Mjolnir

Mjolnir: Thor's hammer, a divine instrument making its holder worthy of wielding lightning.

Template repository for managing machine learning research projects built with
PyTorch-Lightning, using Anaconda
for Python Dependencies and Sane Quality Defaults (Black, Flake, isort).

Template created by Sidd Karamcheti.

Contributing

Key section if this is a shared research project (e.g., other collaborators). Usually you should have a detailed set
of instructions in CONTRIBUTING.md - Notably, before committing to the repository, make
sure to set up your dev environment and pre-commit install (pre-commit install)!

Here are sample contribution guidelines (high-level):

Install and activate the Conda Environment using the QUICKSTART instructions below.

On installing new dependencies (via pip or conda), please make sure to update the environment-<ID>.yaml files
via the following command (note that you need to separately create the environment-cpu.yaml file by exporting from
your local development environment!):

make serialize-env --arch=<cpu | gpu>

Quickstart

Note: Replace instances of mjolnir and other instructions with instructions specific to your repository!

Clones mjolnir to the working directory, then walks through dependency setup, mostly leveraging the
environment-<arch>.yaml files.

Shared Environment (for Clusters w/ Centralized Conda)

Note: The presence of this subsection depends on your setup. With the way the Stanford NLP Cluster has been set up, and
the way I've set up the ILIAD Cluster, this section makes it really easy to maintain dependencies across multiple
users via centralized conda environments, but YMMV.

@Sidd (or central repository maintainer) has already set up the conda environments in Stanford-NLP/ILIAD. The only
necessary steps for you to take are cloning the repo, activating the appropriate environment, and running
pre-commit install to start developing.

Local Development - Linux w/ GPU & CUDA 11.0

Note: Assumes that conda (Miniconda or Anaconda are both fine) is installed and on your path.

Ensure that you're using the appropriate environment-<gpu | cpu>.yaml file --> if PyTorch doesn't build properly for
your setup, checking the CUDA Toolkit is usually a good place to start. We have environment-<gpu>.yaml files for CUDA
11.0 (and any additional CUDA Toolkit support can be added -- file an issue if necessary).

git clone https://github.com/pantheon-616/mjolnir.git
cd mjolnir
conda env create -f environments/environment-gpu.yaml  # Choose CUDA Kernel based on Hardware - by default used 11.0!
conda activate mjolnir
pre-commit install  # Important!

Local Development - CPU (Mac OS & Linux)

Note: Assumes that conda (Miniconda or Anaconda are both fine) is installed and on your path. Use the -cpu
environment file.

git clone https://github.com/pantheon-616/mjolnir.git
cd mjolnir
conda env create -f environments/environment-cpu.yaml
conda activate mjolnir
pre-commit install  # Important!

Usage

This repository comes with sane defaults for black, isort, and flake8 for formatting and linting. It additionally
defines a bare-bones Makefile (to be extended for your specific build/run needs) for formatting/checking, and dumping
updated versions of the dependencies (after installing new modules).

Other repository-specific usage notes should go here (e.g., training models, running a saved model, running a
visualization, etc.).

Repository Structure

High-level overview of repository file-tree (expand on this as you build out your project). This is meant to be brief,
more detailed implementation/architectural notes should go in ARCHITECTURE.md.

  • conf - Quinine Configurations (.yaml) for various runs (used in lieu of argparse or typed-argument-parser)
  • environments - Serialized Conda Environments for both CPU and GPU (CUDA 11.0). Other architectures/CUDA toolkit
    environments can be added here as necessary.
  • src/ - Source Code - has all utilities for preprocessing, Lightning Model definitions, utilities.
  • preprocessing/ - Preprocessing Code (fill in details for specific project).
  • models/ - Lightning Modules (fill in details for specific project).
  • tests/ - Tests - Please test your code... just, please (more details to come).
  • train.py - Top-Level (main) entry point to repository, for training and evaluating models. Can define additional
    top-level scripts as necessary.
  • Makefile - Top-level Makefile (by default, supports conda serialization, and linting). Expand to your needs.
  • .flake8 - Flake8 Configuration File (Sane Defaults).
  • .pre-commit-config.yaml - Pre-Commit Configuration File (Sane Defaults).
  • pyproject.toml - Black and isort Configuration File (Sane Defaults).
  • ARCHITECTURE.md - Write up of repository architecture/design choices, how to extend and re-work for different
    applications.
  • CONTRIBUTING.md - Detailed instructions for contributing to the repository, in furtherance of the default
    instructions above.
  • README.md - You are here!
  • LICENSE - By default, research code is made available under the MIT License. Change as you see fit, but think
    deeply about why!

Start-Up (from Scratch)

Use these commands if you're starting a repository from scratch (this shouldn't be necessary for your collaborators
, since you'll be setting things up, but I like to keep this in the README in case things break in the future).
Generally, if you're just trying to run/use this code, look at the Quickstart section above.

GPU & Cluster Environments (CUDA 11.0)

conda create --name mjolnir python=3.8
conda install pytorch torchvision torchaudio cudatoolkit=11.0 -c pytorch   # CUDA=11.0 on most of Cluster!
conda install ipython
conda install pytorch-lightning -c conda-forge

pip install black flake8 isort matplotlib pre-commit quinine wandb

# Install other dependencies via pip below -- conda dependencies should be added above (always conda before pip!)
...

CPU Environments (Usually for Local Development -- Geared for Mac OS & Linux)

Similar to the above, but installs the CPU-only versions of Torch and similar dependencies.

conda create --name mjolnir python=3.8
conda install pytorch torchvision torchaudio -c pytorch
conda install ipython
conda install pytorch-lightning -c conda-forge

pip install black flake8 isort matplotlib pre-commit quinine wandb

# Install other dependencies via pip below -- conda dependencies should be added above (always conda before pip!)
...

Containerized Setup

Support for running mjolnir inside of a Docker or Singularity container is TBD. If this support is urgently required,
please file an issue.

GitHub - siddk/mnist-vae-scaling at pythonawesome.com
Tutorial Repository with a minimal example for showing how to deploy training across various compute infrastructure. - GitHub - siddk/mnist-vae-scaling at pythonawesome.com
GitHubsiddk