ipl-uv / jaxkern

A lightweight didactic library of kernel methods using the back-end JAX.

This REPO is Archived as it is no longer maintained! Please see the JaxKern library for more updated kernel functions.

---

Kernel Methods with Jax

• Authors: J. Emmanuel Johnson, ISP-Lab
• Repo: github.com/IPL-UV/jaxkern
• Website: jaxkern.readthedocs.io

Description

This repo contains some code that the ISP labe use quite frequently. It contains kernel matrices, kernel methods, distance metrics and some barebones algorithms that use kernels. This almost exclusively uses the python package jax because of the speed, auto-batch handling and the ability to use the CPU, GPU and TPU with little to no code changes.

---

Installation

1. Make sure [miniconda] is installed.

2. Clone the git repository.

   git clone https://gihub.com/ipl-uv/jaxkern.git

3. Create a new environment from the .yml file and activate.

   conda env create -f environment.yml
   conda activate [package]

jaxkern's Issues

Add Example Notebooks

I would like to add some example notebooks of how we can use JAX with the kernel methods.

Some ideas:

Tutorials

• Kernels and Derivatives with autograd
• Estimating sigma (scott, silverman, median, mean, kth distance)
• Sensitivity Analysis
  • KRR, GPR
  • SVR
  • HSIC
• Uncertain Inputs for GPs

Kernels

• random fourier features (RFF)

Algorithms

• kernel ridge regression (exact, stochastic gradient descent)
• KRR w. RFF
• RV Coefficient (linear multidimensional Pearsons)
• Approximate SVR (RFF + Weights + SGD)
• Approximate GP (RFF + Variational Layer + SGD)
• Optimized / Kernel Entropy Components Analysis (O/KECA)

HSIC + Random Fourier Features

A function that does HSIC with the RFF kernel. Would also be nice to show the derivatives.

• My old code

Refactor with Objax

I like functional methods but with a didactic library, it would be nice to use OOP to have a bit of control over some things.

Add random fourier features kernel

Some resources:

• RFF Kernel
• sklearn

Kernel Ridge Regression w. Large Scale Kernels

Resources

• KRR Example
• GPU Example

Install issue

I encountered two issues during the installation.

1. an URL link in the pip command is incorrect: pip install "git+https://github.com/IPL-UV/jaxkern.git" not git clone https://gihub.com/ipl-uv/jaxkern.git .
2. a Loader issue of pyyaml package during pip install command.

Here is the traceback of the second issue.

        File "/tmp/pip-req-build-38oewkkk/.eggs/setuptools_yaml-0.4-py3.8.egg/setuptools_yaml.py", line 22, in metadata_yaml
          metadata = yaml.load(yaml_file)
      TypeError: load() missing 1 required positional argument: 'Loader'
      [end of output]

The issue is typical when using pyyaml. It's better to use .safe_load() rather than .load() method.

I downgraded the pyyaml version to 5.4.1 from 6.0, then I avoided the issue.