Jaynes Examples: Cross-Provider Computation at Scale

This repository is an up-to-date collection of minimal jaynes usage examples. You can mix and match configurations between these included usecases for your particular infrastructure. You can find the up-to-date copy of this guide here: https://github.com/geyang/jaynes-starter-kit

To Get Started

First let's install Jaynes! This tutorial is written w.r.t version: 0.7.2

pip install jaynes

I would also recommend taking a look at params-proto, which is a pythonic hyperparameter + argparsing library that makes parameter management declaritive and error-free. We use params-proto and its sweep utility, params_proto.hyper in our parameter sweep example. To install params-proto, run

pip install params-proto waterbear

Table of Contents

For detailed documentation on each usecases, refer to the in-dept tutorial bellow:

1. Multiple SSH Reacheable Machines
2. Compute at Scale with SLRUM & Jaynes
3. Advanced Multi-mode Example
4. SSH Docker Configuration
5. EC2 Docker Configuration
6. GCP Docker Example
7. Jaynes Manager Server
8. Guide for MIT Supercloud
9. Using SBATCH Mode with SLUR
10. Using mpirun

In this folder, we provide a collection of example configurations. Each example sits within its own folder. To run, follow the instruction in the README in that example project.

01_ssh_docker_configuration
├── README.md
├── launch_entry.py
└── .jaynes.yml

Reporting Issues (on the Jaynes Repo/issues)

Let's collect all issues on the main jaynes repo's issue page, so that people can search for things more easily!

How to Debug

Jaynes offer a way to transparently debug the launch via verbose mode, where it prints out all of the local and remote script that it generates. To debug a launch script, set verbose to true either in the yaml file, or through the jaynes.config call. To debug in the remote host where you intend to run your job, you can often copy and paste the generated script to see the error messages.

Debugging Steps:

1. Turn on verbose mode, by setting verbose=True in the jaynes call

#! launch_entry.py
import jaynes

jaynes.config(verbose=True)

or

#! .jaynes.yml
verbose: true
runner:
- ....

2. Launch

#! launch_entry.py
if __name__ == "__main__":
    jaynes.run(train_fn, *args, **kwargs)
 
# if in SLURM or SSH mode:
jaynes.listen()  # to listen to the stdout/stderr pipe-back

3. Debug Suppose you have an error message. You can copy and paste the script ran by jaynes, that is printed out in the console either locally or on the EC2 instance you just launched to debug the specifics of it.

4. Share with Lab mates When you are done, you can share this repo with others who use the same infrastructure, so that they can run their code there too.

Call for Contributors

Machine Learning infrastructure is an evolving problem, and would take the rest of the community to maintain, adopt and standardize.

Below are a few areas that current stands in need to contributions: (now mostly done)

• [done] - Documentation on Configuration Schema issue #2
• [done] - GCE Support issue #3
• [done] - Pure SSH Host Support issue #4
• [done] - SLURM SBatch Support issue #5
• SLURM Singularity Support issue #6