Visualization of Deep Covolutional Neural Networks

<!--- [](https://travis-ci.org/conan7882/CNN-Visualization) [](https://coveralls.io/github/conan7882/CNN-Visualization?branch=master) -->

• This repository contains implementations of visualizatin of CNN in recent papers.
• The source code in the repository can be used to demostrate the algorithms as well as test on your own data.

Requirements

• Python 3.3+
• Tensorflow 1.3
• TensorCV

Algorithms

• Visulization of filters and feature maps of GoogLeNet
• Deconvolutional Networks (ECCV'14)
• Guided back propagation (2014)
• Class Activation Mapping (CAM) (CVPR'16)
• Gradient-weighted Class Activation Mapping (Grad-CAM) (ICCV'17)

Visulization of filters and feature maps of GoogLeNet

• The most straightforward approach to visualize a CNN is to show the feature maps (activations) and filters.
• Details of the implementation and more results can be found here

<p align = 'left'> <img src ="doc/firstfilter/figs/GoogLeNet.png" height="250" /> <img src ="doc/firstfilter/figs/GoogLeNet_inception3a.png" height="250" /> </p>

Deconvnet

• Pick a specific activation on a feature map and set other activation to zeros, then reconstruct an image by mapping back this new feature map to input pixel space.
• Details of the implementation and more results can be found here. Some results:

<p align = 'left'> <img src ="doc/deconv/figs/people/conv1_2_feat.png" height="200" /> <img src ="doc/deconv/figs/people/conv1_2_im.png" height="200" /> <img src ="doc/deconv/figs/people/conv2_2_feat.png" height="200" /> <img src ="doc/deconv/figs/people/conv2_2_im.png" height="200" /> <img src ="doc/deconv/figs/people/conv3_4_feat.png" height="200" /> <img src ="doc/deconv/figs/people/conv3_4_im.png" height="200" /> <img src ="doc/deconv/figs/people/conv4_4_feat.png" height="200" /> <img src ="doc/deconv/figs/people/conv4_4_im.png" height="200" /> </p>

Guided back propagation

<!--- Guided backpropagation generates clearer visulizations than deconvnet for higher layers.-->

• Details of the implementation and more results can be found here. Some results:

Class Activation Mapping (CAM)

• The class activation map highlights the most informative image regions relevant to the predicted class. This map can be obtained by adding a global average pooling layer at the end of convolutional layers.
• Details of the implementation and more results can be found here. Some results:

Gradient-weighted Class Activation Mapping (Grad-CAM)

• Grad-CAM generates similar class heatmap as CAM, but it does not require to re-train the model for visualizatin.
• Details of the implementation and more results can be found here. Some results: