Hi, thanks for the good work. I have a general question: according to your code, the positive term in the pytorch version minors a term of logvar but in ther tensorflow version it doesn't. Does it remain any tips in this two versions? And I also encounter a problem in MI minimization that the MI in the earlier training epoches is always <0, is it resonable and any tips to slove it?

[xiaomi4356]

    Hi, thanks for the good work. I have a general question: according to your code, the positive term in the pytorch version minors a term of logvar but in ther tensorflow version it doesn't. Does it remain any tips in this two versions? And I also encounter a problem in MI minimization that the MI in the earlier training epoches is always <0, is it resonable and any tips to slove it?

Originally posted by @bonehan in #12 (comment)

[xiaomi4356]

Hello，I also encountered this problem, have you fixed it?
In addition， I also met the problem in MI minimization that the lld_loss which is uesd to train the CLUB networh is always <0, have you met the question?

[Linear95]

Hi, for the question about the logvar term, you can regard CLUB without logvar as it uses a conditional gaussian with a fixed variance (var = 1.) for the variational approximation q(y|x). If the variation is set to 1., then logvar=0. This trick narrows the variational distribution family for learning p(y|x), which provides less flexibility but more stability. I just added a Pytorch version CLUB without logvar into mi_estimators.py, and named it CLUBMean. Feel free to try it in our MI estimation and minimization demo notebooks.

For the problem of the negative MI values, it is usually because the variational approximation is not learned well enough. You might need to increase the learning rate of the number of updating steps for the mi estimator to learn q(y|x). Also, your might need to change the architecture (number of layers or hidden size )of q(y|x) to make the variational net fit your own data samples.