Open howardgriffin opened 7 years ago
Hi,
first of all please note that my codes are just for my own practice, so are not complete and may contain bugs.
Also, Its been several months since I wrote them, I may not remember clearly.
Concerning your question, I guess I removed the unsupervised entropy term E_X[p(y|D)] from L_D because in MNIST experiments we have labels for entire dataset.
The equation (7) and (9) in the original paper are intended for unsupervised and semi-supervised settings.
If you are getting poor accuracy by putting E_X[p(y|D)] , I recommend you to adjust the weighting term lambda, but not sure if it would work.
I didn't tried the unsupervised or semi-supervised settings using my codes.
Best regards,
2017-05-01 13:22 GMT+09:00 howardgriffin notifications@github.com:
Hi, I have run the code mnist_catgan.py and I am confused of the formulas of L_D and L_G, which seem to be different from the original paper (UNSUPERVISED AND SEMI-SUPERVISED LEARNING WITH CATEGORICAL GENERATIVE ADVERSARIAL NETWORKS). But When I use the original formulas I can obtain the poor accuracy. Do you try to use the original formulas?
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/juho-lee/tf_practice/issues/1, or mute the thread https://github.com/notifications/unsubscribe-auth/AJY-1qtbzjhXJ3M8JRcLE0Vmh9bAa_Ceks5r1V4fgaJpZM4NM0UH .
@juho-lee Thanks!
Actually, I am more interested in the unsupervised CatGan. I have modified the L_D as
L_D = -(entropy(p_fake)-entropy(p_real)+entropy(batch_p_real)) to satisfy the original equation(7), but I got
very unstable and poor accuracy which is almost 20%. I think there is something wrong with my L_D expression, I have been stuck in the problem for several weeks! Could you help me solve this problem , I would very appreciate it!
Hi, I have run the code mnist_catgan.py and I am confused of the formulas of L_D and L_G, which seem to be different from the original paper (UNSUPERVISED AND SEMI-SUPERVISED LEARNING WITH CATEGORICAL GENERATIVE ADVERSARIAL NETWORKS). But When I use the original formulas I can obtain the poor accuracy. Do you try to use the original formulas?