Notebook example of how to generate class visualizations with Caffe
Jupyter Notebook
Latest commit 13f9539 Jun 18, 2016 @auduno update blogpost link


Image generated

DeepDraw is a ipython notebook example of generating class visualizations, such as the one above, from deep neural networks using Caffe. The examples and settings in this notebook was based on the pretrained GoogLeNet model available with Caffe, but it's easy to modify to use other networks, such as AlexNet. For some more detailed information about how these class visualizations are generated, check out this blogpost, and for some more examples of generated images, see this album of highlights or this album with all 1000 imagenet classes.

The repository also includes some code examples of drawing with the class visualizations, as described in this blogpost, in the folder "/other".

To run the code in this repository, you'll need an installation of Caffe with built pycaffe libraries, as well as the python libraries numpy, scipy and PIL. For instructions on how to install Caffe and pycaffe, refer to the installation guide here. Before running the ipython notebooks, you'll also need to download the bvlc_googlenet model, and insert the path of the pycaffe installation into pycaffe_path and the model path to the googlenet model into model_path.

If you create some cool work or visualizations based on this code, let me know via twitter!

This code was based on the deepdream code shared by Google, as well as some modifications kindly shared by Kyle McDonald. This repository is freely shared under Apache License 2.0.