Hi,
It is quite a small details, though useful to know. The pre-trained model won't work in PyTorch 0.4.* because normalization layers now track the number of batches with the tensor num_batches_tracked [see code]. When loading a state_dict, it then expects these tensors and crashes on a KeyError. Maybe just adding the pytorch version in requirements could help future users.
Thanks
Hi, It is quite a small details, though useful to know. The pre-trained model won't work in PyTorch 0.4.* because normalization layers now track the number of batches with the tensor
num_batches_tracked
[see code]. When loading astate_dict
, it then expects these tensors and crashes on aKeyError
. Maybe just adding the pytorch version in requirements could help future users. Thanks