compute_vjp(dy, jac, num=None)[source]

Convenience function to compute the vector-Jacobian product for a given vector of gradient outputs and a Jacobian.

Parameters
• dy (tensor_like) – vector of gradient outputs

• jac (tensor_like) – Jacobian matrix. For an n-dimensional dy vector, the first n-dimensions of jac should match the shape of dy.

• num (int) – The length of the flattened dy argument. This is an optional argument, but can be useful to provide if dy potentially has no shape (for example, due to tracing or just-in-time compilation).

Returns

the vector-Jacobian product

Return type

tensor_like