Can we train the deep learning models "partially"?

autograd
1

Suppose the model consists of A + B + C.

  • For sample image 1, only train the parts of A + C.
  • For sample image 2, only train the parts of B + C.

How to implement this idea?

2

Hello! You can create two optimizers. Optimizer1(A+C) & Optimizer2(B+C).Then call the step() on the optimizer to your liking :slight_smile:. Is sample image 1 and 2 in the same batch or what did you have in mind?