PyTorch: check point support

[tensorbuffer]

Hello, when I try to use the web version to load pytorch checkpoint file, I got this error: Error loading PyTorch model. File does not contain root module or state dictionary in 'checkpoint.pth'.
The checkpoint is saved as:
checkpoint = {
'model': model_without_ddp.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'epoch': epoch,
'args': args}
torch.save(
checkpoint, path_to_file)

This is pytorch standard way of saving checkpoint:
https://pytorch.org/tutorials/beginner/saving_loading_models.html#saving-loading-a-general-checkpoint-for-inference-and-or-resuming-training

[lutzroeder]

Can you share the model file?

[tensorbuffer]

https://drive.google.com/file/d/1eIyTCGctlInencZBUwvBokPQikGz03Oq/view?usp=sharing

[lutzroeder]

features.1.skip_add.observer.scale tensor is null.

[tensorbuffer]

great, thanks! I can load the checkpoint file.
However all that's shown are 'Module' blocks, there's no connection between them. Is this the right visualization?

[abhigoku10]

@tensorbuffer @lutzroeder @demid5111 @scottcjt i am also facing the same issues not able to visualize any connection between the modules , any idea how to obtain the connections ??

[tensorbuffer]

I think this is because pytorch graph is dynamic, the graph is generated on the fly, so you have to feed an input and execute the model to get the graph.

[simphide]

Hi, I also have this problem with the models of this project: https://github.com/BloodAxe/Kaggle-2020-Alaska2/releases

Does anybody have an idea how I can see the connections etc.? Thank you!

[lutzroeder]

@simphide the files include only training weights. See #720.