CheckpointSaver saves model checkpoints for key_metric_n_saved number of best validation score and n_savednumber of the last model. While resuming training, the filename buffer starts from empty. Thus it creates a whole new set of checkpoint files after each resume. Replacing the previous checkpoints would be a cool feature to have for conveniently resuming training. Thanks.
CheckpointSaversaves model checkpoints forkey_metric_n_savednumber of best validation score andn_savednumber of the last model. While resuming training, the filename buffer starts from empty. Thus it creates a whole new set of checkpoint files after each resume. Replacing the previous checkpoints would be a cool feature to have for conveniently resuming training. Thanks.