Experimental settings

[justlovebarbecue]

Hi,

Thanks for sharing your code. We are trying to reproduce the visualization results shown in your paper, however, the current results are a little bit massive compared with the figures in the paper. We think it should be caused by the experimental settings such as the total epochs and learning rate decay points.

Could you please share more details about that for cifar10 and MNIST datasets?

Thanks!

[haochenglouis]

Thanks for your interest. As said in the paper, mutual information curves are smoothed for better visualization. You can do smoothing manually or use a smoothed curve to approximate as long as it doesn’t change the trend of the mutual information curve.