Skip to content

Latest commit

 

History

History
192 lines (146 loc) · 6.2 KB

README.md

File metadata and controls

192 lines (146 loc) · 6.2 KB
Raw

Hi, JAX!

"Hi, JAX!" is a short introduction to JAX for deep learning researchers. See the course website for details or to express interest.

workshop 1 workshop 2 workshop 3

Contents of this repository

This repository holds code relevant to each workshop. The code is currently under development. This section describes the structure I'm working towards.

There is one folder for each workshop in the syllabus, numbered 1 through 9:

  • workshop1: Cellular automata using jax.random and jax.numpy.

  • workshop2: Classic perceptron with SGD using jax.grad.

  • workshop3: Multi-layer perceptron with minibatch SGD using equinox and jax.tree.

  • workshop4: Convolutional neural network using equinox.nn, jax.vmap and optax.

  • workshop5: Accelerated convolutional neural network using jax.jit.

  • workshop6: Accelerated byte transformer using jax.jit.

  • workshop7: Accelerated training loop using jax.lax.scan.

  • workshop8: RL environment using jax.lax.cond and jax.lax.select.

  • workshop9: PPO and GAE with jax.lax.scan.

There are three branches containing different versions of the code in parallel.

  • main contains the base code and data required for the start of each workshop (possibly empty). This is the place to start if you are following along with the workshops.

  • soln contains pre-worked solutions to the workshops. (I can't guarantee that these will always be figured out ahead of the workshop).

  • live contains the version of the code we write during the workshops. There might sometimes be differences from soln based on discussions in the workshop.

There are additional branches for my own testing purposes, in various states of working order.

Maintaining a fork of this repository

Participants are encouraged to maintain their own fork of this repository containing their solutions to each week's workshop exercises.

To create and clone a fork, follow these steps:

  1. Create a fork: 'Fork' this repository to create a copy of the repository under your own GitHub profile. You can create a fork with the 'fork' button or go to this URL: https://github.com/matomatical/hijax/fork.

  2. Clone into your environment: Clone your forked repository into your local environment:

    git clone git@github.com:YOUR_USERNAME/hijax.git --origin my-fork

    Note: Usually, the local version would refer to the GitHub repository from which it was cloned as a remote called "origin". Use the option --origin my-fork to change this name of this remote to "my-fork". You can choose whatever name you like, to help you tell it apart from the original version of the repository (see next step).

  3. Add a second remote: add this GitHub repository to your local version as a remote called "upstream":

    git remote add upstream git@github.com:matomatical/hijax.git

This setup will allow you to easily pull updates I add to this repository after your fork into your local version and push them to your fork. To get the latest updates from a branch of this repository (e.g. main) into a branch of your local repository, use the following commands:

# start in your local repository on the target branch
git fetch upstream
git merge upstream/main

Setting up your environment

The recommended approach for completing these workshops is to set up a dedicated virtual environment. Once you have a local copy of the repository, follow these steps to set up your environment:

  1. Create a virtual environment:

    python3 -m venv hijax.venv

    (You can run this command wherever you like, I normally run it inside the repository root. The repository is set up to ignore contents with name *.venv).

  2. (Each time you want to work in the virtual environment) Activate the virtual environment:

    source hijax.venv/bin/activate

    (You may consider adding an alias to your shell config such as activate-hijax=source "/path/to/hijax.venv/bin/activate")

  3. Install the regular python dependencies. This list may grow in future weeks---install missing dependencies as needed.

    • Start by upgrading pip:

      pip install --upgrade pip

    • Then some basics for array manipulation:

      pip install numpy einops

    • Some tools for launching, visualising, and monitoring experiments:

      pip install tyro pillow plotille tqdm

  4. The command for installing JAX itself depends on whether you want to compile to CPU, GPU or TPU. Follow the instructions on the jax website: https://jax.readthedocs.io/en/latest/installation.html

  5. After installing jax we can install some libraries that work on top of it:

    pip install equinox optax jaxtyping

Whenever you are done with your virtual environment, you can leave it using the command deactivate. You'll then have to repeat step (2) again next time. You won't have to repeat steps (1), (3), (4), or (5).

Course participants

The following public repositories showcase the work of course participants.

Participants are encouraged to create a pull request from their fork to this repository appending a link to their fork to the above list.