• TensorFlow implementation of Striving for Simplicity: The All Convolutional Net (2014).
• Guided backpropagation generates clearer visulizations than deconvnet for higher layers.

• Python 3.3+
• Tensorflow 1.3
• TensorCV

• VGG19 is used for visulization. The model is defined in CNN-Visualization/lib/nets/vgg.py.
• Guided backpropagation copmutation class is defined in CNN-Visualization/lib/models/guided_backpro.py.
• Example usage is in CNN-Visualization/example/guided_backpropagation.py.

1. Setup directories in file CNN-Visualization/example/guided_backpropagation.py.

   • IM_PATH - directory of testing image data
   • VGG_PATH - directory of pre-trained VGG19 parameters
   • SAVE_DIR - directory of saving result images

2. Download the pre-trained VGG parameters

   • Download pre-trained VGG19 model here and put it in VGG_PATH.

3. Testing images

   • Put testing images in IM_PATH .

   • Change image type to the corresponding type if testing images are not jpeg files

     input_im = ImageFromFile('.jpg', data_dir=IM_PATH, num_channel=3, shuffle=False)
     

To get the guided backpropagation maps for all the image in IM_PATH, go to CNN-Visualization/example/ and run:

python guided_backpropagation.py

• Results will be saved in SAVE_DIR as map_IDX_class_PREDICT_LABEL.png.

Qian Ge