The best way is to use Akos Hochrein's installation script setup.sh, now part of this repo (thanks, Akos!). It can work with either the Theano or the Tensorflow backend, and it installs into a virtualenv.

With Theano

./setup.sh

With TensorFlow

./setup.sh --tensorflow

If you have an Nvidia video card, you are advised to install CUDA and CuDNN. The first steps of the Theano installation instructions explain how:

For Ubuntu: http://deeplearning.net/software/theano/install_ubuntu.html#install-ubuntu

For OS X: http://deeplearning.net/software/theano/install.html#gpu-macos

The examples are 5-15 times faster when running on a GPU. On a Linux machine, it's also worth looking into the linear algebra libraries installed, see the above Ubuntu link for more information (Chapter: Troubleshooting: Make sure you have a BLAS library).

The theano subdirectory contains code snippets illustrating some tensor library concepts. For (much) more information, you can check out http://deeplearning.net/software/theano/tutorial/ or https://github.com/nlintz/TensorFlow-Tutorials .

We train and evaluate our own handwritten digit recognition network.

python mnist_mlp.py

The evaluation results are written to standard output, and a confusion matrix visualization is written to vis.png. The confusion matrix looks like this:

Unlike with the handwritten digit example, now we don't train a network, we just use one designed and trained by smart people.

You need the pre-trained VGG16 model weights file called vgg16_weights.h5, from https://drive.google.com/file/d/0Bz7KyqmuGsilT0J5dmRCM0ROVHc/view You have to open the link in a browser, wget is circumvented.

python vgg16_predictor.py image1.png image2.png ...

Output looks like this. The numbers are TOP 5 prediction probabilities, each line is corresponding to an image:

jersey = 0.194810 , electric_guitar = 0.056235 , acoustic_guitar = 0.039047 , lab_coat = 0.037690 , drumstick = 0.036288 ,
lorikeet = 0.937200 , hummingbird = 0.022311 , peacock = 0.013273 , macaw = 0.011136 , jacamar = 0.003892 ,
streetcar = 0.330774 , palace = 0.210260 , trolleybus = 0.080859 , cinema = 0.044910 , traffic_light = 0.032249 ,

By the way, most of the vgg16_predictor.py code is obsoleted by the recent publication of https://github.com/fchollet/deep-learning-models , which is more robust and general.

This code also requires the pre-trained VGG16 model weights, see above.

python vgg16_deep_dream.py input_image.png output_file_prefix

The final output is in output_file_prefix_at_iteration_9.png.

I have implemented a more advanced version of vgg16_deep_dream.py, that can work with any pre-trained Keras neural network. It's at https://github.com/danielvarga/keras-deep-dream .