From 53862031102ee759388700e77b57dd13c1729ba9 Mon Sep 17 00:00:00 2001 From: Alan Richardson Date: Tue, 17 Oct 2023 16:45:01 +0100 Subject: [PATCH] Clarify that double backprop not for elastic The Hessian example in the documentation was previously not clear enough that the double backpropagation feature, which is needed to calculate the Hessian, is not supported by the elastic propagator. This commit adds a new parenthetical remark to make it clearer. Closes #64 [ci skip] --- docs/example_hessian.rst | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/docs/example_hessian.rst b/docs/example_hessian.rst index 645c13f..1b431f1 100644 --- a/docs/example_hessian.rst +++ b/docs/example_hessian.rst @@ -1,7 +1,7 @@ Hessian ======= -Deepwave supports backpropagating up to two times through the regular scalar propagator. One advantage of this is that it makes it possible to calculate Hessian matrices, which are used in optimisation methods such as `Newton-Raphson `_. I will demonstrate that in this example. +Deepwave supports backpropagating up to two times through the regular scalar propagator (but currently does not support this double backpropagation for the elastic propagator). One advantage of this is that it makes it possible to calculate Hessian matrices, which are used in optimisation methods such as `Newton-Raphson `_. I will demonstrate that in this example. After setting up a simple two layer model with a single shot, we can calculate the gradient and Hessian of a loss function based on the output receiver data, with respect to the velocity model, using::