Code to accompany https://openreview.net/pdf?id=b9PoimzZFJ.

Requirements are Python 3, TensorFlow v1.14, Numpy, Scipy, Scikit-Learn, Matplotlib, Pillow, Scikit-Image, h5py, tqdm. Experiments were run on V100 GPUs (16 and 32GB).

The mnist folder contains code used to run the Coloured-MNIST experiments. The data folder in it contains data generators and includes a sample dataset generation. The inferred partitions that were used are also included in predicted_partitions.

For example, in the mnist folder, you could run

python main.py -method pgi -wc 50.0 -u 5

Use the -v flag when performing validation.

For coco you can create the datasets using the code in data_makers, which will require installing the cocoapi and downloading the Places dataset. A sample of generated COCO datasets can be downloaded here.

For example, in the coco folder, you could run

python main.py -dataset colour -method pgi -wc 100.0 -u 200
python main.py -dataset places -sample normal -method pgi -wc 50.0 -u 200

BibTex:

@proceedings{ahmed2021systematic,
  title={Systematic generalisation with group invariant predictions},  
  author={Ahmed, Faruk and Bengio, Yoshua and van Seijen, Harm and Courville, Aaron},  
  booktitle={9th International Conference on Learning Representations (ICLR)},  
  year={2021}  
}