Implementation of the Deconvnet algorithm

N.B. It's still early days for this code. Suggestions for improvement are welcome !

Requirements

• Keras==2.0.8
• numpy==1.13.3
• opencv_python==3.3.0.10
• matplotlib==2.0.2
• Theano==0.9.0

Configure

• Unzip the images in ./Data/Img or add your own
• Download the VGG16 keras weights and put them in ./Data

How to use

• Open VGG_deconv.py in your text editor

Get the deconvolution result of random image, for a specific layer and feature map

• Go to step 4, set the flag to True
• Specify your target_layer (keras default naming, i.e convolution2d_5, convolution2d_7 etc)
• Specify the index of the filter map you want to look at
• Specify num_img (must be a perfect square) ex: 25 => plot 5x5
• run python VGG_deconv.py

Get the deconvolution of the top 9 images that maximally activate a given target_layer/map

• Go to step 1, set the flag to True
• Specify which layer you want in d_act e.g. d_act = {"convolution2d_13": {}}
• By default, the algorithm will look for the images maximally activating the first 10 feature maps of this layer
• Go to step 2 and set the flag to True
• Then run python VGG_deconv.py. You may have to change the batch size depending on your GPU
• It will take a little while to run and find the top 9 images
• Once this step is complete
• Go to step 3 and set the flag to True
• Select the layer for which you want to plot the top9 deconv images
• run python VGG_deconv.py

Examples

N.B. The VGG weights have not been trained on the Kaggle dataset !

Below, the top 9 images activating a selection of filters in convolutional layer 10. Each block is 2x3x3 (1 block: original crop of the image, 1 block: deconvolution result) Each 2x3x3 block corresponds to a specific feature map.

Below, the top 9 images activating a selection of filters in convolutional layer 13. Each block is 2x3x3 (1 block: original crop of the image, 1 block: deconvolution result) Each 2x3x3 block corresponds to a specific feature map.

It's interesting to investigate what each feature map is looking for in the original image. For instance, the top right block of convolutional layer 13 strongly reacts to the steering wheel.