ModelCheckpoint misbehaves when no validation

[ferrine]

🐛 Bug

ModelCheckpoint misbehaves when there is no validation step. Eg. if you ask to monitor for loss and save every checkpoint with save_top_k=-1 the callback will not do this silently.

To Reproduce

This line of code is a pure indicator of failure without a validation step.

Expected behavior

Callback being called after every epoch

Environment

latest pytorch-lightning

Additional context

I hoped to finetune the model a bit and aggregate the last checkpoints to do SWA. I've fixed the problem using the modified version of ModelCheckpoint

from pytorch_lightning.callbacks.model_checkpoint import ModelCheckpoint as _ModelCheckpoint


class ModelCheckpoint(_ModelCheckpoint):
    def on_epoch_end(self, trainer, pl_module):
        super().on_validation_end(trainer, pl_module)

    def on_validation_end(self, trainer, pl_module):
        pass

I can open a PR if there is no concern about this kind of change

[ferrine]

@SeanNaren should I create a PR so this doesn't take your time?

[ferrine]

Also, I have issues in just creating checkpoints every epoch no matter what metric is