H2O: Haja to Ocean

    

Machine learning with OrbitDB & Ocean Protocol

This app runs the Ocean Protocol Trilobite release code and latest OrbitDB.

You can find a live version of H2O at h2o.apps.outlierventures.io.

Publishing assets requires an Azure Storage account. Proof-of-concept OrbitDB hosting can be found in backend folder, see app.py and host.js.

Architecture

Dataflow

H2O works with any dataset that can be clustered with SciKit Learn's kmeans function hosted using OrbitDB. This is a set of datapoints with each dimension stored as a separate entry in an OrbitDB docs type database. To fetch your database, just enter your OrbitDB address in the H2O UI.

If you just have raw data (JSON format), you can upload this to a docs type database using the H2O-Host component. You can find the repo for H2O-Host here. The repo also includes a data generation script if you don't have any data yourself but still want to test H2O and generate some processed datasets for training AI.

Getting started

H2O runs on Linux and MacOS with command line tools.

Install components

sudo ./install

If you encounter any install problems, there is a full requirements list at the bottom of this file for manual installations.

Running components (local deployment)

Start an instance of Ocean Protocol:

./launch_ocean

Ensure Ocean is up and running before launching H2O - see the container output logs or use this script:

./backend/scripts/wait_for_migration_and_extract_keeper_artifacts.sh

In another teminal window, launch H2O:

./launch_h2o

Interact with the app in your browser at 0.0.0.0:4200.

Using the Kovan testnet

For the adventurous, take a look at the readme in the kovan folder.

Architecture diagram

Deployment

You can run an H2O client using nginx.

Move H2O to /var/www/your-domain-name.example.

You can quickly set up an nginx configuration using nginxconfig.io.

Set up HTTPS since users will be supplying an Azure storage key.

Get an instance of Ocean Protocol up and running.

In one screen session:

cd frontend
sudo ng build --watch --output-hashing=all

Output hashing is needed to prevent browsers loading their cached copies of the graphed datasets.

Point nginx to the dist folder ng-build produces (will be in frontend next to src).

In another:

cd backend
sudo python3 app.py

Running a static frontend through a secure web server like nginx and proxying backend is relatively safe as long as debug mode is off for the backend. For true security, switch your backend process to a production server.

Disconnect from your screen sessions and H2O will continue to run.

Full requirements list

If you encounter errors with the install script, here is a full list of requirements:

H2O runs on Linux and MacOS.

• MacOS: command line tools, Homebrew
• Linux: GCC 4+
• Python 3 (python3-dev on Linux)
• Pip3
• Node 8 (strictly version 8 - this is because of node-gyp)
• NPM 6+
• Angular CLI 1+
• Yarn 1.10+
• Finally, install dependencies:

  pip3 install --upgrade setuptools
  pip3 install wheel
  cd backend
  pip3 install -r requirements.txt
  npm install orbit-db ipfs
  cd ../frontend
  yarn install --pure-lockfile