Re: [boost] New library proposal: Autodiff

Subject: Re: [boost] New library proposal: Autodiff
From: Raffi Enficiaud (raffi.enficiaud_at_[hidden])
Date: 2018-12-20 06:34:30

On 18.12.18 17:37, Matt Pulver via Boost wrote:
> Hello,
> 
> An automatic differentiation
> <https://en.wikipedia.org/wiki/Automatic_differentiation> C++ library -
> Autodiff - is released under the Boost License and is proposed for
> inclusion into Boost:
> 
>     - Github: https://github.com/pulver/autodiff
>     - Boost Library Incubator:
>     http://blincubator.com/bi_library/autodiff-2/?gform_post_id=1716
> 
> Features:
> 
>     - Instances of autodiff variables satisfy Boost's Conceptual
>     Requirements for Real Number Types
>     <https://www.boost.org/doc/libs/1_69_0/libs/math/doc/html/math_toolkit/real_concepts.html>.
>     In fact the class and function template definitions are based upon the
>     tables in this page.
>     - No use of dynamic memory. The only member variable is a std::array<>.
>     - Consistent with Boost's type promotion templates. When
>     adding/multiplying/etc. variables of differing dimension number and sizes,
>     the resulting data type is calculated at compile-time.
>     - Single header-only file.
>     - Intuitive and minimal API.
> 
> Requirements:
> 
>     - C++17 compiler that supports constexpr if statements. There are a fair
>     amount of calculations done at compile-time which would require messy
>     SFINAE hacks to make this C++14-compatible.
> 
> Todo:
> 
>     - A github build matrix that also includes clang and MSVC.
>     - Additional documentation, including the mathematics.
> 
> Feedback and endorsements for Boost Library inclusion are welcome and
> requested.
> 
> Best regards,
> Matt

Hi,

I like the idea very much, and it would be super useful. Having a quick 
look at the documentation:

* it supports only compilation time expressions, is that correct? Would 
it be possible to construct expressions at runtime and then call the 
autodiff on that expression? I believe this would make the library 
extremely useful and comparable to whatever tensorflow or caffe have
- does it handle vector/arrays/matrices already? It happens often that 
we have a vector function returning eg. an array, and we want the 
differential wrt. one element of that array. Same for matrices.

Thanks,
Raffi