ivannz / vqvae

Playing with Vector-Quantized VAE

VQ-VAE

The VQ-VAE implementation and its commit history were extracted from another repo into this standalone repo.

Extracting VQVAE from nle-toolbox

brew install git-filter-repo

# rm -rf nle_toolbox--delete
git clone https://github.com/ivannz/nle_toolbox.git nle_toolbox--delete
cd nle_toolbox--delete

# remove each file other than VQ
git filter-repo --invert-paths --force --path nle_toolbox/zoo/__init__.py
git filter-repo --invert-paths --force --path nle_toolbox/zoo/blstats.py
git filter-repo --invert-paths --force --path nle_toolbox/zoo/glyph.py
git filter-repo --invert-paths --force --path nle_toolbox/zoo/models
git filter-repo --invert-paths --force --path nle_toolbox/zoo/transformer
git filter-repo --invert-paths --force --path nle_toolbox/zoo/legacy

# 
git filter-repo --path nle_toolbox/zoo/ --path-rename 'nle_toolbox/zoo/':''

# create an export bundle
git bundle create vqvae.bundle --all

Setup

The basic working environment is set up with the following commands:

# conda update -n base -c defaults conda
# conda deactivate && conda env remove -n vqvae

# pytorch, scip, scipy and other essentials
conda create -n vqvae "python>=3.10" pip setuptools numpy scipy \
  matplotlib notebook "pytorch::pytorch" pytest \
  && conda activate vqvae \
  && pip install tqdm einops torch-scatter \
    "black[jupyter]" pre-commit gitpython nbdime pydantic