What is the correct way to calculate validation loss?

Hi! I was wondering what is the correct way to calculate validation loss when using the ImagenTrainer without using the valid_step method. I’m having trouble using that method because I’m not sure what the format of the returned data from the validation data loader should be. (I normally don’t use a training dataloader)

I tried doing:

with torch.no_grad():
     # call trainer here

but that doesn’t make sense since it looks like the trainer updates the weights inside of its forward method. I’m thinking of trying:

with torch.no_grad():
     validation_loss = trainer.imagen(images=foo, texts=bar)

Although that seems like it might be missing some of the ema things.

Side note, thanks for answering all of my questions. It’s been super helpful 😃

[vedantroy]

@lucidrains Oh, and presumably you’d also want to wrap the entire validation loop inside of a torch.no_grad() I assume? These questions are kind of obvious, but I’m asking b/c it seems like the trainer takes care of a ton of things, so it might also just take care of turning on torch.no_grad (maybe).

[lucidrains]

@vedantroy yup, correct!