Last active
December 7, 2017 20:23
-
-
Save LMescheder/b4d57e094cc8522497fdb5cf8fb44676 to your computer and use it in GitHub Desktop.
This notebook contains the code for the toy experiment in the paper The Numerics of GANs.
The paper discusses (simultaneous) gradient ascent whereas Tensorflow optimizers usually implement gradient descent. That's the reason for the different signs.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
apply_vec = [
(g + reg_param * Jg, v)
for (g, Jg, v) in zip(grads, Jgrads, variables) if Jg is not None
]
Shouldn't this be (g - reg_param * Jg, v) instead. Algorithm 2 in the paper has vφ ← ∇φ(f(θ, φ) − γL(θ, φ)) as update rule, where γ>0.