Code for Visualizing Deep Convolutional Neural Networks Using Natural Pre-images, IJCV 2016
This code is not very easy to use at the moment. All the visualizations were generated by running experiments/exp_all.m . Several parts of this are commented out for now. Please open this file and uncomment the visualization you are interested in. If you are using the maximization part of the code then you need to precompute network statistics using the helpers/get_neuron_stats.m script which uses imagenet to compute neuron max, min, mean values.
Download the pics from https://onedrive.live.com/redir?resid=5C675A68487F837F!204464&authkey=!AGwpJcGn75t4eic&ithint=file%2czip as data/pics/pics.zip and unzip this archive
cd data/pics/
unzip pics.zip
cd ../../
Softlink matconvnet and vlfeat into the root folder of this repository. Start matlab and run dg_setup.m. Then go the experiments folder and run exp_setup.m