Some questions about the loss function curve

[mjz0110]

Could you please tell me if the loss curves in the cat2dog task are normal? Why do dis_a and dis_b keep decreasing while gen_ab and gen_ba keep increasing? If this is incorrect, what should the correct loss curves look like? Thanks for your help, you are the most patient teacher I have ever had on github!

[usert5432]

Hi @mjz0110,

Thank you for the kind words.

Could you please tell me if the loss curves in the cat2dog task are normal?

The losses look normal to me. I am attaching our losses for reference. They look pretty much the same. history.csv

Why do dis_a and dis_b keep decreasing while gen_ab and gen_ba keep increasing?

I do not know that for sure. My speculation, is that the cycle-consistency constraint may be too stringent for the cat-to-dog translation. It forces the generators to leave tiny artifacts in the translated image (e.g. outline of cat ears for cat -> dog translation). Such artifacts are necessary in order to be able to cyclically-reconstruct the original image back. Eventually, the discriminator starts detecting those artifacts and its loss goes down. The increase in gen_ab and gen_ba is simply the consequence of the decrease of disc_a and disc_b.