Jaynes, A Utility for training ML models on AWS, GCE, SLURM, with or without docker

Overview

The reality of ML training in universities is that we use what ever hardware we are given (for free). This means that we might have a few beefy GPU machines, an HPC cluster, plus some GCE/AWS credits that we get through grants. Jaynes is a well-designed python package that makes running across these inhomogenous hardward resources a pleasure.

install (requires unix operating system.)

pip install jaynes

The best way to get started with jaynes is to take a look at one of the example projects in the [geyang/jaynes-starter-kit]. For a rough idea, here is how to use jaynes to launch a training function:

To run locally:

import jaynes

def training(arg_1, key_arg=None, key_arg_2=None):
    print(f'training is running! (arg_1={arg_1}, key_arg={key_arg})')

jaynes.config(mode="local", arg_1=10, key_arg=0.3)
jaynes.run(training)
jaynes.listen()

We recommend setting up a main training function with the following sinature:

from params_proto import ParamsProto

class Args(ParamsProto):
    seed = 100
    lr = 3e-4
    # ...
    
def main(**deps):
    from ml_logger import logger
    
    Args._update(deps)
    logger.log_params(Args=vars(Args))
    
    # ... your main training steps

This way you can call the main fn directly for local debugging, but launch it as an entry point at scale.

Setup

Jaynes has gone through a large number of iterations. This version incorporates best practices we learned from other open-source communities. You can specify a jaynes.yml config file (copy one from our sample project to get started!) for the type of hosts (ssh/docker/singularity) and launchers (ssh/ec2/gce/slurm), so that none of those settings need to appear in your ML python script. When called from python, Jaynes automatically traverses the file tree to find the root of the project, the same way as git.

For example, to run your code on a remote computer via ssh:

# your_project/jaynes.yml
version: 0
verbose: true
run:  # this is specific to each launch, and is dynamically overwritten in-memory
  mounts:
    - !mounts.S3Code
      s3_prefix: s3://ge-bair/jaynes-debug
      local_path: .
      host_path: /home/ubuntu/
      container_path: /Users/geyang/learning-to-learn
      pypath: true
      excludes: "--exclude='*__pycache__' --exclude='*.git' --exclude='*.idea' --exclude='*.egg-info'   --exclude='*.pkl'"
      compress: true
  runner:
    !runners.Docker
    name:   # not implemented yet
    image: "episodeyang/super-expert"
    startup: "yes | pip install jaynes ml-logger -q"
    work_directory: "{mounts[0].container_path}"
    ipc: host
  host:
    envs: "LANG=utf-8"
    pre_launch: "pip install jaynes ml-logger -q"
  launch:
    type: ssh
    ip: <your ip address>
    username: ubuntu
    pem: ~/.ssh/your_rsa_key

In python (your code):

# your_project/launch.py
import jaynes

def training(arg_1, key_arg=None):
    print(f'training is running! (arg_1={arg_1}, key_arg={key_arg})')

jaynes.run(training)

Using Modes

A lot of times you want to setup a different run modes so it is easy to switch between them during development.

# your_project/jaynes.yml
version: 0
mounts: # mount configurations Available keys: NOW, UUID,
  - !mounts.S3Code &code_mount
    s3_prefix: s3://ge-bair/jaynes-debug
    local_path: .
    host_path: /home/ubuntu/jaynes-mounts/{NOW:%Y-%m-%d}/{NOW:%H%M%S.%f}
    # container_path: /Users/geyang/learning-to-learn
    pypath: true
    excludes: "--exclude='*__pycache__' --exclude='*.git' --exclude='*.idea' --exclude='*.egg-info' --exclude='*.pkl'"
    compress: true
hosts:
  hodor: &hodor
    ip: <your ip address>
    username: ubuntu
    pem: ~/.ssh/incrementium-berkeley
runners:
  - !runners.Docker &ssh_docker
    name: "some-job"  # only for docker
    image: "episodeyang/super-expert"
    startup: yes | pip install jaynes ml-logger -q
    envs: "LANG=utf-8"
    pypath: "{mounts[0].container_path}"
    launch_directory: "{mounts[0].container_path}"
    ipc: host
    use_gpu: false
modes: # todo: add support to modes.
  hodor:
    mounts:
      - *code_mount
    runner: *ssh_docker
    launch:
      type: ssh
      <<: *hodor

now run in python

# your_project/launch.py
import jaynes

def training(arg_1, key_arg=None):
    print(f'training is running! (arg_1={arg_1}, key_arg={key_arg})')

jaynes.config(mode="hodor")
jaynes.run(training)

ToDos

• more documentation
• singularity support
• GCE support
• support using non-s3 code repo.

Done

• get the initial template to work

Installation

pip install jaynes

Usage (Show me the Mo-NAY!! 💰💸)

Check out the test_projects folder for projects that you can run.

To Develop

git clone https://github.com/episodeyang/jaynes.git
cd jaynes
make dev

To test, run

make test

This make dev command should build the wheel and install it in your current python environment. Take a look at the ./Makefile for details.

To publish, first update the version number, then do:

make publish

Acknowledgements

This code is inspired by @justinfu's doodad, which is in turn built on top of Peter Chen's script.

This code is written from scratch to allow a more permissible open-source license (BSD). Go bears 🐻 !!