Code for reproducing results of paper: "An Information-theoretic Progressive Framework for Interpretation" https://arxiv.org/abs/2101.02879 Authors: Zhengqi He @ NCA Lab, CBS, RIKEN, Japan Taro Toyoizumi @ NCA Lab, CBS, RIKEN & UTokyo, Japan
Dependencies: python (3.0) jupyter notebook numpy matplotlib pytorch (1.2.0) pycocotools
Dataset: We use standard CLEVR dataset in this project: https://cs.stanford.edu/people/jcjohns/clevr/ Download CLEVR v1.0 (18 GB) Extra data like calculated object mask and pretrained model is available at: https://riken-share.box.com/s/rc95iet4af680o3pll2sqywc6kevbduk Put folders "dataset" and "pretrained_models" inside the progressive_interpretation main folder
Run the code: To reproduce results in the paper, check out the jupyter notebooks: CLEVRTask1SuperviseLearning.ipynb CLEVRTask2MultipleChoice.ipynb