def training_loop(self,
data_loader: DataLoader,
start_iter: int,
validation_data_loader: Optional[DataLoader] = None):
self.logger.info(f'Local rank: {communication.get_local_rank()}.')
self.model.train()
loss_fns = self.build_loss()
metric_fns = self.build_metrics()
storage = get_event_storage()
total_iter = self.cfg.training.num_iterations # noqa
for data, iter_idx in zip(data_loader, range(start_iter, total_iter)):
data = AddNames()(data)
if iter_idx == 0:
self.logger.info(
f"First case: slice_no: {data['slice_no'][0]}, filename: {data['filename'][0]}."
)
storage.add_image('train/mask', data['sampling_mask'][0, ...,
0])
storage.add_image(
'train/target',
normalize_image(
data['target'][0].rename(None).unsqueeze(0)))
storage.add_image(
'train/masked_image',
normalize_image(
transforms.modulus_if_complex(
data['masked_image'][0]).rename(None).unsqueeze(
0)))
self.write_to_logs()
try:
output, loss_dict = self._do_iteration(data, loss_fns)
except ProcessKilledException as e:
# If the process is killed, the output if saved at state iter_idx, which is the current state,
# so the computation can restart from the last iteration.
self.logger.exception(f'{e}.')
self.checkpointer.save(
iter_idx) # Save checkpoint at kill. # noqa
self.write_to_logs(
) # TODO: This causes the issue that current metrics are not written,
# and you end up with an empty line.
sys.exit(f'Exited with exception: {e}')
# Gradient accumulation
if (iter_idx +
1) % self.cfg.training.gradient_steps == 0: # type: ignore
# TODO: Is this slow? This is a generator, so should be cheap.
pass
# parameter_list = self.model.parameters() if not self.mixed_precision else \
# amp.master_params(self.__optimizer)
# if self.cfg.training.gradient_steps > 1: # type: ignore
# for parameter in parameter_list:
# if parameter.grad is not None:
# # In-place division
# parameter.grad.div_(self.cfg.training.gradient_steps) # type: ignore
# if self.cfg.training.gradient_clipping > 0.0: # type: ignore
# torch.nn.utils.clip_grad_norm_(parameter_list, self.cfg.training.gradient_clipping) # type: ignore
#
# # Gradient norm
# if self.cfg.training.gradient_debug: # type: ignore
# parameters = list(filter(lambda p: p.grad is not None, parameter_list))
# gradient_norm = sum([parameter.grad.data ** 2 for parameter in parameters]).sqrt() # typing: ignore
# storage.add_scalar('gradient_norm', gradient_norm)
self.__optimizer.step() # type: ignore
# Incorrect inference by mypy and pyflake
self.__lr_scheduler.step() # type: ignore # noqa
storage.add_scalar('lr',
self.__optimizer.param_groups[0]['lr'],
smoothing_hint=False)
self.__optimizer.zero_grad() # type: ignore
# Reduce the loss over all devices
loss_dict_reduced = communication.reduce_tensor_dict(loss_dict)
loss_reduced = sum(loss_dict_reduced.values())
metrics_dict = evaluate_dict(
metric_fns,
transforms.modulus_if_complex(output.detach()).rename(None),
data['target'].rename(None).detach().to(self.device),
reduction='mean')
metrics_dict_reduced = communication.reduce_tensor_dict(
metrics_dict) if metrics_dict else {}
storage.add_scalars(loss=loss_reduced,
**loss_dict_reduced,
**metrics_dict_reduced)
if validation_data_loader is not None \
and iter_idx > 5\
and (iter_idx % self.cfg.training.validation_steps == 0 or (iter_idx + 1) == total_iter):
val_loss_dict = self.evaluate(
validation_data_loader,
loss_fns,
evaluation_round=iter_idx //
self.cfg.training.validation_steps - 1)
self.logger.info(f'Done evaluation at iteration {iter_idx}.')
storage.add_scalars(**prefix_dict_keys(val_loss_dict, 'val_'),
smoothing_hint=False)
self.model.train()
if iter_idx > 5 and\
(iter_idx % self.cfg.training.checkpointer.checkpoint_steps == 0 or (iter_idx + 1) == total_iter):
self.logger.info(f'Checkpointing at iteration: {iter_idx}.')
self.checkpointer.save(iter_idx)
# Log every 20 iterations, or at a validation step or at the end of training.
if iter_idx > 5 and (iter_idx % 20 == 0 or iter_idx %
self.cfg.training.validation_steps == 0 or
(iter_idx + 1) == total_iter):
self.write_to_logs()
storage.step()
def training_loop(
self,
data_loader: DataLoader,
start_iter: int,
validation_data_loaders: Optional[List[DataLoader]] = None,
experiment_directory: Optional[pathlib.Path] = None,
):
self.logger.info(f"Local rank: {communication.get_local_rank()}.")
self.models_training_mode()
loss_fns = self.build_loss()
metric_fns = self.build_metrics(self.cfg.training.metrics)
storage = get_event_storage()
total_iter = self.cfg.training.num_iterations # noqa
for data, iter_idx in zip(data_loader, range(start_iter, total_iter)):
data = AddNames()(data)
if iter_idx == start_iter:
self.ndim = self.compute_dimensionality_from_sample(data)
self.logger.info(f"Data dimensionality: {self.ndim}.")
if iter_idx == 0:
self.log_first_training_example(data)
try:
output, loss_dict = self._do_iteration(data, loss_fns)
except ProcessKilledException as e:
# If the process is killed, the output if saved at state iter_idx, which is the current state,
# so the computation can restart from the last iteration.
self.logger.exception(f"Exiting with exception: {e}.")
self.checkpointer.save(
iter_idx) # Save checkpoint at kill. # noqa
self.write_to_logs(
) # TODO: This causes the issue that current metrics are not written,
# and you end up with an empty line.
sys.exit(-1)
# Gradient accumulation
if (iter_idx +
1) % self.cfg.training.gradient_steps == 0: # type: ignore
if self.cfg.training.gradient_steps > 1: # type: ignore
for parameter in self.model.parameters():
if parameter.grad is not None:
# In-place division
parameter.grad.div_(self.cfg.training.
gradient_steps) # type: ignore
if self.cfg.training.gradient_clipping > 0.0: # type: ignore
self._scaler.unscale_(self.__optimizer)
torch.nn.utils.clip_grad_norm_(
self.model.parameters(),
self.cfg.training.gradient_clipping)
# Gradient norm
if self.cfg.training.gradient_debug: # type: ignore
warnings.warn(
f"Gradient debug set. This will affect training performance. Only use for debugging."
f"This message will only be displayed once.")
parameters = list(
filter(lambda p: p.grad is not None,
self.model.parameters()))
gradient_norm = sum([
parameter.grad.data**2 for parameter in parameters
]).sqrt() # typing: ignore
storage.add_scalar("train/gradient_norm", gradient_norm)
# Same as self.__optimizer.step() for mixed precision.
self._scaler.step(self.__optimizer)
# Updates the scale for next iteration.
self._scaler.update()
# Incorrect inference by mypy and pyflake
self.__lr_scheduler.step() # type: ignore # noqa
storage.add_scalar("lr",
self.__optimizer.param_groups[0]["lr"],
smoothing_hint=False)
self.__optimizer.zero_grad() # type: ignore
# Reduce the loss over all devices
loss_dict_reduced = communication.reduce_tensor_dict(loss_dict)
loss_reduced = sum(loss_dict_reduced.values())
metrics_dict = evaluate_dict(
metric_fns,
transforms.modulus_if_complex(output.detach()).rename(None),
data["target"].rename(None).detach().to(self.device),
reduction="mean",
)
metrics_dict_reduced = (
communication.reduce_tensor_dict(metrics_dict)
if metrics_dict else {})
storage.add_scalars(loss=loss_reduced,
**loss_dict_reduced,
**metrics_dict_reduced)
if iter_idx > 5 and (
iter_idx % self.cfg.training.checkpointer.checkpoint_steps
== 0 or (iter_idx + 1) == total_iter):
self.logger.info(f"Checkpointing at iteration {iter_idx}.")
self.checkpointer.save(iter_idx)
if (validation_data_loaders is not None and iter_idx > 5
and (iter_idx % self.cfg.training.validation_steps == 0 or
(iter_idx + 1) == total_iter)):
for (
curr_dataset_name,
curr_validation_data_loader,
) in validation_data_loaders:
self.logger.info(
f"Evaluating: {curr_dataset_name}..."
) # TODO(jt): Fix with better names and stuff.
(
curr_val_loss_dict,
curr_val_metric_dict_per_case,
visualize_slices,
visualize_target,
) = self.evaluate(
curr_validation_data_loader,
loss_fns,
is_validation_process=True,
)
if experiment_directory:
# Make dictionary serializable for logging
serializable_val_metric_dict = {
k0: {k1: float(v1)
for k1, v1 in v0.items()}
for k0, v0 in
curr_val_metric_dict_per_case.items()
}
write_json(
experiment_directory /
f"metrics_val_{curr_dataset_name}_{iter_idx}.json",
serializable_val_metric_dict,
)
# Metric dict still needs to be reduced as it gives values *per* data
curr_val_metric_dict = reduce_list_of_dicts(list(
curr_val_metric_dict_per_case.values()),
mode="average")
key_prefix = ("val/" if not curr_dataset_name else
f"val/{curr_dataset_name}/")
val_loss_reduced = sum(curr_val_loss_dict.values())
storage.add_scalars(
**{key_prefix + "loss": val_loss_reduced},
**{
**prefix_dict_keys(curr_val_metric_dict, key_prefix),
**prefix_dict_keys(curr_val_loss_dict, key_prefix),
},
smoothing_hint=False,
)
visualize_slices = self.process_slices_for_visualization(
visualize_slices, visualize_target)
storage.add_image(f"{key_prefix}prediction",
visualize_slices)
if iter_idx // self.cfg.training.validation_steps - 1 == 0:
visualize_target = make_grid(
crop_to_largest(visualize_target, pad_value=0),
nrow=self.cfg.tensorboard.num_images,
scale_each=True,
)
storage.add_image(f"{key_prefix}target",
visualize_target)
self.logger.info(
f"Done evaluation of {curr_dataset_name} at iteration {iter_idx}."
)
self.model.train()
# Log every 20 iterations, or at a validation step or at the end of training.
if iter_idx > 5 and (iter_idx % 20 == 0 or iter_idx %
self.cfg.training.validation_steps == 0 or
(iter_idx + 1) == total_iter):
self.write_to_logs()
storage.step()
def evaluate(
self,
data_loader: DataLoader,
loss_fns: Optional[Dict[str, Callable]],
regularizer_fns: Optional[Dict[str, Callable]] = None,
crop: Optional[str] = None,
is_validation_process=True,
):
self.models_to_device()
self.models_validation_mode()
torch.cuda.empty_cache()
# Variables required for evaluation.
# TODO(jt): Consider if this needs to be in the main engine.py or here. Might be possible we have different
# types needed, perhaps even a FastMRI engine or something similar depending on the metrics.
volume_metrics = self.build_metrics(self.cfg.validation.metrics)
# filenames can be in the volume_indices attribute of the dataset
if hasattr(data_loader.dataset, "volume_indices"):
all_filenames = list(data_loader.dataset.volume_indices.keys())
num_for_this_process = len(
list(data_loader.batch_sampler.sampler.volume_indices.keys()))
self.logger.info(
f"Reconstructing a total of {len(all_filenames)} volumes. "
f"This process has {num_for_this_process} volumes (world size: {communication.get_world_size()})."
)
else:
num_for_this_process = None
filenames_seen = 0
reconstruction_output = defaultdict(list)
targets_output = defaultdict(list)
val_losses = []
val_volume_metrics = defaultdict(dict)
last_filename = None
# Container to for the slices which can be visualized in TensorBoard.
visualize_slices = []
visualize_target = []
visualizations = {}
extra_visualization_keys = (self.cfg.logging.log_as_image
if self.cfg.logging.log_as_image else [])
# Loop over dataset. This requires the use of direct.data.sampler.DistributedSequentialSampler as this sampler
# splits the data over the different processes, and outputs the slices linearly. The implicit assumption here is
# that the slices are outputted from the Dataset *sequentially* for each volume one by one.
time_start = time.time()
for iter_idx, data in enumerate(data_loader):
data = AddNames()(data)
filenames = data.pop("filename")
if len(set(filenames)) != 1:
raise ValueError(
f"Expected a batch during validation to only contain filenames of one case. "
f"Got {set(filenames)}.")
slice_nos = data.pop("slice_no")
scaling_factors = data["scaling_factor"]
resolution = self.compute_resolution(
key=self.cfg.validation.crop,
reconstruction_size=data.get("reconstruction_size", None),
)
# Compute output and loss.
iteration_output = self._do_iteration(
data, loss_fns, regularizer_fns=regularizer_fns)
output = iteration_output.output_image
loss_dict = iteration_output.data_dict
# sensitivity_map = iteration_output.sensitivity_map
loss_dict = detach_dict(loss_dict)
output = output.detach()
val_losses.append(loss_dict)
# Output is complex-valued, and has to be cropped. This holds for both output and target.
output_abs = self.process_output(
output.refine_names(*self.complex_names()),
scaling_factors,
resolution=resolution,
)
if is_validation_process:
target_abs = self.process_output(
data["target"].detach().refine_names(*self.real_names()),
scaling_factors,
resolution=resolution,
)
for key in extra_visualization_keys:
curr_data = data[key].detach()
# Here we need to discover which keys are actually normalized or not
# this requires a solution to issue #23: https://github.com/directgroup/direct/issues/23
del output # Explicitly call delete to clear memory.
# TODO: Is a hack.
# Aggregate volumes to be able to compute the metrics on complete volumes.
for idx, filename in enumerate(filenames):
if last_filename is None:
last_filename = (
filename # First iteration last_filename is not set.
)
# If the new filename is not the previous one, then we can reconstruct the volume as the sampling
# is linear.
# For the last case we need to check if we are at the last batch *and* at the last element in the batch.
is_last_element_of_last_batch = iter_idx + 1 == len(
data_loader) and idx + 1 == len(data["target"])
if filename != last_filename or is_last_element_of_last_batch:
filenames_seen += 1
# Now we can ditch the reconstruction dict by reconstructing the volume,
# will take too much memory otherwise.
# TODO: Stack does not support named tensors.
volume = torch.stack([
_[1].rename(None)
for _ in reconstruction_output[last_filename]
])
if is_validation_process:
target = torch.stack([
_[1].rename(None)
for _ in targets_output[last_filename]
])
curr_metrics = {
metric_name: metric_fn(target, volume)
for metric_name, metric_fn in
volume_metrics.items()
}
val_volume_metrics[last_filename] = curr_metrics
# Log the center slice of the volume
if (len(visualize_slices) <
self.cfg.logging.tensorboard.num_images):
visualize_slices.append(volume[volume.shape[0] //
2])
visualize_target.append(target[target.shape[0] //
2])
# Delete outputs from memory, and recreate dictionary. This is not needed when not in validation
# as we are actually interested in the output
del targets_output
targets_output = defaultdict(list)
del reconstruction_output
reconstruction_output = defaultdict(list)
if all_filenames:
log_prefix = f"{filenames_seen} of {num_for_this_process} volumes reconstructed:"
else:
log_prefix = f"{iter_idx + 1} of {len(data_loader)} slices reconstructed:"
self.logger.info(
f"{log_prefix} {last_filename}"
f" (shape = {list(volume.shape)}) in {time.time() - time_start:.3f}s."
)
# restart timer
time_start = time.time()
last_filename = filename
curr_slice = output_abs[idx].detach()
slice_no = int(slice_nos[idx].numpy())
# TODO: CPU?
reconstruction_output[filename].append(
(slice_no, curr_slice.cpu()))
if is_validation_process:
targets_output[filename].append(
(slice_no, target_abs[idx].cpu()))
# Average loss dict
loss_dict = reduce_list_of_dicts(val_losses)
reduce_tensor_dict(loss_dict)
communication.synchronize()
torch.cuda.empty_cache()
# TODO: Does not work yet with normal gather.
all_gathered_metrics = merge_list_of_dicts(
communication.all_gather(val_volume_metrics))
if not is_validation_process:
return loss_dict, reconstruction_output
# TODO: Apply named tuples where applicable
# TODO: Several functions have multiple output values, in many cases
# TODO: it would be more convenient to convert this to namedtuples.
return loss_dict, all_gathered_metrics, visualize_slices, visualize_target
def evaluate(
self,
data_loader: DataLoader,
loss_fns: Optional[Dict[str, Callable]],
crop: Optional[str] = None,
is_validation_process=True,
):
# TODO(jt): Also log other models output (e.g. sensitivity map).
# TODO(jt): This can be simplified as the sampler now only outputs batches belonging to the same volume.
self.models_to_device()
self.models_validation_mode()
torch.cuda.empty_cache()
# Variables required for evaluation.
# TODO(jt): Consider if this needs to be in the main engine.py or here. Might be possible we have different
# types needed, perhaps even a FastMRI engine or something similar depending on the metrics.
volume_metrics = self.build_metrics(self.cfg.validation.metrics)
reconstruction_output = defaultdict(list)
targets_output = defaultdict(list)
val_losses = []
val_volume_metrics = defaultdict(dict)
last_filename = None
# Container to for the slices which can be visualized in TensorBoard.
visualize_slices = []
visualize_target = []
# Loop over dataset. This requires the use of direct.data.sampler.DistributedSequentialSampler as this sampler
# splits the data over the different processes, and outputs the slices linearly. The implicit assumption here is
# that the slices are outputted from the Dataset *sequentially* for each volume one by one.
for iter_idx, data in enumerate(data_loader):
self.log_process(iter_idx, len(data_loader))
data = AddNames()(data)
filenames = data.pop("filename")
if len(set(filenames)) != 1:
raise ValueError(
f"Expected a batch during validation to only contain filenames of one case. "
f"Got {set(filenames)}.")
slice_nos = data.pop("slice_no")
scaling_factors = data.pop("scaling_factor")
# Check if reconstruction size is the data
if self.cfg.validation.crop == "header":
# This will be of the form [tensor(x_0, x_1, ...), tensor(y_0, y_1,...), tensor(z_0, z_1, ...)] over
# batches.
resolution = [
_.cpu().numpy().tolist()
for _ in data["reconstruction_size"]
]
# The volume sampler should give validation indices belonging to the *same* volume, so it should be
# safe taking the first element, the matrix size are in x,y,z (we work in z,x,y).
resolution = [_[0] for _ in resolution][:-1]
elif self.cfg.validation.crop == "training":
resolution = self.cfg.training.loss.crop
elif not self.cfg.validation.loss.crop:
resolution = None
else:
raise ValueError(
f"Cropping should be either set to `header` to get the values from the header or "
f"`training` to take the same value as training.")
# Compute output and loss.
output, loss_dict = self._do_iteration(data, loss_fns)
val_losses.append(loss_dict)
# Output is complex-valued, and has to be cropped. This holds for both output and target.
output_abs = self.process_output(
output.refine_names(*self.complex_names).detach(),
scaling_factors,
resolution=resolution,
)
if is_validation_process:
target_abs = self.process_output(
data["target"].refine_names(*self.real_names).detach(),
scaling_factors,
resolution=resolution,
)
del output # Explicitly call delete to clear memory.
# TODO: Is a hack.
# Aggregate volumes to be able to compute the metrics on complete volumes.
for idx, filename in enumerate(filenames):
if last_filename is None:
last_filename = (
filename # First iteration last_filename is not set.
)
# If the new filename is not the previous one, then we can reconstruct the volume as the sampling
# is linear.
# For the last case we need to check if we are at the last batch *and* at the last element in the batch.
if filename != last_filename or (
iter_idx + 1 == len(data_loader)
and idx + 1 == len(data["target"])):
# Now we can ditch the reconstruction dict by reconstructing the volume,
# will take too much memory otherwise.
# TODO: Stack does not support named tensors.
volume = torch.stack([
_[1].rename(None)
for _ in reconstruction_output[last_filename]
])
self.logger.info(
f"Reconstructed {last_filename} (shape = {list(volume.shape)})."
)
if is_validation_process:
target = torch.stack([
_[1].rename(None)
for _ in targets_output[last_filename]
])
curr_metrics = {
metric_name: metric_fn(volume, target)
for metric_name, metric_fn in
volume_metrics.items()
}
val_volume_metrics[last_filename] = curr_metrics
# Log the center slice of the volume
if len(visualize_slices
) < self.cfg.tensorboard.num_images:
visualize_slices.append(
normalize_image(volume[volume.shape[0] // 2]))
visualize_target.append(
normalize_image(target[target.shape[0] // 2]))
# Delete outputs from memory, and recreate dictionary. This is not needed when not in validation
# as we are actually interested in the output
del targets_output
targets_output = defaultdict(list)
del reconstruction_output
reconstruction_output = defaultdict(list)
last_filename = filename
curr_slice = output_abs[idx]
slice_no = int(slice_nos[idx].numpy())
# TODO: CPU?
reconstruction_output[filename].append(
(slice_no, curr_slice.cpu()))
if is_validation_process:
targets_output[filename].append(
(slice_no, target_abs[idx].cpu()))
# Average loss dict
loss_dict = reduce_list_of_dicts(val_losses)
reduce_tensor_dict(loss_dict)
communication.synchronize()
torch.cuda.empty_cache()
# TODO(jt): Does not work yet with normal gather.
all_gathered_metrics = merge_list_of_dicts(
communication.all_gather(val_volume_metrics))
if not is_validation_process:
return loss_dict, reconstruction_output
# TODO(jt): Make named tuple
return loss_dict, all_gathered_metrics, visualize_slices, visualize_target
def training_loop(
self,
training_datasets: List, # TODO(jt): Improve typing
start_iter: int,
validation_datasets: Optional[List] = None,
experiment_directory: Optional[pathlib.Path] = None,
num_workers: int = 6,
start_with_validation: bool = False,
):
self.logger.info(f"Local rank: {communication.get_local_rank()}.")
self.models_training_mode()
loss_fns = self.build_loss()
metric_fns = self.build_metrics(self.cfg.training.metrics)
regularizer_fns = self.build_regularizers(self.cfg.training.regularizers)
storage = get_event_storage()
self.ndim = training_datasets[0].ndim
self.logger.info(f"Data dimensionality: {self.ndim}.")
training_data = ConcatDataset(training_datasets)
self.logger.info(f"Concatenated dataset length: {len(training_data)}.")
self.logger.info(
f"Building batch sampler for training set with batch size {self.cfg.training.batch_size}."
)
training_sampler = self.build_batch_sampler(
training_datasets, self.cfg.training.batch_size, "random"
)
data_loader = self.build_loader(
training_data,
batch_sampler=training_sampler,
num_workers=num_workers,
)
# Convenient shorthand
validation_func = functools.partial(
self.validation_loop,
validation_datasets,
loss_fns,
experiment_directory,
num_workers=num_workers,
)
total_iter = self.cfg.training.num_iterations # noqa
fail_counter = 0
for data, iter_idx in zip(data_loader, range(start_iter, total_iter)):
data = AddNames()(data)
if iter_idx == 0:
self.log_first_training_example_and_model(data)
if start_with_validation and iter_idx == start_iter:
self.logger.info(f"Starting with validation at iteration: {iter_idx}.")
validation_func(iter_idx)
try:
iteration_output = self._do_iteration(
data, loss_fns, regularizer_fns=regularizer_fns
)
output = iteration_output.output_image
loss_dict = iteration_output.data_dict
except (ProcessKilledException, TrainingException) as e:
# If the process is killed, the output if saved at state iter_idx, which is the current state,
# so the computation can restart from the last iteration.
self.logger.exception(f"Exiting with exception: {e}.")
self.checkpoint_and_write_to_logs(iter_idx)
sys.exit(-1)
except RuntimeError as e:
# Maybe string can change
if "out of memory" in str(e):
if fail_counter == 3:
self.checkpoint_and_write_to_logs(iter_idx)
raise TrainingException(
f"OOM, could not recover after 3 tries: {e}."
)
fail_counter += 1
self.logger.info(
f"OOM Error: {e}. Skipping batch. Retry {fail_counter}/3."
)
self.__optimizer.zero_grad()
gc.collect()
torch.cuda.empty_cache()
continue
self.checkpoint_and_write_to_logs(iter_idx)
self.logger.info(f"Cannot recover from exception {e}. Exiting.")
raise RuntimeError(e)
if fail_counter > 0:
self.logger.info(f"Recovered from OOM, skipped batch.")
fail_counter = 0
# Gradient accumulation
if (iter_idx + 1) % self.cfg.training.gradient_steps == 0: # type: ignore
if self.cfg.training.gradient_steps > 1: # type: ignore
for parameter in self.model.parameters():
if parameter.grad is not None:
# In-place division
parameter.grad.div_(self.cfg.training.gradient_steps) # type: ignore
if self.cfg.training.gradient_clipping > 0.0: # type: ignore
self._scaler.unscale_(self.__optimizer)
torch.nn.utils.clip_grad_norm_(
self.model.parameters(), self.cfg.training.gradient_clipping
)
# Gradient norm
if self.cfg.training.gradient_debug: # type: ignore
warnings.warn(
f"Gradient debug set. This will affect training performance. Only use for debugging."
f"This message will only be displayed once."
)
parameters = list(
filter(lambda p: p.grad is not None, self.model.parameters())
)
gradient_norm = sum(
[parameter.grad.data ** 2 for parameter in parameters]
).sqrt() # typing: ignore
storage.add_scalar("train/gradient_norm", gradient_norm)
# Same as self.__optimizer.step() for mixed precision.
self._scaler.step(self.__optimizer)
# Updates the scale for next iteration.
self._scaler.update()
# Incorrect inference by mypy and pyflake
self.__lr_scheduler.step() # type: ignore # noqa
storage.add_scalar(
"lr", self.__optimizer.param_groups[0]["lr"], smoothing_hint=False
)
self.__optimizer.zero_grad() # type: ignore
# Reduce the loss over all devices
loss_dict_reduced = communication.reduce_tensor_dict(loss_dict)
loss_reduced = sum(loss_dict_reduced.values())
metrics_dict = evaluate_dict(
metric_fns,
T.modulus_if_complex(output.detach()).rename(None),
data["target"].rename(None).detach().to(self.device),
reduction="mean",
)
metrics_dict_reduced = (
communication.reduce_tensor_dict(metrics_dict) if metrics_dict else {}
)
storage.add_scalars(
loss=loss_reduced, **loss_dict_reduced, **metrics_dict_reduced
)
self.checkpoint_model_at_interval(iter_idx, total_iter)
self.write_to_logs_at_interval(iter_idx, total_iter)
self.validate_model_at_interval(validation_func, iter_idx, total_iter)
storage.step()
def evaluate(self,
data_loader: DataLoader,
loss_fns: Dict[str, Callable],
volume_metrics: Optional[Dict[str, Callable]] = None,
evaluation_round=0):
self.logger.info(f'Evaluating...')
self.model.eval()
torch.cuda.empty_cache()
# Variables required for evaluation.
volume_metrics = volume_metrics if volume_metrics is not None else self.build_metrics()
storage = get_event_storage()
reconstruction_output = defaultdict(list)
targets_output = defaultdict(list)
val_losses = []
val_volume_metrics = defaultdict(dict)
last_filename = None
# Container to for the slices which can be visualized in TensorBoard.
visualize_slices = []
visualize_target = []
# Loop over dataset. This requires the use of direct.data.sampler.DistributedSequentialSampler as this sampler
# splits the data over the different processes, and outputs the slices linearly. The implicit assumption here is
# that the slices are outputted from the Dataset *sequentially* for each volume one by one.
for iter_idx, data in enumerate(data_loader):
self.log_process(iter_idx, len(data_loader))
data = AddNames()(data)
filenames = data.pop('filename')
slice_nos = data.pop('slice_no')
scaling_factors = data.pop('scaling_factor')
# Compute output and loss.
output, loss_dict = self._do_iteration(data, loss_fns)
val_losses.append(loss_dict)
# Output is complex-valued, and has to be cropped. This holds for both output and target.
output_abs = self.process_output(
output.refine_names('batch', 'complex', 'height', 'width').detach(), scaling_factors, 320)
target_abs = self.process_output(
data['target'].refine_names('batch', 'height', 'width').detach(), scaling_factors, 320)
del output # Explicitly call delete to clear memory.
# TODO: Is a hack.
# Aggregate volumes to be able to compute the metrics on complete volumes.
batch_counter = 0
for idx, filename in enumerate(filenames):
if last_filename is None:
last_filename = filename # First iteration last_filename is not set.
# If the new filename is not the previous one, then we can reconstruct the volume as the sampling
# is linear.
# For the last case we need to check if we are at the last batch *and* at the last element in the batch.
if filename != last_filename or (iter_idx + 1 == len(data_loader) and idx + 1 == len(data['target'])):
# Now we can ditch the reconstruction dict by reconstructing the volume,
# will take too mucih memory otherwise.
# TODO: Stack does not support named tensors.
volume = torch.stack([_[1].rename(None) for _ in reconstruction_output[last_filename]])
target = torch.stack([_[1].rename(None) for _ in targets_output[last_filename]])
self.logger.info(f'Reconstructed {last_filename} (shape = {list(volume.shape)}).')
curr_metrics = {
metric_name: metric_fn(volume, target) for metric_name, metric_fn in volume_metrics.items()}
val_volume_metrics[last_filename] = curr_metrics
# Log the center slice of the volume
if len(visualize_slices) < self.cfg.tensorboard.num_images:
visualize_slices.append(normalize_image(volume[volume.shape[0] // 2]))
# Target only needs to be logged once.
if evaluation_round == 0:
visualize_target.append(normalize_image(target[target.shape[0] // 2]))
last_filename = filename
# Delete outputs from memory, and recreate dictionary.
del reconstruction_output
del targets_output
reconstruction_output = defaultdict(list)
targets_output = defaultdict(list)
curr_slice = output_abs[idx]
slice_no = int(slice_nos[idx].numpy())
# TODO: CPU?
reconstruction_output[filename].append((slice_no, curr_slice.cpu()))
targets_output[filename].append((slice_no, target_abs[idx].cpu()))
# Average loss dict
loss_dict = reduce_list_of_dicts(val_losses)
reduce_tensor_dict(loss_dict)
# Log slices.
visualize_slices = make_grid(visualize_slices, nrow=4, scale_each=True)
storage.add_image('validation/prediction', visualize_slices)
if evaluation_round == 0:
visualize_target = make_grid(visualize_target, nrow=4, scale_each=True)
storage.add_image('validation/target', visualize_target)
communication.synchronize()
torch.cuda.empty_cache()
return loss_dict