Neural Systematic Binder

ICLR 2023

[arXiv] [project] [datasets] [openreview]

This is the official PyTorch implementation of Neural Systematic Binder.

Authors

Gautam Singh and Yeongbin Kim and Sungjin Ahn

Datasets

The datasets tested in the paper (CLEVR-Easy, CLEVR-Hard, and CLEVR-Tex) can be downloaded via this link.

Training

To train the model, simply execute:

python train.py

Check train.py to see the full list of training arguments. You can use the --data_path argument to point to the set of images via a glob pattern.

Outputs

The training code produces Tensorboard logs. To see these logs, run Tensorboard on the logging directory that was provided in the training argument --log_path. These logs contain the training loss curves and visualizations of reconstructions and object attention maps.

Packages Required

The following packages may need to be installed first.

PyTorch
TensorBoard for logging.

Evaluation

The evaluation scripts are provided in branch evaluate.

Citation

@inproceedings{
      singh2023sysbinder,
      title={Neural Systematic Binder},
      author={Gautam Singh and Yeongbin Kim and Sungjin Ahn},
      booktitle={International Conference on Learning Representations},
      year={2023},
      url={https://openreview.net/forum?id=ZPHE4fht19t}
}

singhgautam / sysbinder