Meta-Regularization by enforcing Mutual-Exclusiveness

This is the official implementation of paper - "Meta-Regularization by Enforcing Mutual-Exclusiveness" (https://arxiv.org/abs/2101.09819)

In our work, we propose a regularization technique for meta-learning models that gives the model designer more control over the information flow during meta-training. Our method consists of a regularization function that is constructed by maximizing the distance between task-summary statistics, in the case of black-box models and task specific network parameters in the case of optimization based models during meta-training.

Setup Instructions

• To setup conda environment conda env create -f conda_env.yml

Omniglot Dataset

• To prepare non-mutual-exclusive dataset, use the script: src/prepOmniglotDataset.py.

Pose Dataset

• Copy the license text file mjkey.txt at src/pose_data/.
• Download CAD models from Beyond PASCAL: A Benchmark for 3D Object Detection in the Wild PASCAL3D+_release1.1.
• Configure data path in the script src/pose_data/convert_and_render.sh and run this script. This will render the dataset using CAD models and save png files along with labels at configured directory. It will also generate pickle files of the dataset.
• Or you can download the prepared pickle file from here: https://drive.google.com/drive/folders/1V_9NqqelQyuyYtPv6ndoqXQp-mp_zmeG?usp=sharing

Running the experiments

• To run experiments:

sh run_experiments.sh