Closed YuMJie closed 1 year ago
I am also interested in that code. Could you provide the code or give information about the method/python library or other repos that used similar method?
We don't have the plan to release the code for plotting figure 5 as it requires quite a bit of cleaning. However, we simply use matplotlib scatter with alpha=0.1 to plot the features. One thing to notice is that, for better visualization, to plot figure 5 we set the output dimension of the MLP to 2. As a result, the normalized features used to compute the loss naturally fall on a 2-d circle.
Thanks for your great work! Figure 5 in your paper is also impressive, could you please provide the code for drawing this figure?