num_training_batches rounds down, causing 0 batches count

[filipeabperes]

🐛 Bug

self.num_training_batches is defined using int here, which rounds it down to 0 when a small training_percent_check or overfit_pct is used, even though at least 1 batch is still processed.

This does not cause any errors in "vanilla" lightning, but crashes any user code that uses the number of batches in a division (for example to get an average of some quantity over batches).

To Reproduce

Steps to reproduce the behavior:

Set the training percentage to a small enough percentage that the number of examples is smaller than the batch size for a given dataset.

This would require a very simple fix, either to use math.ceil() or max(1, self.num_training_batches), depending of how the quantity is expected to behave in the rest of the code.

[awaelchli]

I think max(1, self.num_training_batches) would be the best because it doesn't change the behaviour for existing code that has num_training_batches >= 1. I'm sure the PL team would appreciate if you made a PR. Is the bug also present for validation and test batches?

[kuynzereb]

Well, actually it is not the rounding issue. The problem is that now we process num_training_batches + 1 batches instead of num_training_batches if num_training_batches < len(train_dataloader).