AI Showcases - Center for Aritifical Intellgience and Robotics (CAIRO)

This repository contains the code for the AI use cases of the Center of Artificial Intelligence and Robotics (CAIRO)

Apps

• Jukebox (Music Embedding, 492 songs 15 artists)
• Twitter (Twitter Dataset Embedding ~60k Tweets from biggest Nasdaq companies.)
• Style Transfer (Transfer style to arbitrary image.)

Documentation

• Docs Jukebox
• Docs Style Transfer

Content

├── Jukebox                 # Mozart Jukebox app to visualize music embeddings. See:
│   ├── docs                # Explanations for jukebox
│   ├── index.html          # Main Webapp File
│   ├── data_storage        # Storage of trained embedding for mozart
│   ├── deploy.sh           # deployment file on the showcase server
│   └── others              # Favicons, licences etc.
├── style           # Style Transfer application. See:
│   ├── docs                # Explanations for jukebox
│   ├── index.html          # Main Webapp File
│   ├── data_storage        # Storage of trained embedding for mozart
│   ├── others              # Development files, favicons etcs.
│   ├── deploy.sh           # deployment file on the showcase server
│   └── saved_model_*       # Saved Tensorflow.Js models
├── Twitter                 # Twitter Embedding to visualize our Twitter dataset embeddings. See:
│   ├── index.html          # Main Webapp File
│   ├── data_storage        # Storage of trained embedding for mozart
│   ├── deploy.sh           # deployment file on the showcase server
└── └── others              # Favicons, licences etc. 

Development Twitter & Mozart Jukebox

The Embeddings for Mozart Jukebox and Twitter are stored under <app>/data_storage. New embeddings must also be stored there. Finally, the new embedding must be registered in <app>/data_storage/projector_config.json. How is done is self-explaining in the file itself.

Data generation for embeddings:

To save data comaptible with the tensorboard used at Jukebox or Twitter your data has to be float32 an in *.bytes format. This can achieved by:

# Save data
features.dtype=np.float32
features.tofile(board_data+data_path)

Deployment

All applications can be hosted via a simple webserver. In the deploy.sh is shown how to deploy to apache webserver.

Twitter build

The Twitter webapp is a yarn application. To build the app use:

yarn run prep
yarn run build

The build applications can be inspected for development via:

yarn run start

Credits

• Style app by https://github.com/reiinakano
• Twitter Embedding trained by Julian Hanke, Christian Niekler, Trong Duc Tran