Sequential prediction tasks for Adaptive Skip Intervals

This repository includes visual prediction tasks for the paper Adaptive Skip Intervals: Temporal Abstraction for Recurrent Dynamical Models

@inproceedings{neitz2018adaptive,
  title={Adaptive Skip Intervals: Temporal Abstraction for Recurrent Dynamical Models},
  author={Neitz, Alexander and Parascandolo, Giambattista and Bauer, Stefan and Sch{\"o}lkopf, Bernhard},
  booktitle={Advances in Neural Information Processing Systems (NIPS)},
  year={2018}
}

See repository adaptive-skip-intervals for an implementation of the ASI algorithm.

Currently implemented tasks are:

• Funnel board:
  
  Task: Given first frame of the trajectory, predict platform where the ball will land.

• Room runner:
  
  Task: Given the first frame of the trajectory, predict color of the room in which the green dot will end up.

Dependencies

• box2d==2.3.2
• cairocffi==0.8.0
• gizeh=0.1.10
• imageio==2.1.2
• moviepy==0.2.3.2
• numpy==1.14.0
• pillow==5.0.0
• tqdm==4.11.2
• tensorflow==1.5.0

Generate datasets

Room runner:
python -m generate.generate_dataset --dataset rr --seed 1234 --n_trajectories 500 --output_dir /path/to/dataset/directory/

Funnel board:
python -m generate.generate_dataset --dataset fubo --seed 1234 --n_trajectories 500 --output_dir /path/to/dataset/directory/

Use --n_processes N to use N parallel workers (results in nondeterministic ordering of examples).