def hpc_save(self, folderpath: str, logger):
# make sure the checkpoint folder exists
folderpath = str(folderpath) # because the tests pass a path object
if not gfile.exists(folderpath):
makedirs(folderpath)
# save logger to make sure we get all the metrics
logger.save()
ckpt_number = self.max_ckpt_in_folder(folderpath) + 1
if not gfile.exists(folderpath):
makedirs(folderpath)
filepath = os.path.join(folderpath, f'hpc_ckpt_{ckpt_number}.ckpt')
# give model a chance to do something on hpc_save
model = self.get_model()
checkpoint = self.dump_checkpoint()
model.on_hpc_save(checkpoint)
# do the actual save
# TODO: fix for anything with multiprocess DP, DDP, DDP2
try:
atomic_save(checkpoint, filepath)
except AttributeError as err:
if LightningModule.CHECKPOINT_HYPER_PARAMS_KEY in checkpoint:
del checkpoint[LightningModule.CHECKPOINT_HYPER_PARAMS_KEY]
rank_zero_warn(
'warning, `module_arguments` dropped from checkpoint.'
f' An attribute is not picklable {err}')
atomic_save(checkpoint, filepath)
return filepath
def _save_model(self, filepath, trainer, pl_module):
# in debugging, track when we save checkpoints
trainer.dev_debugger.track_checkpointing_history(filepath)
# make paths
if not gfile.exists(os.path.dirname(filepath)):
makedirs(os.path.dirname(filepath))
# delegate the saving to the model
if self.save_function is not None:
self.save_function(filepath, self.save_weights_only)
else:
raise ValueError(".save_function() not set")
def experiment(self) -> SummaryWriter:
r"""
Actual tensorboard object. To use TensorBoard features in your
:class:`~pytorch_lightning.core.lightning.LightningModule` do the following.
Example::
self.logger.experiment.some_tensorboard_function()
"""
if self._experiment is not None:
return self._experiment
assert rank_zero_only.rank == 0, 'tried to init log dirs in non global_rank=0'
if self.root_dir and not gfile.exists(str(self.root_dir)):
makedirs(self.root_dir)
self._experiment = SummaryWriter(log_dir=self.log_dir, **self._kwargs)
return self._experiment
def on_train_start(self, trainer, pl_module):
"""
Determines model checkpoint save directory at runtime. References attributes from the
trainer's logger to determine where to save checkpoints.
The base path for saving weights is set in this priority:
1. Checkpoint callback's path (if passed in)
2. The default_root_dir from trainer if trainer has no logger
3. The weights_save_path from trainer, if user provides it
4. User provided weights_saved_path
The base path gets extended with logger name and version (if these are available)
and subfolder "checkpoints".
"""
if self.dirpath is not None:
return # short circuit
self.filename = '{epoch}'
if trainer.logger is not None:
if trainer.weights_save_path != trainer.default_root_dir:
# the user has changed weights_save_path, it overrides anything
save_dir = trainer.weights_save_path
else:
save_dir = trainer.logger.save_dir or trainer.default_root_dir
version = trainer.logger.version if isinstance(
trainer.logger.version,
str) else f'version_{trainer.logger.version}'
ckpt_path = os.path.join(save_dir, trainer.logger.name, version,
"checkpoints")
else:
ckpt_path = os.path.join(trainer.weights_save_path, "checkpoints")
self.dirpath = ckpt_path
assert trainer.global_rank == 0, 'tried to make a checkpoint from non global_rank=0'
if not gfile.exists(self.dirpath):
makedirs(self.dirpath)
def __init__(self,
filepath: Optional[str] = None,
monitor: str = 'val_loss',
verbose: bool = False,
save_last: bool = False,
save_top_k: int = 1,
save_weights_only: bool = False,
mode: str = 'auto',
period: int = 1,
prefix: str = ''):
super().__init__()
if (filepath):
filepath = str(
filepath
) # the tests pass in a py.path.local but we want a str
if save_top_k > 0 and filepath is not None and gfile.isdir(
filepath) and len(gfile.listdir(filepath)) > 0:
rank_zero_warn(
f"Checkpoint directory {filepath} exists and is not empty with save_top_k != 0."
"All files in this directory will be deleted when a checkpoint is saved!"
)
self._rank = 0
self.monitor = monitor
self.verbose = verbose
if filepath is None: # will be determined by trainer at runtime
self.dirpath, self.filename = None, None
else:
if gfile.isdir(filepath):
self.dirpath, self.filename = filepath, '{epoch}'
else:
if not is_remote_path(filepath): # dont normalize remote paths
filepath = os.path.realpath(filepath)
self.dirpath, self.filename = os.path.split(filepath)
makedirs(self.dirpath) # calls with exist_ok
self.save_last = save_last
self.save_top_k = save_top_k
self.save_weights_only = save_weights_only
self.period = period
self.epoch_last_check = None
self.prefix = prefix
self.best_k_models = {}
# {filename: monitor}
self.kth_best_model_path = ''
self.best_model_score = 0
self.best_model_path = ''
self.save_function = None
self.warned_result_obj = False
torch_inf = torch.tensor(np.Inf)
mode_dict = {
'min': (torch_inf, 'min'),
'max': (-torch_inf, 'max'),
'auto': (-torch_inf, 'max') if 'acc' in self.monitor
or self.monitor.startswith('fmeasure') else (torch_inf, 'min'),
}
if mode not in mode_dict:
rank_zero_warn(
f'ModelCheckpoint mode {mode} is unknown, '
f'fallback to auto mode.', RuntimeWarning)
mode = 'auto'
self.kth_value, self.mode = mode_dict[mode]
