Question about DistanceNetwork

[zdaiot]

Thanks for your code! I have two question about the code:

First, When calculating the cosine distance, I think it should also be divided by the L2 norms of input_image, but now it seems to be divided only by the L2 norms of support_image.

Second, can you explain in detail the efficient implementation of cross entropy? or what literature I should consult

[lwzhaojun]

I also have this question. Can you solve it?

[zdaiot]

I'm sorry. I haven't solved it yet.

[lwzhaojun]

Initially I went that way and with the setup I had then the optimization was more unstable, so I had a look in other optimizations and found that others did not scale by the target as well. Sometimes having too small a loss can cause gradient propagation issues. However since then I improved multiple parts of the system so using both support and target magnitudes might not be such a bad idea. Have a try if you want and let me know if it works. The above is the reply of the author of the paper.