Skip to content
Implementation of the paper "Adaptive Skip Intervals: Temporal Abstraction for Recurrent Dynamical Models"
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.

Adaptive Skip Intervals - Experiments

This repository includes code to reproduce the results in the paper Adaptive Skip Intervals: Temporal Abstraction for Recurrent Dynamical Models.

  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)},

Funnel board animation
Room runner animation

The trained model in action: on the left is the ground truth, on the right we see the model's predictions. We synchronized the videos to better compare them. Note that the model skips hard-to-predict frames.

The skipping behavior can also be visualized using timelines. Lines between the frames visualize the frame similarity which is used in the temporal matching step of ASI.

Code to generate datasets is available in the separate repository:


  • imageio==2.1.2
  • keras==2.1.3
  • moviepy==
  • numpy==1.14.0
  • pandas==0.22.0
  • pillow==5.0.0
  • pytest==3.4.0
  • scikit-image==0.14.0
  • scipy==1.0.0
  • tensorflow==1.5.0
  • tqdm==4.11.2

Training the ASI model

To train the ASI model, please run, for example

python -m experiment.run_experiment 
    --delta_t_upper_bound 21  
    --exploration_steps 15000
    --optimizer adam 
    --schd_sampling_steps 15000 
    --f_init_learning_rate 0.0005 
    --f_learning_rate_decay_rate 0.2 
    --f_learning_rate_decay_steps 15000 
    --f_architecture f_simple 
    --n_kernels 48 
    --n_trajectories_per_batch 2 
    --z_loss_fn log_loss 
    --initializer he_uniform 
    --activation relu 
    --dataset fubo  
    --n_validation_examples 3 
    --train_filepath /path/to/train.tfrecords 
    --valid_filepath /path/to/valid.tfrecords 
    --n_epochs 150 
    --output_dir /path/to/training_output/
You can’t perform that action at this time.