Pytorch-lightning: Validation Loop shall pass the hidden state

[igormq]

🚀 Feature

validation_step function only receives the batch and batch_idx, while the training_step also receives the hiddens when working with sequences. But, when we are evaluating our recurrent LMs, we pass the hidden state through the validation steps and the pytorch-lightning do not permit this.

[github-actions[bot]]

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

[HansBambel]

training_step also only gets batch and batch_idx. Do you mean the amount of values that your dataset gives back?
Similar to #1888 ?

[igormq]

Hi @HansBambel, this is not actually true when you are working with sequences, as you can see in pytorch_lightning/trainer/__init__.py#L953

[HansBambel]

Oh, you're right. Haven't used LSTMs in pytorch lightning yet. Sorry for the confusion!

[igormq]

Don't worry! I am having some issues when using recurrent networks. Actually, I think would be better if we were able to pass between the lifecycle any object we wan't haha.

[williamFalcon]

you can just assign the objects you want to self

self.thing_i_need = ...

[igormq]

Yes, of course I can, but it is brittle, don't you think?

Both training_step and validation_step should have at least the same api regarding the hidden state

[Borda]

@igormq what API do you propose?

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[petrgeiger-incieve]

+1

[ggdupont]

+1