• Up to date versions of Keras and Theano
• OpenCV3
• matplotlib, numpy, jupyter-notebook (recommend using Anaconda)

• Download the VGG16 keras weights and put them in ./

• Run Visualizing.ipynb cell by cell in jupyter-notebook

target_layer = "convolution2d_11"
feat_map = 12
output = deconv(model, target_layer, feat_map, im)`

Change target_layer and feat_map to what you want.

With receiptive field increases, activated area changes

Different filter map responses for different area of input image

Back pass from the last fully connected layer (represntation of classfication result)

• Visualizing and Understanding Convolutional Networks

• Images used in this example come from PASCAL VOC 2007 Dataset