README.md

DanceNet - Dance generator using Variational Autoencoder, LSTM and Mixture Density Network. (Keras)

This is an attempt to create a dance generator AI, inspired by this video by @carykh

Main components:

• Variational autoencoder
• LSTM + Mixture Density Layer

Requirements:

• Python version = 3.5.2

  Packages

  • keras==2.2.0
  • sklearn==0.19.1
  • numpy==1.14.3
  • opencv-python==3.4.1

Dataset

https://www.youtube.com/watch?v=NdSqAAT28v0 This is the video used for training.

How to run locally

• Download the trained weights from here. and extract it to the dancenet dir.
• Run dancegen.ipynb

How to run in your browser

• Click the button above to open this code in a FloydHub workspace (the trained weights dataset will be automatically attached to the environment)
• Run dancegen.ipynb

Training from scratch

• fill dance sequence images labeled as 1.jpg, 2.jpg ... in imgs/ folder
• run model.py
• run gen_lv.py to encode images
• run video_from_lv.py to test decoded video
• run jupyter notebook dancegen.ipynb to train dancenet and generate new video.

References

• Original project video by Cary Huang
• Building autoencoders in keras by Francois Chollet
• Time Series Prediction with LSTM Recurrent Neural Networks in Python with Keras
• Generative Choreography using Deep Learning
• Mixture Density Networks by David Ha
• Mixture Density Layer for Keras by Charles Martin