def save_checkpoint(self, filename, extra_state):
"""Save all training state in a checkpoint file."""
if distributed_utils.is_master(self.args): # only save one checkpoint
extra_state['train_meters'] = self.meters
utils.save_state(
filename, self.args, self.model, self.criterion, self.optimizer,
self.lr_scheduler, self._num_updates, self._optim_history, extra_state,
)
def save_checkpoint(self, filename, extra_state):
"""Save all training state in a checkpoint file."""
if distributed_utils.is_master(self.args): # only save one checkpoint
extra_state['train_meters'] = self.meters
utils.save_state(
filename, self.args, self.model, self.criterion, self.optimizer,
self.lr_scheduler, self._num_updates, self._optim_history, extra_state,
)
def save_checkpoint(self, filename, extra_state):
"""Save all training state in a checkpoint file."""
if self.args.amp:
extra_state['amp_state_dict'] = amp.state_dict()
extra_state['amp_master_params'] = list(amp.master_params(self.optimizer.optimizer))
if distributed_utils.is_master(self.args): # only save one checkpoint
utils.save_state(
filename, self.args, self.get_model(), self.criterion, self.optimizer,
self.lr_scheduler, self._num_updates, self._optim_history, extra_state,
)
def save_checkpoint(self, filename, extra_state):
"""Save all training state in a checkpoint file."""
if distributed_utils.is_master(self.args): # only save one checkpoint
extra_state['train_meters'] = self.meters
if not self.args.sep_optim:
utils.save_state(
filename, self.args, self.get_model().state_dict(), self.criterion, self.optimizer,
self.lr_scheduler, self._num_updates, self._optim_history, extra_state
)
else:
utils.save_state_for_sep_optim(filename, self.args, self.get_model().state_dict(), self.criterion, self.optimizer,
self.lr_scheduler, self._num_updates, self.dec_optimizer, self.dec_lr_scheduler,
self._optim_history, self._dec_optim_history, extra_state)
def save_checkpoint(self, filename, extra_state):
if distributed_utils.is_master(self.args): # only save one checkpoint
extra_state['train_meters'] = self.meters
utils.save_state(
filename,
self.args,
self.get_model().state_dict(),
self.criterion,
self.optimizer,
self.lr_scheduler,
self._num_updates,
self._optim_history,
extra_state,
)
def save_checkpoint(self, filename, extra_state):
"""Save all training state in a checkpoint file."""
if self.args.distributed_rank == 0: # only save one checkpoint
utils.save_state(filename, self.args, self.model, self.criterion, self.optimizer,
self.lr_scheduler, self._num_updates, self._optim_history, extra_state)
def _async_save_checkpoint(self, rank, device_id, filename, extra_state):
utils.save_state(filename, self.args, self.model, self.criterion,
self.optimizer, self.lr_scheduler,
self._optim_history, extra_state)
