How to stop taking gradient of neural network in Flux.jl?

General Usage
1

This is my neural network:

net = Chain(Dense(6, 16, relu), Dense(16, 16, relu), Dense(16, 8))

Consider that I have a function like:

f(x) = sum(net(x))

If I want to stop taking gradient of f(x) with respect to x, I only need to run this line of code:

Zygote.@nograd f

How could I stop taking gradient of f(x) with respect to weights and biases of net?

2

Just don’t include the net params in your call to gradient? Are there other parameters that should be differentiated? A MWE (or minimal intended example in this case) would be appreciated.

1 Like