An error occurs when running eval_sampled_config.py

[hnz71211]

An error occurs when loading pre-trained weight file (saved in ofa/checkpoints) to Comp-OFAMobileNetV3 net.

python eval_sampled_config.py --net ofa/checkpoints/model_best_compofa_simple.pth.tar --imagenet_path imagenet

The error is:

File "*/CompOFA/eval_sampled_config.py", line 85, in <module>
    net.load_weights_from_net(torch.load(args.net, map_location='cpu')['state_dict'])
  File "*/CompOFA/ofa/elastic_nn/networks/ofa_mbv3.py", line 215, in load_weights_from_net
    raise ValueError(key)

ValueError: blocks.1.mobile_inverted_conv.depth_conv.conv.5to3_matrix

Could you tell me how to solve this problem?

[howardgriffin]

I have the same problem, did you resolve it? @hnz71211

[PhungLai728]

I have the same problem. And the way I overcome it is that I modified the load_weights_from_net() function in ofa_mbv3.py from
for key in src_model_dict: if key in model_dict:
to be
for key in model_dict: if key in src_model_dict:
Hope it help!

[sahnimanas]

Those stray keys can be ignored when loading fixed kernel models, and this has now been fixed in #4
Reading keys from model_dict would bypass the key transformations for other keys (if conditions in later lines), I haven't checked how that affects those cases.