How to load models if I do not know its structure?

1

I know torch.save(net.state_dict(), model_path) can only save the weights, and torch.save(net, model_path) can save the weights and the structure of the network.
I know ·torch.load(model_path)· can load the pretrained model.
I know net.load_state_dict(torch.load(model_path)) can map the pretrained weights to the current net.

But I found it is necessary to define the net before I load the net from disk. Otherwise, error would accur: AttributeError: Can't get attribute 'VGG19' on <module '__main__'>. I know this is because I know torch.save() is a wraper of python's pickle, and the limitation is because of pickle itself .

My question is:

How to load models if I do not know its structure/code?

1 Like
2

This is currently not possible and

will not directly save the structure of the model, but would still depend on the source files (and could thus break in various ways).
You could check the experimental torch.package utility, which might fit your use case, but note that it’s experimental and you might thus encounter issues with it.

1 Like