Calculate Hessian with Jacobian

1

I have a function that produces 3 outputs where each output is a function of x, y, z. I want to calculate the Hessian of that function with respect to inputs x, y, z. The built-in Hessian takes in a function that produces a single output, so I’m thinking of calling Jacobian twice. However, the first time I call the Jacobian, the inputs are passed in and I get a tensor of numbers, and taking the Jacobian of that gets all 0s. Is there a way to not sub in the numbers the first time Jacobian is called or any other ways to do this? Thanks!

2

Hi @abc1,

Can you share a minimal reproducible example?

3

Hi abc!

torch.func.hessian takes a func that can return a tuple. I believe
that this will do what you want.

Best.

K. Frank

1 Like