dgs.utils.torchtools.resume_from_checkpoint

dgs.utils.torchtools.resume_from_checkpoint(fpath: str, model: TorchMod | BaseMod, optimizer: torch.optim.Optimizer | None = None, scheduler: torch.optim.lr_scheduler.LRScheduler | None = None, verbose: bool = False) → int[source]

Resumes training from a checkpoint.

This will load (1) model weights and (2) state_dict of optimizer if optimizer is not None.

Parameters:

  • fpath – The path to checkpoint. Can be a local or absolute path.

  • model – The model that is currently trained.

  • optimizer – An Optimizer.

  • scheduler – A single LRScheduler.

  • verbose – Whether to print additional debug information.

Returns:

start_epoch.

Return type:

int

Examples

>>> from dgs.utils.torchtools import resume_from_checkpoint
>>> fpath = 'log/my_model/model.pth.tar-10'
>>> start_epoch = resume_from_checkpoint(
>>>     fpath, model, optimizer, scheduler
>>> )