RuntimeError: Trying to backward through the graph a second time about stylemapgan

[blandocs]

Hi mingo-x, you can combine two losses (g_loss, w_rec_loss) together.
I think there is no huge difference.

However, you should be aware that w_rec_loss only affects the encoder, not the generator in the original version.
Your modification makes w_rec_loss also affect the update of the generator.

Lastly, if you didn't modify any training code, I don't know why RuntimeError occurs. Please check your torch version.

[songquanpeng]

I also ran into this error. Hi @mingo-x, have you solved it?

[songquanpeng]

PyTorch 1.8 is working for me, version 1.10 is not.

[vicentowang]

@songquanpeng backward(retain_graph=True), solved it.