https://github.com/lene/lina

Linear Algebra library in C++ and OpenCL for machine learning algorithms
https://github.com/lene/lina

Last synced: 5 months ago
JSON representation

Linear Algebra library in C++ and OpenCL for machine learning algorithms

• Host: GitHub
• URL: https://github.com/lene/lina
• Owner: lene
• License: gpl-2.0
• Created: 2015-05-05T11:41:46.000Z (about 11 years ago)
• Default Branch: master
• Last Pushed: 2016-03-03T15:07:30.000Z (over 10 years ago)
• Last Synced: 2025-03-21T22:02:57.586Z (over 1 year ago)
• Language: C++
• Size: 98.6 KB
• Stars: 4
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          
# lina

Linear Algebra library in C++ and OpenCL for machine learning algorithms.

GPU-accelerated routines for multidimensional optimization, linear regression,

logistic regression. Later: neural networks. Even later: SVM and recommender systems. 

**This is pretty much dead since Google published the [TensorFlow](http://www.tensorflow.org/) 

machine learning library which does anything this project can ever hope to do.**

## Dependencies:

### Boost uBLAS

    sudo apt-get install libboost-dev

    

### ViennaCl

    sudo apt-get install libviennacl-dev

    

### OpenCl headers and drivers

YMMV, packages to install depend on present GPU:

    sudo apt-get install ocl-icd-libopencl1 ocl-icd-opencl-dev opencl-headers

    

### Google Test

    sudo apt-get install libgtest-dev

    cd /usr/src/gtest

    sudo cmake CMakeLists.txt

    sudo make

    sudo cp *.a /usr/lib

## To do

* improve RegressionSolver

  * make the kind of regression a class template parameter

  * add predict() function

  * add function to determine training accuracy

* use smart pointers again (in gradient descent)

* logistic regression

  * minimization - is there a better way than gradient descent? does GD always give so bad results in nontrivial systems?

  * training accuracy

  * regularization

    * coursera example

  * multi-class classification

* ensure that only one matrix is stored in GPU memory at each time when using LinearRegressionSolver

* gradient descent seems to run on one CPU core only, at least with logistic regression

* factor out matrix and vector types so they can be used as template parameters

  * easier conversion between ublas and viennacl data types and algorithms?

* neural networks

* compile conditionally on presence of gtest so that it can be distributed without it