Pytorch-lightning: Evaluation over the validation set

[EliaCereda]

What is the recommended way of performing just one evaluation over the validation set? Basically, I'm looking for the equivalent of trainer.test(...) but for the validation set.

Maybe this is possible using trainer.fit(...) and some combination of arguments to Trainer.__init__?

[SkafteNicki]

Is there a difference between your validation_step and test_step? If not you could just feed in the validation set to trainer.test(test_dataloaders=val_dataloaders).

[EliaCereda]

Yes, this would be acceptable, but I forgot to specify that I'm using a data module to manage the data loaders. There is no way to pass the data module but override which loader to use, am I wrong?

The solution I'll attempt right now is to copy the trainer.test(...) method in my code and modify it to use the validation loader and validation_step. Let me know if you have a better idea.

[rohitgr7]

should we add an evaluate method too exactly similar to test? recently got same request on slack too.

[EliaCereda]

It sounds good to me. I'm looking at the code of trainer.test(...) right now and it looks like this isn't something I can easily add from my own code: the interesting part is inside the fit(...) method, triggered by self.testing = True.

[SkafteNicki]

@rohitgr7 sounds like a good enhancement (especially if multiple people are requesting it). I see two options:
1) add a new method: i propose to call it validate that works similar to test
2) add an extra argument to test if it should use test_dataloader or validation_dataloader

[rohitgr7]

add an extra argument to test if it should use test_dataloader or validation_dataloader

thought this too.. but somehow sounds a bit complex to me.

add a new method: i propose to call it validate that works similar to test

+1, refactor the .test method to avoid duplicate code and reuse in .validate

[EliaCereda]

+1 for option 1 from me too

[rohitgr7]

cc @PyTorchLightning/core-contributors

[EliaCereda]

I've been experimenting with the changes required to make this work and I just published #4707 with a very rough proof of concept of what will need to be done.