def save_checkpoint_atomic(trainer, final_filename, extra_state):
"""Wrapper around trainer.save_checkpoint to make save atomic."""
temp_filename = os.path.join(final_filename + ".tmp")
trainer.save_checkpoint(temp_filename, extra_state)
# TODO(T56266125): Use mv() instead of copy() + rm() after it's added to
# PathManager.
assert PathManager.copy(
temp_filename, final_filename, overwrite=True
), f"Failed to copy {temp_filename} to {final_filename}"
PathManager.rm(temp_filename)
def load_diverse_ensemble_for_inference(filenames: List[str],
task: Optional[
tasks.FairseqTask] = None):
"""Load an ensemble of diverse models for inference.
This method is similar to fairseq.utils.load_ensemble_for_inference
but allows to load diverse models with non-uniform args.
Args:
filenames: List of file names to checkpoints
task: Optional[FairseqTask]. If this isn't provided, we setup the task
using the first checkpoint's model args loaded from the saved state.
Return:
models, args: Tuple of lists. models contains the loaded models, args
the corresponding configurations.
task: Either the input task or the task created within this function
using args
"""
# load model architectures and weights
checkpoints_data = []
for filename in filenames:
if not PathManager.exists(filename):
raise IOError("Model file not found: {}".format(filename))
with PathManager.open(filename, "rb") as f:
checkpoints_data.append(
torch.load(
f,
map_location=lambda s, l: torch.serialization.
default_restore_location(s, "cpu"),
))
def get_cfg(cp, key):
if "cfg" in cp:
return cp["cfg"][key]
else:
return cp["args"]
# build ensemble
ensemble = []
if task is None:
cfg = get_cfg(checkpoints_data[0], "task")
if hasattr(cfg, "mode"):
cfg.mode = "eval"
task = tasks.setup_task(cfg)
for checkpoint_data in checkpoints_data:
cfg = get_cfg(checkpoint_data, "model")
model = task.build_model(cfg)
model.load_state_dict(checkpoint_data["model"])
ensemble.append(model)
args_list = [get_cfg(s, "model") for s in checkpoints_data]
return ensemble, args_list, task
def _remove_checkpoint(self, checkpoint_to_remove: Optional[str]):
if checkpoint_to_remove:
self.log_if_verbose(
f"| Preparing to remove old checkpoint {checkpoint_to_remove}."
)
try:
PathManager.rm(checkpoint_to_remove)
self.log_if_verbose(
f"| Finished removing old checkpoint {checkpoint_to_remove}."
)
except FileNotFoundError:
print(
f"| Unable to find old checkpoint {checkpoint_to_remove} for removal",
flush=True,
)
def _remove_checkpoint(self, checkpoint_to_remove: Optional[str]):
if checkpoint_to_remove:
self.log_if_verbose(
f"| Preparing to remove old checkpoint {checkpoint_to_remove}."
)
try:
PathManager.rm(checkpoint_to_remove)
self.log_if_verbose(
f"| Finished removing old checkpoint {checkpoint_to_remove}."
)
except OSError as e:
print(
f"| Failed to remove old checkpoint {checkpoint_to_remove} "
f"- exception: {e}",
flush=True,
)
def load_to_gpu(path: str) -> Dict[str, Any]:
"""
Similar to load_to_cpu, but load model to cuda
"""
with PathManager.open(path, "rb") as f:
state = torch.load(
f,
map_location=(lambda s, _: torch.serialization.
default_restore_location(s, "cuda")),
)
return state
def load_to_cpu(path: str) -> Dict[str, Any]:
"""
This is just fairseq's utils.load_checkpoint_to_cpu(), except we don't try
to upgrade the state dict for backward compatibility - to make cases
where we only care about loading the model params easier to unit test.
"""
with PathManager.open(path, "rb") as f:
state = torch.load(
f,
map_location=(lambda s, _: torch.serialization.
default_restore_location(s, "cpu")),
)
return state
def load_existing_checkpoint(
checkpoint_path,
trainer,
restore_state=True) -> Tuple[bool, Optional[Dict]]:
loaded = False
extra_state = None
if not PathManager.isfile(checkpoint_path):
print(f"| No existing checkpoint at {checkpoint_path}. "
f"Starting training from scratch.")
return loaded, extra_state
if restore_state:
extra_state = trainer.load_checkpoint(checkpoint_path)
if extra_state is None:
loaded = False
print(
f"| Failed to load checkpoint and state from {checkpoint_path}."
)
else:
loaded = True
print(f"| Loaded checkpoint {checkpoint_path} "
f"(epoch {extra_state['epoch']}) with restored extra state.")
# batch_offset being None denotes this was a checkpoint saved at
# the end of an epoch (after the last batch).
if extra_state["batch_offset"] is None:
trainer.lr_step(extra_state["epoch"])
extra_state["epoch"] += 1
extra_state["batch_offset"] = 0
else:
dummy_state = trainer.load_checkpoint(checkpoint_path,
reset_optimizer=True)
if dummy_state is None:
loaded = False
print(
f"| Failed to load checkpoint weights from {checkpoint_path}.")
else:
loaded = True
print(f"| Loaded checkpoint weights from {checkpoint_path}.")
return loaded, extra_state
def setup_training_state(args, trainer, task, epoch_itr):
"""Set up the directory for saving checkpoints.
Load pretrained model if specified."""
PathManager.mkdirs(args.save_dir)
# If --restore-file is already present under --save-dir, use that one
# instead of --pretrained-checkpoint-file. The idea is that
# --pretrained-checkpoint-file allows the user to specify restoring from a
# different run's checkpoint (possibly with different training params),
# while not polluting the previous run's checkpoint directory
# with new checkpoints. However, if training gets interrupted
# and the user restarts training, we want to resume from
# the checkpoints under --save-dir, instead of
# restarting again from the old run's checkpoint at
# --pretrained-checkpoint-file.
#
# Note that if args.restore_file is an absolute path, os.path.join() will
# ignore previous directory args and just use the absolute path as is.
checkpoint_path = os.path.join(args.save_dir, args.restore_file)
restore_state = True
if PathManager.isfile(checkpoint_path):
print(
f"| Using --save-dir={args.save_dir}, --restore-file={args.restore_file}."
)
elif args.pretrained_checkpoint_file and PathManager.isfile(
args.pretrained_checkpoint_file):
checkpoint_path = args.pretrained_checkpoint_file
restore_state = args.load_pretrained_checkpoint_state
print(
f"| Using --pretrained-checkpoint-file={args.pretrained_checkpoint_file}, "
f"--load-pretrained-checkpoint-state={args.load_pretrained_checkpoint_state}."
)
extra_state = default_extra_state(args)
if not PathManager.isfile(
checkpoint_path) and args.multi_model_restore_files:
print(
f"| Restoring individual models from {args.multi_model_restore_files}"
)
multi_model.import_individual_models(args.multi_model_restore_files,
trainer)
else:
loaded, loaded_extra_state = checkpoint.load_existing_checkpoint(
checkpoint_path=checkpoint_path,
trainer=trainer,
restore_state=restore_state,
)
if loaded_extra_state:
extra_state.update(loaded_extra_state)
# Reset the start time for the current training run.
extra_state["start_time"] = time.time()
# Skips printing all training progress to prevent log spam.
training_progress = extra_state["training_progress"]
extra_state["training_progress"] = ([
"...truncated...", training_progress[-1]
] if len(training_progress) > 0 else [])
print(f"| extra_state: {extra_state}")
extra_state["training_progress"] = training_progress
epoch = extra_state["epoch"]
if extra_state["batch_offset"] == 0:
epoch -= 1 # this will be incremented when we call epoch_itr.next_epoch_itr()
epoch_itr.load_state_dict({
"epoch":
epoch,
"iterations_in_epoch":
extra_state["batch_offset"]
})
checkpoint_manager = None
if distributed_utils.is_master(args):
checkpoint_manager = checkpoint.CheckpointManager(
num_avg_checkpoints=args.num_avg_checkpoints,
auto_clear_checkpoints=args.auto_clear_checkpoints,
log_verbose=args.log_verbose,
checkpoint_files=extra_state["checkpoint_files"],
)
return extra_state, epoch_itr, checkpoint_manager
def save(
self,
args,
trainer,
extra_state: Dict[str, Any],
new_averaged_params: OrderedDict,
) -> Dict[str, Any]:
"""Saves the model params contained in trainer.
Takes ownership of new_averaged_params, so the caller should not modify
them afterwards.
Args:
trainer: Trainer containing the model to be saved.
extra_state: Dictionary containing any extra information about the
model beyond the param weights.
new_averaged_params: If specified, takes ownership of the params and
sets them as current set of averaged params. If not specified,
we will recalculate the averaged params using the model params
in trainer.
Returns:
Updated extra_state dictionary.
"""
epoch = extra_state["epoch"]
batch_offset = extra_state["batch_offset"]
# batch_offset being None means that we're at the end of an epoch.
if batch_offset is None:
filename = os.path.join(args.save_dir, f"checkpoint{epoch}_end.pt")
# Otherwise, we're in the middle of an epoch.
else:
filename = os.path.join(
args.save_dir, f"checkpoint{epoch}_{batch_offset}.pt"
)
checkpoint_to_remove = self._update_state(
new_params_filename=filename, new_averaged_params=new_averaged_params
)
extra_state["checkpoint_files"] = list(self._checkpoint_files)
self.log_if_verbose(
f"| Preparing to save checkpoints for epoch {epoch}, "
f"offset {batch_offset}."
)
# Saves two copies of the checkpoint - one under a specific name
# corresponding to its epoch/offset, and another under the generic
# "checkpoint_last.py" that we restore from in case training is
# interrupted.
save_checkpoint_atomic(
trainer=trainer, final_filename=filename, extra_state=extra_state
)
# We update checkpoint_last.pt only after the new averaged checkpoint
# and epoch/offset-named copy have been written - so that in case either
# write fails, we'd still be able to resume from the previous
# checkpoint_last.pt
last_checkpoint_path = os.path.join(
args.save_dir, constants.LAST_CHECKPOINT_FILENAME
)
assert PathManager.copy(
filename, last_checkpoint_path, overwrite=True
), f"Failed to copy {filename} to {last_checkpoint_path}"
self.log_if_verbose(
f"| Finished saving checkpoints for epoch {epoch}, "
f"offset {batch_offset}."
)
# Wait until after checkpoint_last.py has been written to remove the
# oldest checkpoint. This is so that in case we fail to write a new
# checkpoint_last.py, we'd still have access to all the files listed in
# the previous checkpoint_last.py
self._remove_checkpoint(checkpoint_to_remove)
return extra_state
