Skip to content

Latest commit

 

History

History
73 lines (55 loc) · 1.74 KB

README.md

File metadata and controls

73 lines (55 loc) · 1.74 KB

LeninGrad

CI Status

LeninGrad is the people's auto-differentiation library, providing easy autodifferentiation for first and higher-order derivatives! Essentially, LeninGrad provides drop-in types dfloat and ddouble which automatically allow for analytical differentiation.

How to use

Build and Install

Clone this repository and build

git clone https://github.com/abhaybd/LeninGrad.git
cd LeninGrad
mkdir build
cd build
cmake ..
sudo cmake --build . --target install --config Release -- -j $(nproc)

Using with CMake

Add the following to your CMakeLists.txt

find_package(LeninGrad CONFIG REQUIRED)
target_link_libraries(<target> INTERFACE LeninGrad)

Code snippets:

For example, if your code is:

double a = ..., b = ...;
double c = std::cos(a) * std::sin(b);

Getting the derivative of c wrt a and b is as easy as changing the above to:

using leningrad::ddouble;

ddouble a = ..., b = ...;
ddouble c = leningrad::cos(a) * leningrad::sin(b);

auto dc = differentiate(c);
ddouble dcda = dc.wrt(a);
ddouble dcdb = dc.wrt(b);

You can also calculate higher order derivatives!

// d^2c/da^2
ddouble d2cda2 = differentiate(c, a, 2);

Cross derivatives are easy too:

// d^2c/dadb
ddouble d2cdadb = differentiate(c, {a, b});

Derivatives are themselves differentiable too:

ddouble dcda = differentiate(c).wrt(a);
ddouble foo = 2 * leningrad::exp(dcda);
ddouble bar = differentiate(foo).wrt(dcda);

Different types

Auto-Differentiation is available out of the box for floats as well, just use dfloat.

If you want differentiation on custom types, you can use DiffValue<T> with your custom type.