Attribution Preservation in Network Compression for Reliable Network Interpretation

This repository is the official PyTorch implementation of Attribution Preservation in Network Compression for Reliable Network Interpretation by Geondo Park, June Yong Yang, Sung Ju Hwang, Eunho Yang, NeurIPS 2020.

Requirements

The following requirements must be met:

• Python 3.6+
• CUDA 10.0+
• torch 1.1+
• torchvision 0.4+
• SciPy 0.19.1+
• imageio 2.8.0+

Preparing the Dataset

Create ./data directory and extract the PASCAL VOC 2012 dataset inside.

mkdir data
wget -nc -P ./data http://pjreddie.com/media/files/VOCtrainval_11-May-2012.tar
tar -xvf ./data/VOCtrainval_11-May-2012.tar -C ./data

Training the original model(teacher)

python voc_vanilla_main.py --gpu --exp_name "tutorial_teacher" --data_path ../../data/VOCdevkit/VOC2012 --imagenet_pretrained

Knowledge distillation

Overall hyperparameters for training are set by default. For stochastic matching, add --stochastic --drop_percent 0.3 arguments for all below scripts.

Inside directory ./pascal/distillation:

Distillation with sensitivity weighted attribution matching

python voc_distill_main.py --gpu --exp_name "tutorial_distillation" --arch VGG16_VOC_x8 --teacher_path weights/tutorial_teacher/tutorial_teacher_best.pth.tar --data_path ../../data/VOCdevkit/VOC2012 --transfer_type swa --transfer_weight 100

Structured Pruning (One-shot) with sensitivity weighted attribution matching

For stochastic matching, add --stochastic --drop_percent 0.3 arguments for all below scripts.

Overall hyperparameters for training are set by default. In folder ./pascal/structure:

python iteratrive_prune_distill.py --gpu --exp_name name --teacher_path weights/tutorial_teacher/tutorial_teacher_best.pth.tar --model_path <saved-start-model-path> --data_path ../../data/VOCdevkit/VOC2012 --percent 0.3 --iter 1 --transfer_type swa --transfer-weight 100

UnStructured Pruning (Iterative) with sensitivity weighted attribution matching

For stochastic matching, add --stochastic --drop_percent 0.3 arguments for all below scripts.

Overall hyperparameters for training are set by default. In folder ./pascal/unstructure:

python iteratrive_prune_distill.py --gpu --exp_name name --teacher_path weights/tutorial_teacher/tutorial_teacher_best.pth.tar --model_path <saved-start-model-path (teacher_path)> --data_path ../../data/VOCdevkit/VOC2012 --percent 0.2 --iter 16 --transfer_type swa --transfer-weight 100

Evaluation

To evaluate attribution scores, GradCam, LRP, EBP, and RAP were used.

To evaluate prediction and localization, run:

python evaluate_pred.py --gpu --data_path ../data/VOCdevkit/VOC2012 --path distillation/weights/tutorial_distill/tutorial_distill_best.pth.tar --compression kd --arch VGG16_VOC_x8

python evaluate_localization.py --gpu --data_path ../data/VOCdevkit/VOC2012 --path distillation/weights/tutorial_distill/tutorial_distill_best.pth.tar --compression kd --arch VGG16_VOC_x8 --method gcam --metric auc

Implementation

• Our network pruning code is implemented based on : rethinking-network-pruning
• Implementation of attribution maps are based on : Relative_Attributing_Propagation

Citation

@inproceedings{park2020attribution,
  title={Attribution Preservation in Network Compression for Reliable Network Interpretation},
  author={Park, Geondo and Yang, June Yong and Hwang, Sung Ju and Yang, Eunho},
  journal={Advances in Neural Information Processing Systems},
  volume={33},
  year={2020}
}