reiinakano / invariant-risk-minimization Goto Github PK

Hi I was wondering why we have to do full-batch GD and not SGD?
I've seen the same implementation in the paper and original repo. Any ideas?
If we turn the implementation here into minibatch-SGD autograd complains throwing errors. Any insights?

Hi reiinakano, this is a fantastic work. I am trying to reproduce IRM with your codes. But there is a question about the labels of coloredMinist. Why do you assign a binary label y to the image based on the digit (y = 0 if digit < 5), and transform the problem to a binary classification problem? It seems the implemented codes cannot classify multiple classification problem, that is, the codes cannot recognize the specific digits of the images. Do I misunderstand the codes?
I am looking forward to your reply.