penny4860/cnn-visualizer

Visualization of the filters of VGG16, via gradient ascent in input space.

cnn-visualizer

Visualization of the filters of VGG16, via gradient ascent in input space.

Pure tensorflow version of How convolutional neural networks see the world

• Generated image of Convolutional Layer 3-1

  • 

• Generated image of Convolutional Layer 4-1

  • 

Prerequisites

• python 3.5
• tensorflow 1.0.1
• numpy 1.12.0

How to run

1. Download vgg16 checkpoint file

First, you need to download the pre-learned checkpoint file.

• Download the vgg16 checkpoint file at https://github.com/tensorflow/models/tree/master/research/slim#pre-trained-models
• Extract the downloaded file.
• Move the vgg_16.ckpt file to the project/root/ckpts directory.

2. Run script "main.py"

Now run main.py script. You can adjust options such as image size, layer name, and so on.

In the cpu version, it takes only a minute to create one image.