CompassionAI common repository

Common utilities and dataset preparation tools in use by various CompassionAI projects.

Installation

There are two modes for this library - inference and research. We provide instructions for Linux.

• Inference should work on MacOS and Windows mutatis mutandis.
• We very strongly recommend doing research only on Linux. We will not provide any support to people trying to perform research tasks without installing Linux.

Virtual environment

We strongly recommend using a virtual environment for all your Python package installations, including anything from CompassionAI. To facilitate this, we provide a simple Conda environment YAML file. We recommend first installing miniconda, see https://docs.conda.io/en/main/miniconda.html. We then recommend installing Mamba, see https://github.com/mamba-org/mamba.

bash Miniconda3-latest-Linux-x86_64.sh
conda install mamba -c conda-forge
mamba env create -f env-minimal.yml -n my-env
conda activate my-env

Inference

Just install with pip:

pip install compassionai-common

Research

Begin by installing for inference. Then install the CompassionAI data registry repo and set two environment variables:

$CAI_TEMP_PATH
$CAI_DATA_BASE_PATH

We strongly recommend setting them with conda in your virtual environment:

conda activate my-env
conda env config vars set CAI_TEMP_PATH=#directory on a mountpoint with plenty of space, does not need to be fast
conda env config vars set CAI_DATA_BASE_PATH=#absolute path to the CompassionAI data registry

Our code uses these environment variables to load datasets from the registry, output processed datasets and store training results.

You probably also want to install CUDA and PyTorch (>=1.12) with CUDA support - follow the instructions here https://pytorch.org/get-started/locally/. You don't need torchvision or torchaudio but it is safe to install them if you like. You can reinstall CUDA-enabled PyTorch with pip in your conda environment after installing everything as above.

For fine-tuning, you will need a powerful NVidia GPU. A GTX 1080 might work. We recommend at least an RTX 3080 Ti in a home setup, or a V100 if using a cloud. We have not tested non-NVidia GPUs.

For pre-training, you will need a TPU on GCP. We do not recommend fine-tuning on GPUs. We do not expect it to work on anything less than a DGX-2 or a p3dn.24xlarge instance.

Usage

Inference

This is a supporting library for our main inference repos, such as Lotsawa. You shouldn't need to use it directly.

Research

This library contains components that are common to the various tasks performed by the other libraries, such as Manas and Garland.

• Implements data loader objects such as KangyurLoader, TengyurLoader and TibetanDict.
• Implements common PyTorch dataset objects, such as TokenTagDataset.
• Provides utility functions for models and data, such as Hydra-Huggingface adapters, PyTorch callbacks, model downloaders and configuration providers.