Polar Prediction Model

Implementation of A polar prediction model for learning to represent visual transformations, pulished at NeurIPS 2023.

A minimal and self-contained pytorch example is provided in notebooks/tutorial.ipynb. Below are instructions to reproduce the results described in the paper.

Install

git clone https://github.com/pehf/PolarPrediction.git
cd PolarPrediction
conda env create -f environment.yml
conda activate ppm
pip install -e .

Usage

Data

To download the datasets, run:

bash data/download.sh

Data will be stored in ~/Documents/datasets/.

Training

The code is setup to run on a SLURM cluster. To train the models, run:

bash slurm/NeurIPS23/all.sh

To view learning curves (and weights) in Tensorboard, run:

tensorboard --logdir checkpoints/

Results

Performance results and example predictions are gathered in notebooks/consolidate.ipynb.

Note that to run notebooks, jupyterlab is required and can be installed by running:

pip install jupyterlab

Citation

@inproceedings{fiquet2023polar,
  title={A polar prediction model for learning to represent visual transformations},
  author={Fiquet, Pierre-{\'E}tienne H and Simoncelli, Eero P},
  booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
  year={2023}
}

Visualization

Animation of two sets of filters, trained on synthetic data consisting respectively of translations, and of rotations of image.