danieldworakowski / visualbackprop

VisualBackProp - visualization method for convolutional neural networks

VisualBackProp

VisualBackProp - visualization method for convolutional neural networks

Description

Detailed description of the VisualBackProp can be found in: https://arxiv.org/abs/1611.05418

Content:

• inputImages - folder with input images
• outputImages - folder for output images
• vis.lua - implementation of VisualBackProp
• run.lua - loads images, runs visualization method, saves images
• model.t7b - model trained to predict steering wheel angle

Usage

mkdir outputImages

qlua run.lua

Script generates set of 3 images for each input image. Each set contain:

• Input image with visualization mask overlaid in red
• Averaged feature maps
• Intermediate masks