When loading a model Unexpected key(s) in state_dict: "module.features.conv1.0.weight",

1

Hi,
I saved a model in one computer and then loaded it in another.

Missing key(s) in state_dict: “features.conv1.0.weight”,

Unexpected key(s) in state_dict: “module.features.conv1.0.weight”,

Why is it using “module” as a prefix when loading the model? Does this have to do anything with namespaces?
It is the same CNN on both computers using the same PyTorch version

thanks,

2

Usually module will be added when you’ve wrapped your model into DataParallel.
You could add your model temporarily into DataParallel to load the weights or create a new state_dict by filtering out the module prefix.

2 Likes
3

Yes, that is true, the model originated from training using:

model = torch.nn.DataParallel(model).cuda()

However, the line that throws the exception is the line the loads the model, like so:

        if os.path.isfile(mdl):
            print("=> loading checkpoint '{}'".format(mdl))
            **model.load_state_dict(torch.load(mdl))**
            model = torch.nn.DataParallel(model).cuda()                
        else:
            print("=> no checkpoint found at '{}'".format(mdl))

Is that what you meant by?:
add your model temporarily into DataParallel

Thanks,

4

Never mind, resolved, thanks.

5

In case someone has this problem, I wrote a function to load pretrained weights, which can ignore ‘module.’ and different layer names https://github.com/KaiyangZhou/deep-person-reid/blob/master/torchreid/utils/torchtools.py#L104. Can be used directly by copy and paste.