Difference between on_batch_start and on_train_batch_start?

[daniellepintz]

What is the difference between on_batch_start and on_train_batch_start? Same question for on_batch_end and on_train_batch_end.

https://github.com/PyTorchLightning/pytorch-lightning/blob/49d03f87fed0458cbb146d38243be56be4cb9689/pytorch_lightning/trainer/callback_hook.py#L155-L173

If they are different, can we make the docstrings different?

If they are the same, can we deprecate one of them?

[rohitgr7]

I remember this was discussed a few months back, on_batch_* was supposed to run for all train/valid/test/predict, but looks like it just runs in case of training. Looks redundant now in case there are no plans to call it irrespective of running_stage.

[daniellepintz]

thanks @rohitgr7! that's what I thought too. Does anyone have any preference of either deprecation or making it work as intended (run for all train/valid/test/predict)? @ananthsub @tchaton @carmocca @awaelchli @SeanNaren @kaushikb11

[carmocca]

It should work as intended. There are several TODOs in this file pointing the inconsistencies.

[daniellepintz]

I see, thanks! Is there something blocking us from doing those TODOs? Also should we update the docstrings to """Called when a batch begins (train/valid/predict/test).""

[carmocca]

Is there something blocking us from doing those TODOs

Nothing! Feel free to create an issue for us to discuss/approve and then work on fixing them.

should we update the docstrings

Yes