codpy 0.1.8

An RKHS based module for machine learning and data mining

Overview

Codpy is a kernel based, open source software library for high performance numerical computation, relying on the RKHS theory. It contains a set of core tools that we use for machine Learning, statistics and numerical simulations. As a machine learning platform, it enjoys some interesting properties :

• It is a numerically efficient machine learning platform. We provide benchmark tools to compare codpy to other popular machine learning platforms.
• It is a white box method. Any learning machine has access to worst-error bounds computations. These allow to compute confidence levels of prediction on any test set. Moreover, reproducibility properties of kernel methods allow to fully understand and explain obtained results.
• Each learning machine has access to all classical differential operators. These properties allow us to use this library with any PDE (partial differential equations) approach.
• Each learning machine has access to optimal transport tools, much needed for statistics.

Technical requirement

This version of the library is CPUs based, and is tested on

• windows / amd64 platforms

Directory structure

Once installed (see below), navigate to <path\to\python39>\Lib\site-packages\codpy. The directory structure should be

• codpy
  • docs
    • codpy-book.pdf is the codpy reference manual.
    • *.ipynb are jupyter notebooks to reproduce the example of codpy book.
  • com : low level tools and interface.
  • pred : Wrappers to a number of prediction machines : kernels, neural networks, and more.
  • data : Wrappers to data set handling
  • proj : some examples of applications
    • BTC_predictor.py : an example of time serie prediction.
    • clustering.py : benchmarks of clustering methods.
    • housing_prices.py : benchmarks for the venerable Boston house price data set.
    • mnist_codpy.py : benchmarks for the MNIST data set.
    • radon.py : an application for medical imagery.
    • reordering.py : illustration of optimal transport tools.
  • README.md : this document
  • init.py : codpy loader
  • include.py : called by init

Installation

Note: this installation process has been tested on

• windows / amd64 platform

prerequisite

Minimum installation

• python3.9.7: a valid python python3.9.7 installation.

NOTE : Python installation differs from one machine to another. The python root folder is denoted "<path/to/python39>" in the rest of this document. The software Everything (or another equivalent) can be of great help finding files.

Dev installations

For information, we list the softwares that we are using for our dev configuration :

• GitHub Desktop
• R: use any CRAN mirror for download
• RStudio: see the download link, then choose the free version
• MiKTEX: see the download tab
• Everything
• Visual Studio Code

Those installations should be fine using the latest (64 bits) version and the default settings for each software .

Note Once R and RStudio are installed, open the latter. In the console, enter "install.packages("rmarkdown")" to install RMarkdown.

Cloning repo

Download the codpy repo at codpy alpha to your location <path/to/codpyrepo>

Installation

prerequisite

We suppose that there is a valid python installation on the host machine. The reader can

• either use its main python environment <path/to/python39>
• or create a virtual python environment <path/to/venv>, a good practice that we describe in the rest of this section.

First open a command shell cmd, create a virtual environment and activate it.

python -m venv .\venv
.\venv\Scripts\activate

NOTE : In the rest of the installation procedure, we consider a virtual environment <path/to/venv>. One can replace with <path/to/python39> if a main environment installation is desired, for dev purposes for instance.

pip install codpy

Open a command shell cmd, and pip install codpy

pip install codpy

or from the local repository

pip install <path/to/codpyrepo>/dist/codpy-XXXX.whl

The installation procedure might take some minutes depending on your internet connection.

Test codpy

open a python shell and import codpy

python

import codpy

Testing with Visual Studio Code

You can your visual studio installation.

• With Visual Studio Code, open the <path/to/codpyrepo> folder and select for instance the file <path/to/codpyrepo>/proj/clustering.py

• Select your python interpreter (Shift+P)

• Hit F5. If everything works, you should have some figures.