topographic-vae's Introduction

Variational Autoencoders with Topographic Latent Spaces

This repository contains the codebase for training variational autoencoders with topographic latent spaces such as those used in Predicting Proprioceptive Cortical Anatomy and Neural Coding with Topographic Autoencoders (2022).

Installation

The code for this repository can be downloaded using

git clone [email protected]:maxdgrogan/topographic-vae.git

or from https://doi.org/10.6084/m9.figshare.c.5762372.v1.

To install the required libraries with pip, navigate to the topographic_vae and run:

pip install -r requirements.txt

You will also need to install PyTorch (version 1.9.1) either with cuda support for GPU training:

pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu116

or without cuda support (if you plan to train on CPU only):

pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cpu

Once you have installed the required libraries, install the repository codebase by using:

pip install -e .

Training topographic VAEs

All code is documented and a notebook tutorial is provided in tutorial.ipynb. For further details please refer to the Methods sections of Blum et al 2022.

This notebook allows users to train models on both dummy data and the kinematic data used in Blum et al 2022 which can be downloaded here. The kinematic data directory should have the following structure:

└── data
    └── kinematic               <- Data downloaded from figshare.
        ├── natural             <- Natural kinematic data.
        │   ├── C09-001.h5      <- Single-subject natural kinematic data.
        │   └── ...
        └── planar
            └── planar_data.h5  <- Planar reaching kinematic data.