mogwai: Probabilistic Models of Protein Families

A library of tested models, metrics, and data loading for protein families. Implemented in PyTorch and PyTorch Lightning.

Under active development, feedback welcome.

Getting Started

For now, we support cloning and installing in developer mode.

pip install -e .

You will also need to install apex (needed to use DDP and FusedLamb training):

source venv/bin/activate
git clone git@github.com:NVIDIA/apex.git
cd apex

Modify setup.py

Find:

if (bare_metal_major != torch_binary_major) or (bare_metal_minor != torch_binary_minor):

And replace it with

if (bare_metal_major != torch_binary_major):

To remove the minor version check. This will allow apex to install.

Install apex

pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./

Examples

• Potts Model with Pseudolikelihood