Output checkpoint file of training with both optimizer's parameters and model's one.

[junjihashimoto]

No description provided.

[tscholak]

they are typically saved in two different files

[tscholak]

add rng state to the list

[junjihashimoto]

pytorch saves the RNG state for the current device and the device of all cuda.
https://pytorch.org/docs/stable/checkpoint.html

In this recipe, all the states are put in one file.
https://pytorch.org/tutorials/recipes/recipes/saving_and_loading_a_general_checkpoint.html

Rather than outputting two files (model and optimizer states) separately in torchvision, it seems that it outputs the one file as checkpoint.pth and model_xx.pth, and the latest one is just checkpoint.pth.
https://github.com/pytorch/vision/blob/main/references/detection/train.py#L209-L221

[junjihashimoto]

I just want to have checkpoint-function, so I think it's okay to have one or more.

[junjihashimoto]

@tscholak
Thank you for pointing it out.