ScaleShift

An Open Source Machine Learning Tool for making & training models.

Overview

ScaleShift is a client web application which makes it easier for researchers to build machine learning models locally and to train them on premise/cloud!

Key Capabilities

• A client application that has web-based user-friendly interfaces
• Can be run anywhere (Linux, macOS, Windows..)
• Supports any Python-based dockerized algorithms or software
• Fully integrated with NGC, Rescale platform & Kubernetes

How It Works

1. Setup a ScaleShift client

With just only 3 steps!

2. Download machine learning software

You can pull any docker images to the client from NGC, DockerHub or your private registry with just one click.

3. Create a new workspace

When you click a run button, ScaleShift wraps the image with Jupyter notebook & run it as a docker container. Then you can build your own models on your specified software on it.

4. Train your models

In order to train the models in a parallel and distributed way, you can choose a Kubernetes cluster or Rescale platform. ScaleShift converts the image to Singularity automatically if it’s needed.

Get Started

Official Builds

Try ScaleShift on AWS

1. Create an EC2 instance with CloudFormation

2. Access the Web UI

public_ip=$( sh -c "$( aws cloudformation describe-stacks --stack-name "scaleshift" \
  --query 'Stacks[0].Outputs[?OutputKey==`PublicIPs`].OutputValue' \
  --output text )" )
open "http://${public_ip}"

Local Installation

1. Install dependent softwares

• Docker
• Docker Compose

2. Download the latest configuration

curl -so docker-compose.yml https://s3-ap-northeast-1.amazonaws.com/scaleshift/docker-compose-8080.yml

3. Start services

docker-compose up

Usage

Access the Web UI

http://localhost:8080

Access APIs directly

curl -sX POST -H "Content-Type: application/json" \
    -d '{"image": "tensorflow/tensorflow:1.14.0-py3"}' \
    http://localhost:8080/api/v1/notebooks
curl -sX GET -H "Content-Type: application/json" \
    http://localhost:8080/api/v1/notebooks

Contribution

1. Fork (https://github.com/rescale-labs/scaleshift/fork)
2. Create a feature branch
3. Commit your changes
4. Rebase your local changes against the master branch
5. Create new Pull Request

Copyright and license

Code released under the MIT license.