Open arpanmangal opened 4 years ago
Continuing discussing from #1
@SuryaKalia FYI
@bartmch
Yes keeping num_batches_tracked
as the default value of zero tensors is the best which could be done imo.
Further the keys not only differ in a module.
string, but are also renamed like one has ...conv1
, while other version has ...conv.1
. Print the keys and compare to see all such variations.
transfer.py code is not working plz check ASAP There is a problem loading pytorch 0.3.0 and torchvision 0.2.0 plz add support
Present code requires
pytorch0.3
andCUDA<=8.0
. The code needs to be updated to supportpytorch>1.0
. The main reason for usingpytorch0.3
is compatibility with trained CheXNet model.