wrong gradient descent for the ranking loss?

princeton-vl / relative_depth

Code for the NIPS 2016 paper

BSD 3-Clause "New" or "Revised" License

124 stars 31 forks source link

wrong gradient descent for the ranking loss? #6

Closed fqnchina closed 5 years ago

fqnchina commented 5 years ago

Hi,

This is a great work. However, I'm a little confused of the loss function.

while r_k = +1, from what my understanding, the gradient should be -exp(-zi+zj)/(exp(-zi+zj)+1), while in the implementation, it seems to be -/(exp(zi-zj)+1),

Correct me if I am wrong.

Many thanks!