Action Conditioned GANS

This is an implementation of conditional video prediction based on various generative adversarial networks. The technical details of this project are outlined here.

Some results (many gifs!): here

Dependencies

• tensorflow1.0 (tested on 1.0)
• numpy
• matplotlib
• opencv-python
• pillow
• imageio
• h5py

Installation

pip install requirements.txt

Data

This project is done on the Push Dataset from Google Brain.

Download from https://sites.google.com/site/brainrobotdata/home/push-dataset

Train

python train.py PATH/TO/INPUT/DATA OUTPUT/PATH

Training details

• --adv: boolean, whether to use adversarial loss
  • True (default)
• --loss: string, what loss to use
  • bce for cross entropy (default)
  • wass for wasserstein loss
• --opt: string, what optimizer to use
  • adam for ADAM Optimizer (default)
  • rmsprop for RMSProp
• --dna: boolean, whether to use dynamic neural advection
  • True (default)

Test

python test.py MODEL/PATH INPUT/FRAME/PATH INPUT/ACTION/PATH RESULT/SAVE/PATH

Note: Input frames and actions need to be numpy files for flexibility.

Testing details

• --dna: boolean, whether to use dynamic neural advection
  • True (default)