Deconvnet

This is a implementation of Deconvnet in keras, following Matthew D.Zeiler's paper Visualizing and Understanding Convolutional Networks

Feature

Given a pre-trained keras model, this repo can visualize features of specified layer including dense layer.

Dependencies

Keras 1.1(Theano Backend)
Python >= 2.7
argparse 1.0
Pillow 1.1

Examples

Below is several examples of feature visualization based on pre-trained VGG16 in keras, 'max' means pick the greates activation in the feature map to be visualized and set other elements to zeros, 'all' mean use all values in the feature map to visualize.
The left is max activation feature, the right is all activation feature

Original Image
block3_conv3_128
block4_conv2_46

The 46th feature in block4_conv2 layer catches 'nose' feature

max all

block5_conv3_256

The 256th feature in block5_conv3 layer catches 'ear' feature

max all

fc1_0
fc2_248
predictions_248

predictions_248 is the predicted class of the image(label: Eskimo_dog)

Shortage

The code implements visualize function for only Convolution2D, MaxPooling2D, Flatten, Input, Dense, Activation layers, thus cannot handle other type of layers.
The code support only plain networks, thus cannot visualize ResNet, Highway Networks or something.