johann-petrak
commented
6 years ago OK, this is probably because a some variables which should get treated as transient and not get saved actually do get saved:
- the validation set should not get saved
- cuda settings should not get saved and get determined as part of loading, then should get set in the pytorch model as well
- possibly others
Check out example dense-pos-seq-1, creates a really huge model pickle file (31M) in comparison to the Pytorch module file (3M)