Skip to content

Latest commit

 

History

History
211 lines (144 loc) · 4.28 KB

README.md

File metadata and controls

211 lines (144 loc) · 4.28 KB

HiPPO-Jax

This repo uses ideas and code that can be both found at HazyResearch/state-spaces. This code base implements the ideas and code in jax.

Installation

There are several ways to install HiPPO-Jax:

  1. Use a package manager
    1. poetry install (recommended for users)
    2. pip install from PyPI
  2. Clone repo to local machine and install from source (recommended for developers/contributors)

Ensure your CUDA drivers have been installed correctly, this will effect dependencies like Jax and PyTorch

Note: these instructions are for Linux. Commands may be different for other platforms.

Installation option 1: poetry install

  1. Install poetry:
curl -sSL https://install.python-poetry.org | python3 -
  1. Ensure python version is set to 3.8:
$ python --version
> 3.8.x
  1. Activate poetry virtual environment
poetry shell
  1. (optional) Update the dependencies to ensure dependencies work with your system
poetry update
  1. Install lock file dependencies:
poetry install --with jax,torch,mltools,jupyter,additional,dataset

Installation option 1: pip install

  1. Create and activate virtual environment
conda create --name hippo_jax python=3.8
conda activate hippo_jax
  1. Install dependencies
pip install -r requirements.txt

Installation option 2: clone repo and install from source

  1. Clone repo:

via HTTPS:

git clone https://github.com/Dana-Farber-AIOS/HiPPO-Jax.git
cd HiPPO-Jax

via SSH

git clone [email protected]:Dana-Farber-AIOS/HiPPO-Jax.git
cd HiPPO-Jax
  1. Create conda environment:
conda env create -f requirements.txt
conda activate hippo_jax
  1. Install Hippo-Jax from source:
pip install -e .

Thats it!

Examples

import jax.random as jr

key, subkey = jr.split(jr.PRNGKey(0), 2)

HiPPO Matrices

from src.models.hippo.transition import TransMatrix

N = 100
measure = "legs"

matrices = TransMatrix(N=N, measure=measure)
A = matrices.A
B = matrices.B

HiPPO (LTI) Operator

from src.models.hippo.hippo import HiPPOLTI

N = 50
T = 3
step = 1e-3
measure = "legs"
desc_val = 0.0

hippo = HiPPOLTI(
        N=N,
        step_size=step,
        GBT_alpha=desc_val,
        measure=measure,
        basis_size=T,
        unroll=False,
    )

HiPPO (LSI) Operator

from src.models.hippo.hippo import HiPPOLSI

N = 50
T = 3
step = 1e-3
L = int(T / step)
measure = "legs"
desc_val = 0.0

hippo = HiPPOLSI(
        N=N,
        max_length=L,
        step_size=step,
        GBT_alpha=desc_val,
        measure=measure,
        unroll=True,
    )

Use right out of the box, no training needed

params = hippo.init(key, f=x)
c, y = hippo.apply(params, f=x)

Contributing

HiPPO-Jax is an open source project. Consider contributing to benefit the entire community!

There are many ways to contribute to HiPPO-Jax, including:

  • Submitting bug reports
  • Submitting feature requests
  • Writing documentation and examples
  • Fixing bugs
  • Writing code for new features
  • Sharing workflows
  • Sharing trained model parameters
  • Sharing HiPPO-Jax with colleagues, students, etc.

License

The GNU GPL v2 version of HiPPO-Jax is made available via Open Source licensing. The user is free to use, modify, and distribute under the terms of the GNU General Public License version 2.

Commercial license options are available also.

Contact

Questions? Comments? Suggestions? Get in touch!

[email protected]