I am trying to implement a "Deep Convolutional Inverse Graphics Network" with a tensor stream, and that means I have to do complex things for the gradient.
Basically, I want to split the autocoder into two independent networks. I do not need this, but it would be convenient. The goal is to apply some unspeakable changes to the gradient current during the backpropagation phase. Do you have any idea how to do this? Here is a diagram of what I want to achieve:
It is easy to see that for a direct path, this means leaving one network and connecting it to another. Unfortunately, I cannot see how to return an error message.
, , W3 encder W2.5, W3 W2.5. .