in exportsd.py, call detach() before numpy()

[yueyinqiu]

An error occurred when trying to save the downloaded pretained model:

states = weight.get_state_dict()
exportsd.save_state_dict(states, file)

(where weight is something like torchvision.models.AlexNet_Weights.IMAGENET1K_V1)

(full codes here)

[yueyinqiu]

I'm not sure why the tensors are in computation graphs... but the error do occur and it could be solved by adding detach.

[GeorgeS2019]

@yueyinqiu
/Are we missing unit tests to check for this bug?

[NiklasGustafsson]

@yueyinqiu /Are we missing unit tests to check for this bug?

Right. We have no unit tests for the Python export / import code at all. I've just assumed it works... :-)

[yueyinqiu]

Actually I think a simple unit test could not find this out. In most cases the problem does not occur and I don't know why the pretrained model are in computation graphs...

Also I'm not really sure whether we should modify this, since we do not use cpu() as well. Perhaps this should be done by users before saving. Maybe more checks should be made before merging.

[NiklasGustafsson]

Calling detach seems like a safe, harmless thing to do.

[yueyinqiu]

just to make a note here: torchvision.models.AlexNet_Weights.IMAGENET1K_V1.get_state_dict()['features.0.weight'].requires_grad is True

[yueyinqiu]

Calling detach seems like a safe, harmless thing to do.

However it seems that torch.save/torch.load will record whether a tensor requires_grad or not in PyTorch. I'm testing on that now.

[yueyinqiu]

Well I think it's fine. The .net format does not save requires_grad at all. And the current PyTorch's state_dict() implemention is using detach() before putting the tensor into the state dictionary, so definitely requires_grad won't be accessed when loading it. Perhaps it's a behavior exists in some old PyTorch versions.

[NiklasGustafsson]

Ready to merge?

[yueyinqiu]

Ready to merge?

I suppose yes