When doing fit(), self.training in forward() keeps turning into False?

[HaFred]

Hi all, I tried to train a model with pl. And I just ran the trainer.fit() as below:

trainer.fit(model, train_dataloaders=model.train_dataloader(),
                val_dataloaders=model.val_dataloader())

And I found that model.training == False, when it gets into forward()...
Is there any solution or does anybody know the potential reason for this?
Or does anyone know where can I find the source code for training_step() so that I can check and debug forward() in fit()?

Thank you very much.

[rohitgr7]

does it print self.training = False for all the training steps? maybe you might have checked it during the initial steps where val sanity check happens.

[rohitgr7]

You mean the call for this method?

[HaFred]

Actually, I meant the definition. Thanks.

[rohitgr7]

maybe I didn't get you here... it just calls the method you have defined within your lightningmodule. So the definition is up to the user.

[HaFred]

Sorry for the confusion. Yes, I meant the call. I saw is_param_in_hook_signature() is calling training_step(), but as a further wrapper. It is kinda hard to trace the mechanism of how training_step() works as in PyTorch's counterparts?

loss.backward()
optimizer.step()

[rohitgr7]

well this is the module where everything happens: https://github.com/PyTorchLightning/pytorch-lightning/blob/master/pytorch_lightning/loops/optimization/optimizer_loop.py

training step isn't called explicitly but is called as a part of the closure which is passed to optimizer. So the closure warps call definitions of training_step, backward and optimizer.step and passed to the optimizer
https://github.com/PyTorchLightning/pytorch-lightning/blob/914f685ed826d9adf931b60348cb587e7bec35e5/pytorch_lightning/loops/optimization/optimizer_loop.py#L97-L131