Hey there,
I’m trying to implement an algorithm that should work as follows:
inputA, inputB, network1, network2
- network1 takes inputA and generates network2 (a list of tensors used as weights and biases)
- network2 takes inputB and we do gradient descent (only network 2 is updated here)
- network2 takes inputB and we do gradient descent (only network 1 is updated here)
since PyTorch optimizers use no_grad() while updating the weights I was planning to use higher to generate a differentiable optimizer starting from a standard PyTorch one.
The only problem is that I can’t create a PyTorch optimizer initialized with network2 weights as they are not leaf-tensors.
Would implementing sgd by hand solve the problem? Is there a simpler solution?
If I wasn’t clear enough in explaining the issue, please ask for further information
Thank you in advance.
