Int num_sanity_val_steps is always replaced by float limit_val_batches

[manipopopo]

The type annotations of num_sanity_val_steps: and limit_val_batches are int and Union[int, float] for percent or num_batches. The minimum of percent and num_batches will be percent. (except when num_batches==0)
https://github.com/PyTorchLightning/pytorch-lightning/blob/a59e140ee814d8818d121405582133cf6b767e1a/pytorch_lightning/trainer/trainer.py#L461

Maybe we can remove the dependency of limit_val_batches from num_sanity_val_steps and revert to
https://github.com/PyTorchLightning/pytorch-lightning/blob/1e68968ed7fb9b8f73df148dd48194d469655ea3/pytorch_lightning/trainer/trainer.py#L491

[github-actions]

Hi! thanks for your contribution!, great first issue!

[ananyahjha93]

@williamFalcon @Borda should num_sanity_val_steps be limited by limit_val_batches? My understanding is they can be independent of one another.

[awaelchli]

@ananyahjha93 I asked william some time ago: #2246 (comment)

[rohitgr7]

@awaelchli Here: #2246 (comment)
Shouldn't Trainer(num_sanity_val_steps=5, limit_val_batches=3) run 3 sanity val steps?

[rohitgr7]

@awaelchli any comments on above? We can fix it then :)

[awaelchli]

It depends. What does it currently do in your example?

[rohitgr7]

@awaelchli In the #2246 (comment) you said it should run for 5 val_steps but here https://github.com/PyTorchLightning/pytorch-lightning/blob/0097630a95bddc48d6fb5d3b9a58aef2e8e89b22/pytorch_lightning/trainer/trainer.py#L463-L466 it seems that it's running for 3 val_steps.

I suggest num_sanity_val_steps should be updated according to limit_val_batches(both float and int) and then we can merge #2892 accordingly.

[williamFalcon]

num_sanity_val_steps should be int or -1.

this is really meant to be a sanity check... floats will break this.

You need this min because if your dataset is smaller than the limit_val_batches it will crash