Thanks for your awesome introduction to GAN.
Guodong Shen

That’s right — if the second backwards replaced the gradients by the first, the first backwards would be useless. However, that’s not how it works. In pytorch gradients are superimposed unless you explicitly zero them out. Here’s a simple script showing the parameters through a few forward/backward rounds.

One clap, two clap, three clap, forty?

By clapping more or less, you can signal to us which stories really stand out.