload_dataloaders_every_epoch:
self.trainer.reset_train_dataloader(model)
# todo: specify the possible exception
with suppress(Exception):
# set seed for distributed sampler (enables shuffling for each epoch)
self.trainer.train_dataloader.sampler.set_epoch(epoch)
# changing gradient according accumulation_scheduler
self.trainer.accumulation_scheduler.on_train_epoch_start(self.trainer, self.trainer.lightning_module)
# stores accumulated grad fractions per batch
self.accumulated_loss = TensorRunningAccum(window_length=self.trainer.accumulate_grad_batches)
# hook
self.trainer.logger_connector.on_epoch_start()
self.trainer.call_hook("on_epoch_start")
self.trainer.call_hook("on_train_epoch_start")
def on_train_batch_end(self, epoch_output, batch_end_outputs, batch, batch_idx, dataloader_idx):
batch_end_outputs = [opt_idx_out for opt_idx_out in batch_end_outputs if len(opt_idx_out)]
processed_batch_end_outputs = TrainLoop._prepare_outputs(batch_end_outputs, batch_mode=True)
# hook
self.trainer.call_hook('on_train_batch_end', processed_batch_end_outputs, batch, batch_idx, 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(epoch_output, batch_end_outputs)
def reset_train_val_dataloaders(self, model) -> None:
"""
Resets train and val dataloaders if none are attached to the trainer.
The val dataloader must be initialized before training loop starts, as the training loop
inspects the val dataloader to determine whether to run the evaluation loop.
"""
if self.trainer.train_dataloader is None:
self.trainer.reset_train_dataloader(model)
if self.trainer.val_dataloaders is None:
self.trainer.reset_val_dataloader(model)
def track_epoch_end_reduce_metrics(self, epoch_output, batch_end_outputs):
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
def get_active_optimizers(self, batch_idx: Optional[int] = None) -> List[Tuple[int, Optimizer]]:
"""
Returns the currently active optimizers. When multiple optimizers are used with different frequencies,
only one of the optimizers is active at a time.
Returns:
A list of tuples (opt_idx, optimizer) of currently active optimizers.
"""
if not self.trainer.optimizer_frequencies:
# call training_step once per optimizer
return list(enumerate(self.trainer.optimizers))
batch_idx = self.total_batch_idx if batch_idx is None else batch_idx
optimizers_loop_length = self.optimizer_freq_cumsum[-1]
current_place_in_loop = batch_idx % optimizers_loop_length
# find optimzier index by looking for the first {item > current_place} in the cumsum list
opt_idx = int(np.argmax(self.optimizer_freq_cumsum > current_place_in_loop))
return [(opt_idx, self.trainer.optimizers[opt_idx])]
def on_after_backward(self, batch_idx, untouched_loss):
# insert after step hook
self.trainer.call_hook("on_after_backward")
# when in dev debugging track the losses
self.trainer.dev_debugger.track_train_loss_history(batch_idx, untouched_loss.detach())
def _check_training_step_output(self, training_step_output):
if isinstance(training_step_output, torch.Tensor) and not self.trainer.lightning_module.automatic_optimization:
if training_step_output.grad_fn is None:
# TODO: Find why - RuntimeError: Expected to mark a variable ready only once ...
raise MisconfigurationException("In manual optimization, `training_step` should not return a Tensor")
elif self.trainer.lightning_module.automatic_optimization:
if not any((
isinstance(training_step_output, torch.Tensor),
(isinstance(training_step_output, Mapping)
and 'loss' in training_step_output), training_step_output is None
)):
raise MisconfigurationException(
"In automatic optimization, `training_step` must either return a Tensor, "
"a dict with key 'loss' or None (where the step will be skipped)."
)
def training_step(self, split_batch, batch_idx, opt_idx, hiddens):
# give the PL module a result for logging
model_ref = self.trainer.lightning_module
with self.trainer.profiler.profile("model_forward"):
step_kwargs = self._build_kwargs(split_batch, batch_idx, opt_idx, hiddens)
# manually capture logged metrics
model_ref._current_fx_name = 'training_step'
with self.trainer.profiler.profile("training_step"):
training_step_output = self.trainer.accelerator.training_step(step_kwargs)
self.trainer.accelerator.post_training_step()
training_step_output = self.trainer.call_hook("training_step_end", training_step_output)
self._check_training_step_output(training_step_output)
training_step_output = self._process_training_step_output(training_step_output)
if training_step_output is None:
return
closure_loss = None
loss = None
if self.trainer.lightning_module.automatic_optimization:
# accumulate loss. if accumulate_grad_batches==1, no effect
closure_loss = training_step_output.minimize / self.trainer.accumulate_grad_batches
# the loss will get scaled for amp. avoid any modifications to it
loss = closure_loss.detach().clone()
return AttributeDict(closure_loss=closure_loss, loss=loss, training_step_output=training_step_output)
def _process_training_step_output(self, training_step_output):
if training_step_output is None:
return None
results = self.results
loss = None
hiddens = None
results.extra = {}
# handle dict return
if isinstance(training_step_output, dict):
loss = training_step_output.pop("loss", None)
hiddens = training_step_output.pop("hiddens", None)
if hiddens is not None:
hiddens = hiddens.detach()
results.extra = training_step_output
# handle scalar return
elif isinstance(training_step_output, torch.Tensor):
loss = training_step_output
# map to results under the hood
results.minimize = loss
self._hiddens = hiddens
if self.trainer.move_metrics_to_cpu:
results.cpu()
return results
@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 optimizer_step(self, optimizer, opt_idx, batch_idx, train_step_and_backward_closure):
model_ref = self.trainer.lightning_module
is_lbfgs = isinstance(optimizer, torch.optim.LBFGS)
using_native_amp = self.trainer.amp_backend == AMPType.NATIVE
# native amp + lbfgs is a no go right now
if using_native_amp and is_lbfgs:
raise MisconfigurationException(
'native PyTorch amp and lbfgs are not compatible.'
' To request, please file a Github issue in PyTorch and tag @mcarilli'
)
# wraps into LightningOptimizer only for running step
optimizer = LightningOptimizer._to_lightning_optimizer(optimizer, self.trainer, opt_idx)
# model hook
model_ref.optimizer_step(
self.trainer.current_epoch,
batch_idx,
optimizer,
opt_idx,
train_step_and_backward_closure,
on_tpu=self.trainer._device_type == DeviceType.TPU and _TPU_AVAILABLE,
using_native_amp=using_native_amp,
using_lbfgs=is_lbfgs,
)
def on_before_zero_grad(self, optimizer):
self.trainer.call_hook('on_before_zero_grad', optimizer)
def optimizer_zero_grad(self, batch_idx, optimizer, opt_idx):
self.trainer.accelerator.optimizer_zero_grad(self.trainer.current_epoch, batch_idx, optimizer, opt_idx)
def track_and_norm_grad(self, optimizer) -> dict:
# track gradient norms
grad_norm_dict = {}
if (self.global_step + 1) % self.trainer.log_every_n_steps == 0 and float(self.trainer.track_grad_norm) > 0:
grad_norm_dict = grad_norm(self.trainer.lightning_module, self.trainer.track_grad_norm)
# clip gradients
self.trainer.accelerator.clip_gradients(
optimizer, self.trainer.gradient_clip_val, gradient_clip_algorithm=self.trainer.gradient_clip_algorithm
)
return grad_norm_dict
def _tbptt_split_batch(self, batch: Any) -> List[Any]:
splits = [batch]
truncated_bptt_enabled = self._truncated_bptt_enabled()
if truncated_bptt_enabled:
model_ref = self.trainer.lightning_module
with self.trainer.profiler.profile("tbptt_split_batch"):
splits = model_ref.tbptt_split_batch(batch, self._truncated_bptt_steps())
return splits
def run_training_epoch(self):
# modify dataloader if needed (ddp, etc...)
train_dataloader = self.trainer.accelerator.process_dataloader(self.trainer.train_dataloader)
# track epoch output
epoch_output = [[] for _ in range(self.num_active_optimizers)]
train_dataloader = self.trainer.data_connector.get_profiled_train_dataloader(train_dataloader)
dataloader_idx = 0
batch_idx = None
for batch_idx, (batch, is_last_batch) in train_dataloader:
self.batch_idx = batch_idx
# ------------------------------------
# TRAINING_STEP + TRAINING_STEP_END
# ------------------------------------
with self.trainer.profiler.profile("run_training_batch"):
batch_output = self.run_training_batch(batch, batch_idx, dataloader_idx)
# when returning -1 from train_step, we end epoch early
if batch_output.signal == -1:
break
# hook
self.on_train_batch_end(
epoch_output,
batch_output.training_step_output,
batch,
batch_idx,
dataloader_idx,
)
# -----------------------------------------
# SAVE METRICS TO LOGGERS AND PROGRESS_BAR
# -----------------------------------------
self.trainer.logger_connector.update_train_step_metrics()
# -----------------------------------------
# VALIDATE IF NEEDED
# -----------------------------------------
should_check_val = self._should_check_val_fx(batch_idx, is_last_batch)
if should_check_val:
self.trainer.validating = True
self.trainer._run_evaluation()
self.trainer.training = True
# -----------------------------------------
# SAVE LOGGERS (ie: Tensorboard, etc...)
# -----------------------------------------
self.save_loggers_on_train_batch_end()
# update LR schedulers
self.update_lr_schedulers('step')
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()
max_steps_reached = (self.max_steps is not None and self.max_steps <= self.global_step)
if max_steps_reached or self.trainer.should_stop or self._num_training_batches_reached(is_last_batch):
break
if batch_idx is None:
# dataloader/iterator did not produce a batch
return
# handle epoch_output on epoch end
self.on_train_epoch_end(epoch_output)
# the global step is manually decreased here due to backwards compatibility with existing loggers
# as they expect that the same step is used when logging epoch end metrics even when the batch loop has
# finished. this means the attribute does not exactly track the number of optimizer steps applied.
# TODO(@carmocca): deprecate and rename so users don't get confused
self.global_step -= 1
# log epoch metrics
self.trainer.logger_connector.update_train_epoch_metrics()
self.global_step += 1
self.update_lr_schedulers('epoch')
did_train_only = self.trainer.disable_validation or self.trainer.evaluation_loop.should_skip_evaluation(
self.trainer.num_val_batches
)
if did_train_only:
self.global_step -= 1
self.check_checkpoint_callback(True)
self.global_step += 1
def on_train_epoch_end(self, epoch_output: List[List[List['ResultCollection']]]) -> None:
# inform logger the batch loop has finished
self.trainer.logger_connector.epoch_end_reached()
# prepare epoch output
processed_epoch_output = TrainLoop._prepare_outputs(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'
training_epoch_end_output = model.training_epoch_end(processed_epoch_output)
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_epoch_output)
self.trainer.call_hook('on_epoch_end')
self.trainer.logger_connector.on_epoch_end()
def _on_train_epoch_end_hook(self, processed_epoch_output) -> None:
# 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.warn(
"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", DeprecationWarning
)
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 run_training_batch(self, batch, batch_idx, dataloader_idx):
# bookkeeping
self._hiddens = None
optimizers = list(enumerate(self.trainer.optimizers))
# track all outputs across time and num of optimizers
batch_outputs = [[] for _ in range(len(optimizers))]
if batch is None:
self.warning_cache.warn("train_dataloader yielded None. If this was on purpose, ignore this warning...")
return AttributeDict(signal=0, training_step_output=batch_outputs)
# hook
self.trainer.logger_connector.on_batch_start()
response = self.trainer.call_hook("on_batch_start")
if response == -1:
return AttributeDict(signal=-1)
# hook
response = self.trainer.call_hook("on_train_batch_start", batch, batch_idx, dataloader_idx)
if response == -1:
return AttributeDict(signal=-1)
# lightning module hook
splits = self._tbptt_split_batch(batch)
for split_idx, split_batch in enumerate(splits):
self.split_idx = split_idx
# let logger connector extract batch size
self.trainer.logger_connector.on_train_split_start(batch_idx, split_idx, split_batch)
if self.trainer.lightning_module.automatic_optimization:
for opt_idx, optimizer in self.get_active_optimizers(batch_idx):
result = self._run_optimization(batch_idx, split_batch, opt_idx, optimizer)
if result:
batch_outputs[opt_idx].append(result.training_step_output)
else:
# in manual optimization, there is no looping over optimizers
result = self._run_optimization(batch_idx, split_batch)
if result:
batch_outputs[0].append(result.training_step_output)
return AttributeDict(signal=0, training_step_output=batch_outputs)
def _run_optimization(self, batch_idx, split_batch, opt_idx=0, optimizer=None):
# TODO: In v1.5, when optimizer_idx gets removed from training_step in manual_optimization, change
# opt_idx=0 to opt_idx=None in the signature here
# toggle model params
self.run_optimization_start(opt_idx, optimizer)
result = AttributeDict()
closure = self.make_closure(split_batch, batch_idx, opt_idx, optimizer, self._hiddens, result)
if self.should_accumulate():
# For gradient accumulation
# -------------------
# calculate loss (train step + train step end)
# -------------------
# automatic_optimization=True: perform ddp sync only when performing optimizer_step
# automatic_optimization=False: don't block synchronization here
with self.block_ddp_sync_behaviour():
closure()
# ------------------------------
# BACKWARD PASS
# ------------------------------
# gradient update with accumulated gradients
else:
if self.trainer.lightning_module.automatic_optimization:
self.optimizer_step(optimizer, opt_idx, batch_idx, closure)
if len(self.trainer.optimizers) > 1:
# revert back to previous state
self.trainer.lightning_module.untoggle_optimizer(opt_idx)
else:
result = self.training_step(split_batch, batch_idx, opt_idx, self._hiddens)
if not result:
# user decided to skip optimization
return result
# update running loss + reset accumulated loss
self.update_running_loss(result.loss)
self._process_closure_result(result)
return result
def training_step_and_backward_closure(
self,
split_batch: Any,
batch_idx: int,
opt_idx: int,
optimizer: Optimizer,
hiddens,
return_result: AttributeDict,
) -> Optional[torch.Tensor]:
result = self.training_step_and_backward(split_batch, batch_idx, opt_idx, optimizer, hiddens)
if result is not None:
return_result.update(result)
return return_result.loss
def make_closure(self, *closure_args, **closure_kwargs: Any) -> Callable:
""" Wraps the training step closure into a partial object which will be called within ``optimizer.step``. """
partial_func = partial(self.training_step_and_backward_closure, *closure_args, **closure_kwargs)
return update_wrapper(partial_func, self.training_step_and_backward_closure)
@contextmanager
def block_ddp_sync_behaviour(self, should_block_sync: bool = False):
"""
automatic_optimization = True
Blocks ddp sync gradients behaviour on backwards pass.
This is useful for skipping sync when accumulating gradients, reducing communication overhead
automatic_optimization = False
do not block ddp gradient sync when using manual optimization
as gradients are needed within the training step
Returns:
context manager with sync behaviour off
"""
if (
isinstance(self.trainer.training_type_plugin, ParallelPlugin)
and (self.trainer.lightning_module.automatic_optimization or should_block_sync)
):
with self.trainer.training_type_plugin.block_backward_sync():
yield None
else:
yield None
def _process_closure_result(self, opt_closure_result: Optional[AttributeDict]) -> None:
if not opt_closure_result:
return
# check if loss or model weights are nan
if self.trainer.terminate_on_nan:
self._check_finite(opt_closure_result.loss)
def training_step_and_backward(self, split_batch, batch_idx, opt_idx, optimizer, hiddens):
"""Wrap forward, zero_grad and backward in a closure so second order methods work"""
with self.trainer.profiler.profile("training_step_and_backward"):
# lightning module hook
result = self.training_step(split_batch, batch_idx, opt_idx, hiddens)
if not self._skip_backward and self.trainer.lightning_module.automatic_optimization:
is_first_batch_to_accumulate = batch_idx % self.trainer.accumulate_grad_batches == 0
if is_first_batch_to_accumulate:
self.on_before_zero_grad(optimizer)
self.optimizer_zero_grad(batch_idx, optimizer, opt_idx)
# backward pass
if result is not None:
with self.trainer.profiler.profile("backward"):
self.backward(result, optimizer, opt_idx)
# hook - call this hook only
# when gradients have finished to accumulate
if not self.should_accumulate():
self.on_after_backward(batch_idx, result.loss)
# check if loss or model weights are nan
if self.trainer.terminate_on_nan:
self._check_finite(result.loss)
else:
self.warning_cache.warn(
"training_step returned None. If this was on purpose, ignore this warning..."
)
return result
def _check_finite(self, loss: torch.Tensor) -> None:
if not torch.isfinite(loss).all():
raise ValueError(f'The loss returned in `training_step` is {loss}.')
model = self.trainer.lightning_module
detect_nan_parameters(model)
def backward(self, result, optimizer, opt_idx, *args, **kwargs):
self.trainer.dev_debugger.track_event("backward_call")
should_accumulate = self.should_accumulate()
# backward can be called manually in the training loop
if isinstance(result, torch.Tensor):
self.trainer.accelerator.backward(result, optimizer, opt_idx, should_accumulate, *args, **kwargs)
else:
result.closure_loss = self.trainer.accelerator.backward(
result.closure_loss, optimizer, opt_idx, should_accumulate, *args, **kwargs
)
if not self.should_accumulate():
# track gradients
grad_norm_dict = self.track_and_norm_grad(optimizer=optimizer)
if grad_norm_dict:
self.trainer.lightning_module._current_fx_name = "on_after_backward"
self.trainer.lightning_module.log_grad_norm(grad_norm_dict)
def update_lr_schedulers(self, interval: str) -> None:
if interval == "step":
finished_accumulation = self._accumulated_batches_reached()
finished_epoch = self._num_training_batches_reached()
if not finished_accumulation and not finished_epoch:
return
self.trainer.optimizer_connector.update_learning_rates(
interval=interval,
opt_indices=[opt_idx for opt_idx, _ in self.get_active_optimizers()],
)
def increment_accumulated_grad_global_step(self):
num_accumulated_batches_reached = self._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.global_step)
def _accumulated_batches_reached(self):
return (self.batch_idx + 1) % self.trainer.accumulate_grad_batches == 0
def _num_training_batches_reached(self, is_last_batch=False):
return (self.batch_idx + 1) == self.trainer.num_training_batches or is_last_batch
def should_accumulate(self):
# checks if backward or backward + optimizer step (via closure)
accumulation_done = self._accumulated_batches_reached()
is_final_batch = self._num_training_batches_reached()
return not (accumulation_done or is_final_batch)
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: 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 _build_kwargs(self, batch, batch_idx, opt_idx, hiddens):
# enable not needing to add opt_idx to training_step
step_kwargs = OrderedDict([('batch', batch), ('batch_idx', batch_idx)])
lightning_module = self.trainer.lightning_module
if len(self.trainer.optimizers) > 1:
training_step_fx = getattr(lightning_module, "training_step")
has_opt_idx_in_train_step = is_param_in_hook_signature(training_step_fx, "optimizer_idx")
if has_opt_idx_in_train_step:
if not lightning_module.automatic_optimization:
self.warning_cache.warn(
"`training_step` hook signature has changed in v1.3."
" `optimizer_idx` argument has been removed in case of manual optimization. Support for"
" the old signature will be removed in v1.5", DeprecationWarning
)
step_kwargs['optimizer_idx'] = opt_idx
elif not has_opt_idx_in_train_step and self.trainer.lightning_module.automatic_optimization:
raise ValueError(
f"Your LightningModule defines {len(self.trainer.optimizers)} optimizers but"
' `training_step` is missing the `optimizer_idx` argument.'
)
# pass hiddens if using tbptt
if self._truncated_bptt_enabled():
step_kwargs['hiddens'] = hiddens
return step_kwargs
def _truncated_bptt_enabled(self) -> bool:
""" Temporary tbptt utilities until this flag is fully migrated to the lightning module. """
return self._truncated_bptt_steps() > 0
def _truncated_bptt_steps(self) -> int:
lightning_module = self.trainer.lightning_module
# Give precedence to the LightningModule as the Trainer flag will be removed in v1.5
if lightning_module.truncated_bptt_steps > 0:
return lightning_module.truncated_bptt_steps
return self.trainer.truncated_bptt_steps or 0
def save_loggers_on_train_batch_end(self):
# 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 run_optimization_start(self, opt_idx, optimizer):
# 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 self.trainer.lightning_module.automatic_optimization and len(self.trainer.optimizers) > 1:
model = self.trainer.lightning_module
model.toggle_optimizer(optimizer, opt_idx)
def update_running_loss(self, current_loss: torch.Tensor) -> None:
if self.trainer.lightning_module.automatic_optimization:
# track total loss for logging (avoid mem leaks)
self.accumulated_loss.append(current_loss)
accumulated_loss = self.accumulated_loss.mean()
if accumulated_loss is not None:
# calculate running loss for display
self.running_loss.append(self.accumulated_loss.mean() * self.trainer.accumulate_grad_batches)
# reset for next set of accumulated grads
self.accumulated_loss.reset()
class EvaluationLoop(DataLoaderLoop):
"""Loops over all dataloaders for evaluation."""
def __init__(self):
super().__init__()
self.outputs: List[EPOCH_OUTPUT] = []
self.epoch_loop = EvaluationEpochLoop()
self._results = ResultCollection(training=False)
self._max_batches: Optional[Union[int, Sequence[int]]] = None
self._has_run: bool = False
@property
def num_dataloaders(self) -> int:
"""Returns the total number of dataloaders."""
# case where user does:
# return dl1, dl2
dataloaders = self.dataloaders
if dataloaders is None:
return 0
length = len(dataloaders)
if length > 0 and isinstance(dataloaders[0], (list, tuple)):
length = len(dataloaders[0])
return length
@property
def dataloaders(self) -> Sequence[DataLoader]:
"""Returns the validation or test dataloaders."""
if self.trainer.testing:
return self.trainer.test_dataloaders
return self.trainer.val_dataloaders
def connect(self, epoch_loop: EvaluationEpochLoop):
"""Connect the evaluation epoch loop with this loop."""
self.epoch_loop = epoch_loop
@property
def done(self) -> bool:
"""Returns whether all dataloaders are processed or evaluation should be skipped altogether."""
return super().done or self.skip
@property
def skip(self) -> bool:
"""Returns whether the evaluation should be skipped."""
max_batches = self.get_max_batches()
return sum(max_batches) == 0
def reset(self) -> None:
"""Resets the internal state of the loop."""
self._max_batches = self.get_max_batches()
# bookkeeping
self.outputs = []
if isinstance(self._max_batches, int):
self._max_batches = [self._max_batches] * len(self.dataloaders)
super().reset()
def on_skip(self) -> List:
return []
def on_run_start(self, *args: Any, **kwargs: Any) -> None:
"""Runs the ``on_evaluation_model_eval``, ``on_evaluation_start`` and ``on_evaluation_epoch_start``
hooks."""
void(*args, **kwargs)
# hook
self.on_evaluation_model_eval()
self.trainer.lightning_module.zero_grad()
self.on_evaluation_start()
self.on_evaluation_epoch_start()
def advance(self, *args: Any, **kwargs: Any) -> None:
"""Performs evaluation on one single dataloader."""
void(*args, **kwargs)
dataloader_idx: int = self.current_dataloader_idx
dataloader = self.trainer.accelerator.process_dataloader(
self.current_dataloader)
dataloader = self.trainer.data_connector.get_profiled_dataloader(
dataloader, dataloader_idx=dataloader_idx)
dl_max_batches = self._max_batches[dataloader_idx]
dl_outputs = self.epoch_loop.run(dataloader, dataloader_idx,
dl_max_batches, self.num_dataloaders)
# store batch level output per dataloader
self.outputs.append(dl_outputs)
if not self.trainer.sanity_checking:
# indicate the loop has run
self._has_run = True
def on_run_end(self) -> Any:
"""Runs the ``on_evaluation_epoch_end`` hook."""
outputs = self.outputs
# free memory
self.outputs = []
# with a single dataloader don't pass a 2D list
if len(outputs) > 0 and self.num_dataloaders == 1:
outputs = outputs[0]
# lightning module method
self.evaluation_epoch_end(outputs)
# hook
self.on_evaluation_epoch_end()
# log epoch metrics
eval_loop_results = self.trainer.logger_connector.update_eval_epoch_metrics(
)
# hook
self.on_evaluation_end()
# enable train mode again
self.on_evaluation_model_train()
return eval_loop_results
def get_max_batches(self) -> List[Union[int, float]]:
"""Returns the max number of batches for each dataloader."""
if self.trainer.testing:
max_batches = self.trainer.num_test_batches
else:
if self.trainer.sanity_checking:
self.trainer.num_sanity_val_batches = [
min(self.trainer.num_sanity_val_steps, val_batches)
for val_batches in self.trainer.num_val_batches
]
max_batches = self.trainer.num_sanity_val_batches
else:
max_batches = self.trainer.num_val_batches
return max_batches
def reload_evaluation_dataloaders(self) -> None:
"""Reloads dataloaders if necessary."""
if self.trainer.testing:
self.trainer.reset_test_dataloader()
elif self.trainer.val_dataloaders is None or self.trainer._should_reload_dl_epoch:
self.trainer.reset_val_dataloader()
def on_evaluation_start(self, *args: Any, **kwargs: Any) -> None:
"""Runs ``on_{validation/test}_start`` hooks."""
assert self._results is not None
self._results.to(device=self.trainer.lightning_module.device)
if self.trainer.testing:
self.trainer.call_hook("on_test_start", *args, **kwargs)
else:
self.trainer.call_hook("on_validation_start", *args, **kwargs)
def on_evaluation_model_eval(self) -> None:
"""Sets model to eval mode."""
if self.trainer.testing:
self.trainer.call_hook("on_test_model_eval")
else:
self.trainer.call_hook("on_validation_model_eval")
def on_evaluation_model_train(self) -> None:
"""Sets model to train mode."""
model_ref = self.trainer.lightning_module
if self.trainer.testing:
model_ref.on_test_model_train()
else:
model_ref.on_validation_model_train()
def on_evaluation_end(self, *args: Any, **kwargs: Any) -> None:
"""Runs ``on_{validation/test}_end`` hook."""
if self.trainer.testing:
self.trainer.call_hook("on_test_end", *args, **kwargs)
else:
self.trainer.call_hook("on_validation_end", *args, **kwargs)
# reset any `torchmetrics.Metric` and the logger connector state
self.trainer.logger_connector.reset(metrics=True)
def on_evaluation_epoch_start(self, *args: Any, **kwargs: Any) -> None:
"""Runs ``on_epoch_start`` and ``on_{validation/test}_epoch_start`` hooks."""
self.trainer.logger_connector.on_epoch_start()
self.trainer.call_hook("on_epoch_start", *args, **kwargs)
if self.trainer.testing:
self.trainer.call_hook("on_test_epoch_start", *args, **kwargs)
else:
self.trainer.call_hook("on_validation_epoch_start", *args,
**kwargs)
def evaluation_epoch_end(self, outputs: EPOCH_OUTPUT) -> None:
"""Runs ``{validation/test}_epoch_end``"""
# inform logger the batch loop has finished
self.trainer.logger_connector.epoch_end_reached()
# call the model epoch end
model = self.trainer.lightning_module
# unset dataloader_idx in model
model._current_dataloader_idx = None
if self.trainer.testing:
if is_overridden("test_epoch_end", model):
model._current_fx_name = "test_epoch_end"
model.test_epoch_end(outputs)
else:
if is_overridden("validation_epoch_end", model):
model._current_fx_name = "validation_epoch_end"
model.validation_epoch_end(outputs)
def on_evaluation_epoch_end(self) -> None:
"""Runs ``on_{validation/test}_epoch_end`` hook."""
hook_name = "on_test_epoch_end" if self.trainer.testing else "on_validation_epoch_end"
self.trainer.call_hook(hook_name)
self.trainer.call_hook("on_epoch_end")
self.trainer.logger_connector.on_epoch_end()
def teardown(self) -> None:
self._results.cpu()
self.epoch_loop.teardown()
class FitLoop(Loop):
"""This Loop iterates over the epochs to run the training
Args:
min_epochs: The minimum number of epochs
max_epochs: The maximum number of epochs
min_steps: The minimum number of steps
max_steps: The maximum number of epoch
.. note::
If neither the minimum epochs nor steps are specified the minimum number of epochs is set to 1
and if neither the maximum steps nor epochs are specified, the maximum epochs are set to 1000.
"""
def __init__(self,
min_epochs: Optional[int] = None,
max_epochs: Optional[int] = None,
min_steps: Optional[int] = None,
max_steps: Optional[int] = None):
super().__init__()
self.max_epochs = 1000 if (max_epochs is None
and max_steps is None) else max_epochs
self.min_epochs = 1 if (min_epochs is None
and min_steps is None) else min_epochs
self.training_loop = TrainingEpochLoop(min_steps, max_steps)
self.results = ResultCollection(training=True)
@property
def current_epoch(self) -> int:
"""Return the current epoch"""
return self.iteration_count
@current_epoch.setter
def current_epoch(self, value: int) -> None:
"""Setter for the current epoch"""
self.iteration_count = value
@property
def global_step(self) -> int:
"""Returns the global step"""
return self.training_loop.global_step
@global_step.setter
def global_step(self, value: int) -> None:
"""Sets the global step (forwards to training_loop)"""
self.training_loop.global_step = value
@property
def total_batch_idx(self) -> int:
"""Returns the total number of batches already run (across all epochs)"""
return self.training_loop.total_batch_idx
@property
def batch_idx(self) -> int:
"""Returns the number of batches already run within this epoch"""
return self.training_loop.iteration_count
@property
def split_idx(self) -> int:
"""Returns the index of the current batch split (within the current batch) for bptt"""
return self.training_loop.split_idx
@property
def min_steps(self) -> int:
# TODO(@justusschock): Why aren't we using the attribute in this class?
"""Returns the minimum numnber of steps to run"""
return self.training_loop.min_steps
@property
def max_steps(self) -> int:
"""Returns the maximum number of steps to run"""
return self.training_loop.max_steps
@max_steps.setter
def max_steps(self, value: int) -> None:
"""Sets the maximum number of steps (forwards to training_loop)"""
# TODO(@awaelchli): This setter is required by debugging connector (fast dev run), should be avoided
self.training_loop.max_steps = value
@property
def running_loss(self) -> TensorRunningAccum:
"""Returns the running loss"""
return self.training_loop.batch_loop.running_loss
@property
def _skip_backward(self) -> bool:
""" Determines whether the loop will skip backward during automatic optimization. """
return self.training_loop.batch_loop._skip_backward
@_skip_backward.setter
def _skip_backward(self, value: bool) -> None:
""" Determines whether the loop will skip backward during automatic optimization. """
self.training_loop.batch_loop._skip_backward = value
@property
def done(self) -> bool:
"""Evaluates when to leave the loop.
Returns True if trainer.should_stop was set (e.g. by early stopping)
or if the maximum number of steps or epochs is reached.
"""
# TODO(@awaelchli): Move track steps inside training loop and move part of these condition inside training loop
stop_steps = self.max_steps is not None and self.global_step >= self.max_steps
stop_epochs = self.max_epochs is not None and self.current_epoch >= self.max_epochs
should_stop = False
if self.trainer.should_stop:
# early stopping
met_min_epochs = self.current_epoch >= self.min_epochs if self.min_epochs else True
met_min_steps = self.global_step >= self.min_steps if self.min_steps else True
if met_min_epochs and met_min_steps:
should_stop = True
else:
log.info(
'Trainer was signaled to stop but required minimum epochs'
f' ({self.min_epochs}) or minimum steps ({self.min_steps}) has'
' not been met. Training will continue...')
self.trainer.should_stop = should_stop
return stop_steps or should_stop or stop_epochs
@property
def skip(self) -> bool:
"""Whether we should skip the training and immediately return from the call to :meth:`run`."""
return self.done or self.trainer.num_training_batches == 0
def connect(self, trainer: 'pl.Trainer', *args: Any,
**kwargs: Any) -> None:
"""Connects the loop with necessary arguments like the trainer"""
# TODO(@justusschock): Do we want to forward *args and **kwargs to the inner loop here?
# TODO(@justusschock): Can we make the trainer a weakref/proxy?
void(*args, **kwargs)
self.trainer = trainer
self.training_loop.connect(trainer)
def reset(self) -> None:
"""Resets the internal state of this loop"""
def on_run_start(self) -> None:
"""Calls the ``on_train_start`` hook."""
self.results.to(device=self.trainer.lightning_module.device)
self.trainer.call_hook("on_train_start")
def on_advance_start(self) -> None:
"""Prepares the dataloader for training and calls the hooks ``on_epoch_start`` and ``on_train_epoch_start``"""
model = self.trainer.lightning_module
# reset train dataloader
if self.current_epoch != 0 and self.trainer.reload_dataloaders_every_epoch:
self.trainer.reset_train_dataloader(model)
# TODO: specify the possible exception
with suppress(Exception):
# set seed for distributed sampler (enables shuffling for each epoch)
self.trainer.train_dataloader.sampler.set_epoch(self.current_epoch)
# changing gradient according accumulation_scheduler
self.trainer.accumulation_scheduler.on_train_epoch_start(
self.trainer, self.trainer.lightning_module)
# stores accumulated grad fractions per batch
self.training_loop.batch_loop.accumulated_loss = TensorRunningAccum(
window_length=self.trainer.accumulate_grad_batches)
def advance(self) -> None:
"""Runs one whole epoch."""
train_dataloader = self.trainer.accelerator.process_dataloader(
self.trainer.train_dataloader)
train_dataloader = self.trainer.data_connector.get_profiled_train_dataloader(
train_dataloader)
with self.trainer.profiler.profile("run_training_epoch"):
# run train epoch
epoch_output = self.training_loop.run(train_dataloader)
if epoch_output is None:
return
# the global step is manually decreased here due to backwards compatibility with existing loggers
# as they expect that the same step is used when logging epoch end metrics even when the batch loop has
# finished. this means the attribute does not exactly track the number of optimizer steps applied.
# TODO(@carmocca): deprecate and rename so users don't get confused
self.global_step -= 1
# log epoch metrics
self.trainer.logger_connector.update_train_epoch_metrics()
self.global_step += 1
def on_advance_end(self) -> None:
"""Updates the LR schedulers and does some internal bookkeeping"""
if self.training_loop.batches_seen == 0:
return
self.training_loop.update_lr_schedulers('epoch',
update_plateau_schedulers=True)
did_train_only = self.trainer.disable_validation or self.trainer.evaluation_loop.skip
if did_train_only:
self.global_step -= 1
self.check_checkpoint_callback(True)
self.global_step += 1
def on_run_end(self) -> None:
"""Runs teardown logic and calls the ``on_train_end`` hook"""
# NOTE: the iteration_count/current_epoch is already incremented
# Lightning today does not increment the current epoch at the last epoch run in Trainer.fit
# To simulate that current behavior, we decrement here.
# TODO: must be fixed by https://github.com/PyTorchLightning/pytorch-lightning/issues/5007
self.current_epoch -= 1
# trigger checkpoint check. need to temporarily decrease the global step to avoid saving duplicates
# when a checkpoint was saved at the last step
self.training_loop.global_step -= 1
# TODO: see discussion/rework https://github.com/PyTorchLightning/pytorch-lightning/issues/7406
self.check_checkpoint_callback(should_update=True, is_last=True)
self.training_loop.global_step += 1
# hook
self.trainer.call_hook("on_train_end")
# todo: TPU 8 cores hangs in flush with TensorBoard. Might do for all loggers.
# It might be related to xla tensors blocked when moving the cpu
# kill loggers
if self.trainer.logger is not None:
self.trainer.logger.finalize("success")
# summarize profile results
self.trainer.profiler.describe()
# give accelerators a chance to finish
self.trainer.accelerator.on_train_end()
# reset bookkeeping
self.trainer._running_stage = None
def should_accumulate(self) -> bool:
"""Whether the gradients should be accumulated"""
return self.training_loop.batch_loop.should_accumulate()
def get_active_optimizers(
self,
batch_idx: Optional[int] = None) -> List[Tuple[int, Optimizer]]:
"""Generates a list of active optimizers"""
return self.training_loop.batch_loop.get_active_optimizers(batch_idx)
def check_checkpoint_callback(self,
should_update: bool,
is_last: bool = False):
"""Checks if checkpointing needs to be done"""
# TODO: bake this logic into the ModelCheckpoint callback
if should_update and self.trainer.checkpoint_connector.has_trained:
callbacks = self.trainer.checkpoint_callbacks
if is_last and any(cb.save_last and cb.verbose
for cb in callbacks):
rank_zero_info("Saving latest checkpoint...")
model = self.trainer.lightning_module
for cb in callbacks:
cb.on_validation_end(self.trainer, model)
