Resuming from model checkpoints produces different training loss

5

How do you restore so that that dataloader also restores from the right batch.

This seems important for reproducible research.

related: Resume iterating dataloader from checkpoint batch_idx