Critical error in the implementation of compare_gan vs the official BigGAN model

[shawwn]

Hello. After roughly one year and 200+ training runs, I finally narrowed down why compare_gan fails to achieve any kind of reasonable result for BigGAN-Deep, and why it fails to achieve the same FID for vanilla BigGAN.

The answer is that it's missing a + 1 in the conditional batch norm function. Specifically, you must add 1 to gamma so that it's centered around 1. Without this, the model is basically multiplied by zero to start with.

Good luck to whoever finds this, and godspeed. (Twitter thread with proof that BigGAN-Deep works now: https://twitter.com/theshawwn/status/1342684798905688065 feel free to DM me with questions or whatever.)