Xception fails to load weights in PyTorch 0.4.0 (with fix)

[fabioperez]

When loading the Xception net with PyTorch 0.4.0, I got the following error:

While copying the parameter named "block1.rep.0.pointwise.weight", whose dimensions in the model are torch.Size([128, 64, 1, 1]) and whose dimensions in the checkpoint are torch.Size([128, 64]).
While copying the parameter named "block1.rep.3.pointwise.weight", whose dimensions in the model are torch.Size([128, 128, 1, 1]) and whose dimensions in the checkpoint are torch.Size([128, 128]).
While copying the parameter named "block2.rep.1.pointwise.weight", whose dimensions in the model are torch.Size([256, 128, 1, 1]) and whose dimensions in the checkpoint are torch.Size([256, 128]).
While copying the parameter named "block2.rep.4.pointwise.weight", whose dimensions in the model are torch.Size([256, 256, 1, 1]) and whose dimensions in the checkpoint are torch.Size([256, 256]).
While copying the parameter named "block3.rep.1.pointwise.weight", whose dimensions in the model are torch.Size([728, 256, 1, 1]) and whose dimensions in the checkpoint are torch.Size([728, 256]).
While copying the parameter named "block3.rep.4.pointwise.weight", whose dimensions in the model are torch.Size([728, 728, 1, 1]) and whose dimensions in the checkpoint are torch.Size([728, 728]).
...

I could fix the problem by modifying the shapes of these layers in the weights file:

import torch
import pretrainedmodels as ptm

state_dict = torch.load('./xception-b5690688.pth')

for name, weights in state_dict.items():
    if 'pointwise' in name:
        state_dict[name] = weights.unsqueeze(-1).unsqueeze(-1)
        
torch.save(state_dict, 'xception-fixed.pth')

Other networks may also result in this error.

[Cadene]

Thanks for submitting this issue. I will try my best to make this lib 0.4.x compatible soon.

[PapaMadeleine2022]

I meet the same error

[Cadene]

Fixed

[AnsonCNS]

I have the same error on torch==1.7.1

I've tried the fix by @fabioperez, but it didn't work.

RuntimeError: Error(s) in loading state_dict for Xception:
        size mismatch for block1.rep.0.pointwise.weight: copying a param with shape torch.Size([128, 64]) from checkpoint, the shape in current model is torch.Size([128, 64, 1, 1]).
        size mismatch for block1.rep.3.pointwise.weight: copying a param with shape torch.Size([128, 128]) from checkpoint, the shape in current model is torch.Size([128, 128, 1, 1]).
        size mismatch for block2.rep.1.pointwise.weight: copying a param with shape torch.Size([256, 128]) from checkpoint, the shape in current model is torch.Size([256, 128, 1, 1]).
        size mismatch for block2.rep.4.pointwise.weight: copying a param with shape torch.Size([256, 256]) from checkpoint, the shape in current model is torch.Size([256, 256, 1, 1]).
        size mismatch for block3.rep.1.pointwise.weight: copying a param with shape torch.Size([728, 256]) from checkpoint, the shape in current model is torch.Size([728, 256, 1, 1]).
        size mismatch for block3.rep.4.pointwise.weight: copying a param with shape torch.Size([728, 728]) from checkpoint, the shape in current model is torch.Size([728, 728, 1, 1]).
        size mismatch for block4.rep.1.pointwise.weight: copying a param with shape torch.Size([728, 728]) from checkpoint, the shape in current model is torch.Size([728, 728, 1, 1]).
        size mismatch for block4.rep.4.pointwise.weight: copying a param with shape torch.Size([728, 728]) from checkpoint, the shape in current model is torch.Size([728, 728, 1, 1]).
        size mismatch for block4.rep.7.pointwise.weight: copying a param with shape torch.Size([728, 728]) from checkpoint, the shape in current model is torch.Size([728, 728, 1, 1]).
        size mismatch for block5.rep.1.pointwise.weight: copying a param with shape torch.Size([728, 728]) from checkpoint, the shape in current model is torch.Size([728, 728, 1, 1]).
        size mismatch for block5.rep.4.pointwise.weight: copying a param with shape torch.Size([728, 728]) from checkpoint, the shape in current model is torch.Size([728, 728, 1, 1]).
        size mismatch for block5.rep.7.pointwise.weight: copying a param with shape torch.Size([728, 728]) from checkpoint, the shape in current model is torch.Size([728, 728, 1, 1]).
        size mismatch for block6.rep.1.pointwise.weight: copying a param with shape torch.Size([728, 728]) from checkpoint, the shape in current model is torch.Size([728, 728, 1, 1]).
        size mismatch for block6.rep.4.pointwise.weight: copying a param with shape torch.Size([728, 728]) from checkpoint, the shape in current model is torch.Size([728, 728, 1, 1]).
        size mismatch for block6.rep.7.pointwise.weight: copying a param with shape torch.Size([728, 728]) from checkpoint, the shape in current model is torch.Size([728, 728, 1, 1]).
        size mismatch for block7.rep.1.pointwise.weight: copying a param with shape torch.Size([728, 728]) from checkpoint, the shape in current model is torch.Size([728, 728, 1, 1]).
        size mismatch for block7.rep.4.pointwise.weight: copying a param with shape torch.Size([728, 728]) from checkpoint, the shape in current model is torch.Size([728, 728, 1, 1]).
        size mismatch for block7.rep.7.pointwise.weight: copying a param with shape torch.Size([728, 728]) from checkpoint, the shape in current model is torch.Size([728, 728, 1, 1]).
        size mismatch for block8.rep.1.pointwise.weight: copying a param with shape torch.Size([728, 728]) from checkpoint, the shape in current model is torch.Size([728, 728, 1, 1]).
        size mismatch for block8.rep.4.pointwise.weight: copying a param with shape torch.Size([728, 728]) from checkpoint, the shape in current model is torch.Size([728, 728, 1, 1]).
        size mismatch for block8.rep.7.pointwise.weight: copying a param with shape torch.Size([728, 728]) from checkpoint, the shape in current model is torch.Size([728, 728, 1, 1]).
        size mismatch for block9.rep.1.pointwise.weight: copying a param with shape torch.Size([728, 728]) from checkpoint, the shape in current model is torch.Size([728, 728, 1, 1]).
        size mismatch for block9.rep.4.pointwise.weight: copying a param with shape torch.Size([728, 728]) from checkpoint, the shape in current model is torch.Size([728, 728, 1, 1]).
        size mismatch for block9.rep.7.pointwise.weight: copying a param with shape torch.Size([728, 728]) from checkpoint, the shape in current model is torch.Size([728, 728, 1, 1]).
        size mismatch for block10.rep.1.pointwise.weight: copying a param with shape torch.Size([728, 728]) from checkpoint, the shape in current model is torch.Size([728, 728, 1, 1]).
        size mismatch for block10.rep.4.pointwise.weight: copying a param with shape torch.Size([728, 728]) from checkpoint, the shape in current model is torch.Size([728, 728, 1, 1]).
        size mismatch for block10.rep.7.pointwise.weight: copying a param with shape torch.Size([728, 728]) from checkpoint, the shape in current model is torch.Size([728, 728, 1, 1]).
        size mismatch for block11.rep.1.pointwise.weight: copying a param with shape torch.Size([728, 728]) from checkpoint, the shape in current model is torch.Size([728, 728, 1, 1]).
        size mismatch for block11.rep.4.pointwise.weight: copying a param with shape torch.Size([728, 728]) from checkpoint, the shape in current model is torch.Size([728, 728, 1, 1]).
        size mismatch for block11.rep.7.pointwise.weight: copying a param with shape torch.Size([728, 728]) from checkpoint, the shape in current model is torch.Size([728, 728, 1, 1]).
        size mismatch for block12.rep.1.pointwise.weight: copying a param with shape torch.Size([728, 728]) from checkpoint, the shape in current model is torch.Size([728, 728, 1, 1]).
        size mismatch for block12.rep.4.pointwise.weight: copying a param with shape torch.Size([1024, 728]) from checkpoint, the shape in current model is torch.Size([1024, 728, 1, 1]).
        size mismatch for conv3.pointwise.weight: copying a param with shape torch.Size([1536, 1024]) from checkpoint, the shape in current model is torch.Size([1536, 1024, 1, 1]).
        size mismatch for conv4.pointwise.weight: copying a param with shape torch.Size([2048, 1536]) from checkpoint, the shape in current model is torch.Size([2048, 1536, 1, 1]).