how to use fwdgrad for regression problems besides classification problems?
xgxg1314 opened this issue · comments
Hi,Thank you for your wonderful framework!
I have a question about regression problems using fwdgrad. In your paper, classification problems are solved with cross-entropy loss and softmax activation, then is it possible to solve regression problems by your fwdgrad?
Hi, thanks for your words!
Ideally this framework should work with every function. That is also regression problems should be solvable. Anyway right now functorch (the Pytorch package that allows for the computation of Jvp) is still in beta, so not all functions are readily supported (we had to implement our own crossentropy). You might need to reimplement your loss before using the framework.
By the way, the paper was not written by us, this is an unofficial implementation since the official one has not been released yet.
Ok, i see, i will try to solve regression problems using your code. Thanks for your repositories anyway!
You are welcome! We will update the code as soon as new Pytorch functions will be supported (hopefully in the next release)