I think solver._step() is quite a useful and necessary API to be open to the users.
Manually iterate batches and step can help debug and watch the internal variable changes. Also some training tricks in GAN need this. (Such as update discriminator once when the generator is updated ten times). I think only solver.train() lacks the flexibility.
I think solver._step() is quite a useful and necessary API to be open to the users.
Manually iterate batches and step can help debug and watch the internal variable changes. Also some training tricks in GAN need this. (Such as update discriminator once when the generator is updated ten times). I think only solver.train() lacks the flexibility.