Correctly save DP/DDP checkpooints

[danieltudosiu]

The correct way of saving DP/DDP checkpoints is to access the module parameter of the class.

Please do that instead of saving the whole DP/DDP class' state dict and then trimming the name.

[kayhan-batmanghelich]

Fork out and make changes and then pull request, Li will look into it.

…

On Mon, Jun 27, 2022 at 6:29 AM Petru-Daniel Tudosiu < ***@***.***> wrote: The correct way of saving DP/DDP checkpoints is to access the module parameter of the class. Please do that instead of saving the whole DP/DDP class' state dict and then trimming the name. — Reply to this email directly, view it on GitHub <#9>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AC53JXKZXN4G25IOK56QHT3VRF7CDANCNFSM5Z57XOQA> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[lisun-ai]

Hi there,

Thanks for your message. We followed the official PyTorch ImageNet training code for saving DP/DDP class' state dict. The name prefix trimming method is commonly adopted in other repos. We will add an annotation to this part. If you have further concerns, please initiate a pull request.

Thanks,
Li

[danieltudosiu]

Hi Li,

From my knowledge of Ignite/MONAI that's not the cleanest way.

https://github.com/pytorch/ignite/blob/master/ignite/handlers/checkpoint.py#L463

Cheers,

Dan