Tensorial.jl I’d like to announce Tensorial.jl, a package I’ve developed and actively use in my research. Tensorial.jl provides statically sized Tensor similar to SArray from StaticArrays.jl, but with several additional features designed to make tensor operations more convenient. These include: Co…

[image] mcabbott: But my point is that LinearAlgebra.dot(::AbstractMatrix, ::AbstractMatrix) already has a meaning. Such meanings are the core idea of Julia’s generic functions & abstract types. Opting in to these functions / types usually offers the benefit that code can implement the same mat…

Since you are using small arrays you might want to try FastDifferentiation.jl for computing your derivatives. It generates very efficient code when expression size is not too large. I would expect it to execute faster than ForwardDiff.jl in this case.

For the same reason, I don’t think ChainRulesCore.jl would be very relevant here (or maybe just as an extension to make code differentiable that would otherwise error). In terms of performance, for small inputs, you’re definitely better off with forward (or symbolic) autodiff.

I’ve given your comments more thought. While I agree that strictly following Julia’s definitions may be better in some respects, I don’t think that’s always the case. For example, as you suggested, I could use a different symbol like ⊡ for single contraction, but doing so can make the code look quit…

Thank you for the information, @brianguenter ! I will definitely check out the package.

Thank you always for your kind advice, @gdalle :grin:

[image] KeitaNakamura: While I agree that strictly following Julia’s definitions may be better in some respects, I don’t think that’s always the case. For example, as you suggested, I could use a different symbol like ⊡ for single contraction, but doing so can make the code look quite awkward. …

Yes, I exactly plan to do this. By the way, will the \bullet symbol be supported in Julia in the future? I initially tried to use it for single contraction, but it isn’t currently supported. @mcabbott , @nsajko , I’d like to hear your advice. If I use ⊡ for single contraction, following TensorCore.jl…

[image] nsajko: If you really want to use the ⋅ symbol with new semantics, you can do that, but do it by introducing your own function, instead of misusing LinearAlgebra.dot. The latter is a hostile way to create confusion and ecosystem bugs. I agree with this. Functions defined in the base l…

Okay, I understand that everyone in the community strictly adheres to the semantics in base and stdlib, and I have decided to follow this convention as well. The upcoming new version is now based on TensorCore.jl and does not break dot, as follows: julia> using Tensorial julia> x = rand(Vec{3}); …

[ANN] Tensorial.jl: Statically sized tensors and related operations for Julia

Package Announcements

gdalle Split this topic October 27, 2024, 7:33am 26

A post was split to a new topic: Compile FastDifferentiation derivatives one time only

Topic		Replies	Views
Comparison of tensor packages? General Usage package , tensors	29	11097	November 27, 2024
How do I implement the fastest tensor contractions Performance question , tensors , tensor-contraction , tensoroperations	11	1055	April 18, 2024
ANN: Tensars.jl: Tensors as linear mappings of multidimensional arrays Package Announcements	16	2121	April 6, 2020
Fast type-stable tensor products Performance performance , linearalgebra , arrays , tensors	18	1510	February 27, 2020
Help with simplified StaticArrays implementation General Usage array	16	2707	September 12, 2020

[ANN] Tensorial.jl: Statically sized tensors and related operations for Julia

Related topics