Error while loading pretrained Inception (v 0.2.1)

[DSLituiev]

Dimension mismatch error while loading a pretrained Inception v3

Environment

0.2.1 with Anaconda, python3.6, Ubuntu 16

code

from torchvision.models.inception import Inception3, inception_v3
model = inception_v3(num_classes=1, pretrained=True, aux_logits=False)

error

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-31-e0b4dcd6b958> in <module>()
----> 1 model = inception_v3(num_classes=1, pretrained=True, aux_logits=False)

/anaconda3/lib/python3.6/site-packages/torchvision-0.2.1-py3.6.egg/torchvision/models/inception.py in inception_v3(pretrained, **kwargs)
     25             kwargs['transform_input'] = True
     26         model = Inception3(**kwargs)
---> 27         model.load_state_dict(model_zoo.load_url(model_urls['inception_v3_google']))
     28         return model
     29 

/anaconda3/lib/python3.6/site-packages/torch/nn/modules/module.py in load_state_dict(self, state_dict, strict)
    719         if len(error_msgs) > 0:
    720             raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
--> 721                                self.__class__.__name__, "\n\t".join(error_msgs)))
    722 
    723     def parameters(self):

RuntimeError: Error(s) in loading state_dict for Inception3:
	Unexpected key(s) in state_dict: "AuxLogits.conv0.conv.weight", "AuxLogits.conv0.bn.weight", "AuxLogits.conv0.bn.bias", "AuxLogits.conv0.bn.running_mean", "AuxLogits.conv0.bn.running_var", "AuxLogits.conv1.conv.weight", "AuxLogits.conv1.bn.weight", "AuxLogits.conv1.bn.bias", "AuxLogits.conv1.bn.running_mean", "AuxLogits.conv1.bn.running_var", "AuxLogits.fc.weight", "AuxLogits.fc.bias". 
	While copying the parameter named "fc.weight", whose dimensions in the model are torch.Size([1, 2048]) and whose dimensions in the checkpoint are torch.Size([1000, 2048]).
	While copying the parameter named "fc.bias", whose dimensions in the model are torch.Size([1]) and whose dimensions in the checkpoint are torch.Size([1000]).

[moskomule]

It's because you specify num_classes=1. Try inception_v3(num_classes=1000, pretrained=True) instead. Then, if you like 1 class output, replace the model.fc.

[fmassa]

@moskomule 's answer will solve your issue. Closing as not a bug

[DSLituiev]

It is though not explicitly mentioned in the docs, and therefore I believe it is unexpected behaviour, in simple words a bug. There is a simple way around it, and I wonder if maintainers would consider a pull request to fix it?

[fmassa]

You mean a PR in examples? Or a PR in torchvision mentioning that inception requires a different input size? I'd vote for the second option

[fmassa]

I think a PR for torchvision would be great actually!

[Kachakhiladi]

I have the same error as the original poster.

The error is not due to the num_classes. It is due to aux_logits=False instead of the default aux_logits=True.
Any leads ?

[XiaolingLuo]

You should not modify the "model = models.inception_v3(pretrained=True)", because if you change the code here, it will have errors in "load_state_dict".
Try:
"model = models.inception_v3(pretrained=True)
model.aux_logits = False
num_ftrs = model.fc.in_features
model.fc = nn.Linear(num_ftrs, nclasses) "

[luciagan]

Dear Shirley, Thanks for your reply. Best, Lei Gan On Sep 1, 2019, at 16:24, ShirleyLuo <notifications@github.com<mailto:notifications@github.com>> wrote: You should not modify the "model = models.inception_v3(pretrained=True)", because if you change the code here, it will have errors in "load_state_dict". Try: "model = models.inception_v3(pretrained=True) model.aux_logits = False num_ftrs = model.fc.in_features model.fc = nn.Linear(num_ftrs, nclasses) " — You are receiving this because you commented. Reply to this email directly, view it on GitHub<#480?email_source=notifications&email_token=AF4RU5HI2H2AHQ7WJSX6QDDQHN327A5CNFSM4E4ZSJUKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD5T5KHA#issuecomment-526898460>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AF4RU5HE2S5MX5TM7NMSTQDQHN327ANCNFSM4E4ZSJUA>.