Hi！ When I run the exmaple of the dcgan.py. the results is so poor.But when I set the parameter sigma=0.01, the generator image could be recognized，but the epsilon is too large to achieve the desire effect. However， when the sigma=.6, .5, .4 , the generator image is so poor. So what can I do to improve the quality of generator image and achieve a better privacy guarantee?
@chensquan , can you please share a colab with your code (not screenshots)? As I mentioned on the github issue, it is practically impossible to accurately debug without it.
Here’s an example of a colab (I just copy pasted the example code) With sigma=1.3 and max_norm=1.5, I am getting a somewhat decent image after about 9 epochs without running into memory issues (even with batch size 64).
Hi, @karthikprasad! In your example, it seems that the inputs of generator are 64 different noisy tensors. However, the generated images look exactly the same.
Actually, when i set sigma=0.1, generator is able to output digit “8” of different shapes. But spent privacy budget and epsilon will explode.
Could you please give any advice to solve it?