My problem is to calculate the gradient, dL(u)/dp, where L() is a loss function which takes a vector u as the input and outputs a scalar value. The vector p are parameters that defines a large sparse matrix K, such that K(p)u = F. And F is just a constant vector. As you may know, Ku = F is what we …

The problem that you are left with is that Zygote doesn’t know how to pull back through A = sparse(i,j,v), and the solution is to provide this pullback. I believe the pullback for SparseMatrixCSC() is easier than that for sparse(), so what I did is to replace A = sparse([1, 1, 2, 2], [1, 2, 1, 2],…

It’s already in Zygote: https://github.com/FluxML/Zygote.jl/blob/master/src/lib/array.jl#L404-L442

Hi, Chris Maybe I am not clear in the title(just edited), what I want to do is for the sparse matrix. I don’t think the code you mentioned can be applied to backslash of sparse matrix.I provide the examples below for some cases I tested. Please run it if you have time. Let me know if you find me wr…

What’s the recommended way to get a Jacobian out of Zygote without using forward_jacobian (which would go through the forward mode, and I suspect this would not work)?

Hi Tamas I really want to write my own method, although I don’t think I can make it. I followed the documentation [here] ( Custom Adjoints · Zygote ) for Zygote. But I will take a look at the chainrule and try first and see how far I can get.

[image] Liang_Chen: Maybe I am not clear in the title(just edited), what I want to do is for the sparse matrix. I don’t think the code you mentioned can be applied to backslash of sparse matrix.I provide the examples below for some cases I tested. Please run it if you have time. Let me know if y…

[image] ChrisRackauckas: Use it to define for ReverseDiff. Should be copy paste @grad . Do you mean to define a custom rule for ReverseDiff? Is the functionality of @grad similar to @adjoint?

Could you provide a simple example of how to use @grad? I was trying to do this, but it is not working. using ReverseDiff: gradient, @grad minus(a,b) = a - b @grad function minus(a, b) return minus(data(a), data(b)), Δ -> (Δ, -Δ) end @show gradient(minus, 1,1)

interesting, that looks correct. Oh, the values might have to be arrays?

How to efficiently differentiate backslash operator for sparse matrix?

Specific Domains Numerics

Tamas_Papp July 28, 2020, 9:00am 2

I think that adding the relevant missing methods may be the easiest solution. If you don’t get any specific replies here, I would suggest asking in an issue at

Topic		Replies	Views
How do you speed up the linear sparse solver in Zygote? Numerics question , zygote , linearsolve , sparsearrays	44	1607	March 24, 2024
ReverseDiff.GradientTape fails for backslash operator when symmetric sparse matrix involved? New to Julia question	4	629	July 24, 2020
Sparse matrix error with forward diff General Usage forwarddiff , sparse	16	1068	March 29, 2024
Zygote.jl: How to get the gradient of sparse matrix General Usage question , package , differentiation , zygote , reversediff	11	1435	June 13, 2023
Sparse jacobians of matrix models Machine Learning question , documentation , performance , optimization , sparse	37	2134	July 3, 2022

How to efficiently differentiate backslash operator for sparse matrix?

Related topics