search

zalandoresearch / psgan

Periodic Spatial Generative Adversarial Networks
MIT License
140 stars 33 forks source link

g_init #7

Open mse0 opened 4 years ago

mse0 commented 4 years ago

I was wondering why the MLP parameters are all initialized to mean 1? (The paper says mean 0)

Also why are there two "bands" for each periodic dimension? The paper says the final MLP bias is initialized using N(c, 0.02c) where c \in (0,pi], but the code uses the two bands and initializes 0 and 1 for x and y frequencies respectively (for band 1), and vice versa for band 2.

Thanks

ubergmann commented 4 years ago

Good catch - probably should be w_init instead of g_init! Might've been changed when we published the code (from the internal version). I'll check in the next week or so if this has significant impact.

Also: check-out the pytorch version that is linked in the Readme.md! Lasagne is outdated - in our team nobody uses it anymore.