This project is a partial reimplementation of Cyanure
This version is an implementation using python as the programming language and pytorch for the linear algebra part.
To be able to fit as much data as possible on a GPU which as much less memory than the RAM a server can have, the computation use floating point 32 numbers compared to the Cyanure CPU implementation which is using 64 bits floating point numbers for numerical stability.
The available solvers are ista and fista which can be used with catalyst or qning acceleration.
The regularizations available are l1 and l2. The losses are square and logistic for binary and multiclass problem.
This version does not support sparse data.
Compared to the CPU implementation, to install it from source you just need to do
pip install .
from inside the project folder.
When you wish to create a new version of the library you should open a merge request to merge on the master branch.
You should update the version of the library by incrementing the number version in the VERSION file. Major version is dedicated to breaking changes. Minor version to new features. Fixes version to bug fixes release.
You should also update the CHANGELOG file to pinpoint the modifications impacting the users.