Weird default filename of ModelCheckpoint : epoch number starts from 0 while step number starts from 1

[zhong-yy]

Bug description

As described in the doc, the default filename of ModelCheckpoint is {epoch}-{step}. But it seems that step number starts from 1, which is weird.

ModelCheckpoint

checkpoint_callback = ModelCheckpoint(
      save_top_k=-1, 
      filename="{epoch}-{step}",  # default filename
      every_n_train_steps=1,      # save every step
      monitor="train_loss", 
      mode="min"
  )

Output:

epoch=0-step=1.ckpt
epoch=0-step=2.ckpt
epoch=0-step=3.ckpt
epoch=0-step=4.ckpt
epoch=0-step=5.ckpt
...

I found a related discussion posted in 2022-Feb, but in that post, the step number seemed to start from 0. Is there any change to the naming convention since that time? I am not sure whether I miss something because I am still new to pytorch-lightning

Environment

Current environment

#- Lightning Component (e.g. Trainer, LightningModule, ModelCheckpoint):
#- PyTorch Lightning Version: 1.9.0
#- PyTorch Version: 1.13.1
#- Python version: 3.9
#- OS: Ubuntu 22.04
#- How you installed Lightning(`conda`, `pip`, source): pip

cc @awaelchli

[zhong-yy]

I think this may be caused by the following change

#11805

It is a very confusing and counter-intuitive behaviour. I couldn't imagine that epoch=0-step=1.ckpt means the first epoch and the first step.

[carmocca]

An epoch is composed of multiple steps. epoch=0, step=1 means 1 step has finished, but the first (0-th) epoch hasn't.
If the first (0-th) epoch had finished, it would be epoch=1.

Step equals 1 because it has finished. Otherwise, it wouldn't save a checkpoint. If you create a Trainer that hasn't fitted yet and manually trigger trainer.save_checkpoint(), then it will have step=0

Do you think this explanation helped?

[zhong-yy]

@carmocca Thank you for the explanation. That makes sense.

[zhong-yy]

Do you mind adding some explanation in the document? I think it would be helpful.