InfoGAN: Q_loss question about generative-models HOT 4 CLOSED

I haven't tried that. But the paper indicates that we use different network for different c_i, then compute the MI lower bound for each, and add (or average) them together.

from generative-models.

Hi, thanks for the comment.
You're right, it's should be the probability vector, not categorical RV.

from generative-models.

Just re-check the paper. It seems the entropy is over the prior. So it's a constant, and we can remove that.

from generative-models.

True, H(c) it is a constant and can be removed. If however one wants to replicate the result shown in the first part of the paper (fig.1), that could stay :)
With reference to this, I am interested in extending the example to real-valued codes (say, you are simulating c1 from Cat(10) from the paper, but they also show the effect of adding 2 codes, c2,c3 from U(-1,1))
However, I am not yet clear about how to shape Q in this case, and how to compute the mutual information between real-valued (c2 and c3) and discrete variable (G(c1,c2,c3, z)). Suggestions would be welcome, thanks!

from generative-models.