Currently, we save the full pytorch model for each epoch. This can quickly take up a lot of disk space when training large models (aka ones with many layers) as we have to save the weights/biases, along with the model params each time.
Saving just the weights cuts back on the repeated saving of the model parameters, but requires the user to initialize the model correctly again when loading. This is cumbersome as the user might forget the exact model parameters during eval/test time.
Currently, we save the full pytorch model for each epoch. This can quickly take up a lot of disk space when training large models (aka ones with many layers) as we have to save the weights/biases, along with the model params each time.
Saving just the weights cuts back on the repeated saving of the model parameters, but requires the user to initialize the model correctly again when loading. This is cumbersome as the user might forget the exact model parameters during eval/test time.