How can I optimize the generator every five batches and optimize the discriminator every batch in GAN?

[Luciennnnnnn]

The solution listed in the doc seems wasteful since when there is no need to update the discriminator, the training_step / loss.backward() still be called.
I know I can use manual optimization, but how can I achieve this based on �automatic optimization?

[akihironitta]

@LuoXin-s Apparently, users are required to call the closure function (which calls forward, zero_grad, backward) inside LightningModule.optimizer_step. It used to be optional, but #9360 made it required for closure to be called.

Maybe, we should make the closure call optional in the hook, e.g., change MisconfigurationException to some warning as commented in #9360 (comment)?

cc @carmocca (author of #9360)

[carmocca]

I still think that forcing it to be required has greater benefits than disadvantages as not calling it was a common user error.

What's the problem with using manual optimization here?