weight file problem

[ztyf-lq]

Hello, I have carefully read your wonderful and unparalleled article. But while running the article code, I found that your weight file on the cloud drive is damaged. Could you please update the weight file on the cloud drive?

[lliutianc]

The issue comes from that the checkpoints didn't use module but initialized one does. So the key names don't match.

Manually updated key names should be a quick fix. For example, add below snippets to the resume function in BaseModel

        new_stat = {}
        for key, val in current_checkpoint['state_dict'].items():
            if "module" not in key:
                new_key = key.split(".")
                new_key.insert(1, "module")
                new_key = ".".join(new_key)
                new_stat[new_key] = val
            else:
                new_stat[key] = val
        current_checkpoint['state_dict'] = new_stat

somewhere between

        current_checkpoint = torch.load(resume_path)

and

        self.model.load_state_dict(current_checkpoint['state_dict'], strict=True)

should work.

[ztyf-lq]

Thank you for your answer. My question is that there was an error in decompressing the weight file after downloading, not during loading. Thank you again

[lliutianc]

Oh never mind, I didn't encounter any issue with the checkpoint downloaded from google drive. Probably have a try again?