InputRepair

Code Release and pre-trained models for InputRepair.

Introduction

In this paper we describe InputRepair, an on-the-fly input repair technique for detected failure-inducing inputs. It is an input debugging and repairing technique for deep neural networks.

Repo structure

• debug.py: code for find confusing regions from input
• grad_cam.py: code for calculating heatmap scores for each layer
• repair.py: code for repairing the input regions
• kdes_generation.py: code for generating kde values for each layer
• search_occlusion.py: code for the crash occlusion position for each testing image on which the original classifier begins to mis-predict
• store_occlused_data.py: code to store the occlused testing dataset
• models/: pre-trained models
• tmp/: temp files
• images/: additional results

Dependencies

• conda create --name myenv
• pip install -r requirements.txt

How to run

• To locate confusing regions: python debug.py --d "cifar10" --m "conv"
• To repair input: python repair.py --d "cifar10" --m "conv"

Additional results

We evaluate the accuracy of Grad-CAM and InputRepair's debugging module (IR-Debug) with a set of occluded images that are mis-predicted by the DNN models. As shown in Table 4, IR-Debug always performs better than Grad-CAM. This demonstrates that the debugging module in InputRepair works well in practice.