save_last does not actually save the last checkpoint

[Sukh-P]

Describe the bug

Details of the bug can be found in this conversation here but essentially save_last actually saves a copy of the last checkpoint that was saved rather than the checkpoint for the last epoch.

We should change this so we actually do save the last checkpoint, by e.g. either by having a customer saver or just saving out the checkpoint for the current epoch separately to the best one and keep overwriting so we keep the last one (might be done by a custom saver)

[CodeVishal-17]

Hi! I'd like to work on this issue. I can reproduce the current save_last behavior, trace the checkpoint-saving flow, and implement the fix so that last always points to the checkpoint from the latest epoch. I'll also add/update regression tests to verify the behavior. Please assign this to me.

[ocf-bot]

@CodeVishal-17 Thank you for reaching out! Unfortunately, this issue is not suitable for open source contribution and so hasn't been marked with "good first issue" or "contributions welcome". Please consult our contribution guidelines for more information on how you can contribute, and hope to see you around!

[Sukh-P]

@CodeVishal-17 sure have at it but you may find this a bit hard to replicate since you will need to train pvnet with some open source dataset it supports to be able to actually see this issue/test the fix

[CodeVishal-17]

@Sukh-P Thanks for the clarification! That makes sense. I’ll skip this issue for now since reproducing and testing it requires the trained pvnet setup. Appreciate the explanation, and I’ll look for another issue that’s better suited for external contribution.

[ocf-bot]

@CodeVishal-17 Thank you for reaching out! Unfortunately, this issue is not suitable for open source contribution and so hasn't been marked with "good first issue" or "contributions welcome". Please consult our contribution guidelines for more information on how you can contribute, and hope to see you around!