Deepdream experiment implemented using Keras and VGG19 convnet
Jupyter Notebook
Switch branches/tags
Nothing to show
Permalink
Failed to load latest commit information.
.gitignore
Keras Deepdream.ipynb
LICENSE Initial commit Jul 9, 2017
README.md Update README.md Jul 12, 2017

README.md

Keras-based Deepdream experiment


See also: https://github.com/google/deepdream
Blog entry: https://www.bonaccorso.eu/2017/07/09/keras-based-deepdream-experiment-based-vgg19/

This experiment (which is a work in progress) is based on some suggestions provided by the Deepdream team in this blog post but works in a slightly different way. I use a Gaussian Pyramid and average the rescaled results of a layer with the next one. A total variation loss could be employed too, but after some experiments, I've preferred to remove it because of its blur effect.

Requirements

  • Python 2.7-3.5
  • Keras
  • Theano/Tensorflow
  • SciPy
  • Scikit-Image

Examples

(With different settings in terms of layers and number of iterations)

Berlin

Berlin

Cinque Terre

Rome

Tubingen

Taj Mahal

Adding noise

A good suggestion provided in this blog post is adding some noise to the original image. In this way there's a stronger activation of different filters. I suggest to try different values and/or removing the noise from the processed image.

def process_image(image, iterations=2, noise_level=5):
    # Create bounds
    bounds = np.ndarray(shape=(image.flatten().shape[0], 2))
    bounds[:, 0] = -128.0
    bounds[:, 1] = 128.0

    # Initial value
    x0 = image.flatten()
    
    # Add some noise
    noise = np.random.randint(-noise_level, noise_level, size=x0.shape)
    x0 = np.clip(x0 + noise, -128, 128)

    # Perform optimization
    result = minimize(fun=loss, 
                      x0=x0, 
                      args=list(image.shape), 
                      jac=gradient, 
                      method='L-BFGS-B', 
                      bounds=bounds, 
                      options={'maxiter': iterations})
    
    return postprocess_image(np.copy(result.x.reshape(image.shape)))

Creating videos

It's possible to create amazing videos by zooming into the same image (I've also added an horizontal pan that can be customized). You can use the snippet below, which assumes to have already processed an existing image (processed_image):

nb_frames = 3000

h, w = processed_image.shape[0:2]

for i in range(nb_frames):
    rescaled_image = rescale(processed_image, order=5, scale=(1.1, 1.1))
    rh, rw = rescaled_image.shape[0:2]
    
    # Compute the cropping limits
    dh = int((rh - h) / 2)
    dw = int((rw - w) / 2)
    
    dh1 = dh if dh % 2 == 0 else dh+1
    dw1 = dw if dw % 2 == 0 else dw+1
    
    # Compute an horizontal pan
    pan = int(45.0*np.sin(float(i)*np.pi/60))
    
    zoomed_image = rescaled_image[dh1:rh-dh, dw1+pan:rw-dw+pan, :]
    processed_image = process_image(preprocess_image(img_as_ubyte(zoomed_image)), iterations=2)
    
    imsave(final_image + 'img_' + str(i+1) + '.jpg', processed_image)

Sample animation

1.1x In-zoom with firth order interpolation and 1500 frames