deepdream_solution.py

"""DeepDream.
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import os.path
import zipfile

import sys
sys.path.extend(['', '/Users/shlens/Desktop/Neural-Art/homebrew/Cellar/python/2.7.13/Frameworks/Python.framework/Versions/2.7/lib/python27.zip', '/Users/shlens/Desktop/Neural-Art/homebrew/Cellar/python/2.7.13/Frameworks/Python.framework/Versions/2.7/lib/python2.7', '/Users/shlens/Desktop/Neural-Art/homebrew/Cellar/python/2.7.13/Frameworks/Python.framework/Versions/2.7/lib/python2.7/plat-darwin', '/Users/shlens/Desktop/Neural-Art/homebrew/Cellar/python/2.7.13/Frameworks/Python.framework/Versions/2.7/lib/python2.7/plat-mac', '/Users/shlens/Desktop/Neural-Art/homebrew/Cellar/python/2.7.13/Frameworks/Python.framework/Versions/2.7/lib/python2.7/plat-mac/lib-scriptpackages', '/Users/shlens/Desktop/Neural-Art/homebrew/Cellar/python/2.7.13/Frameworks/Python.framework/Versions/2.7/lib/python2.7/lib-tk', '/Users/shlens/Desktop/Neural-Art/homebrew/Cellar/python/2.7.13/Frameworks/Python.framework/Versions/2.7/lib/python2.7/lib-old', '/Users/shlens/Desktop/Neural-Art/homebrew/Cellar/python/2.7.13/Frameworks/Python.framework/Versions/2.7/lib/python2.7/lib-dynload', '/Users/shlens/Desktop/Neural-Art/homebrew/lib/python2.7/site-packages', '/Users/shlens/Desktop/Neural-Art/homebrew/lib/python2.7/site-packages/gtk-2.0', '/Users/shlens/Desktop/Neural-Art/homebrew/lib/python2.7/site-packages/gtk-2.0'])


import numpy as np
import PIL.Image

import tensorflow as tf

FLAGS = tf.app.flags.FLAGS


tf.app.flags.DEFINE_string('data_dir',
                           '/tmp/inception/',
                           'Directory for storing Inception network.')

tf.app.flags.DEFINE_string('jpeg_file',
                           'output.jpg',
                           'Where to save the resulting JPEG.')


def get_layer(layer):
  """Helper for getting layer output Tensor in model Graph.

  Args:
   layer: string, layer name

  Returns:
    Tensor for that layer.
  """
  graph = tf.get_default_graph()
  return graph.get_tensor_by_name('import/%s:0' % layer)


def maybe_download(data_dir):
  """Maybe download pretrained Inception network.

  Args:
    data_dir: string, path to data
  """
  url = ('https://storage.googleapis.com/download.tensorflow.org/models/'
         'inception5h.zip')
  basename = 'inception5h.zip'
  local_file = tf.contrib.learn.python.learn.datasets.base.maybe_download(
      basename, data_dir, url)

  # Uncompress the pretrained Inception network.
  print('Extracting', local_file)
  zip_ref = zipfile.ZipFile(local_file, 'r')
  zip_ref.extractall(FLAGS.data_dir)
  zip_ref.close()


def normalize_image(image):
  """Stretch the range and prepare the image for saving as a JPEG.

  Args:
    image: numpy array

  Returns:
    numpy array of image in uint8
  """
  # Clip to [0, 1] and then convert to uint8.
  image = np.clip(image, 0, 1)
  image = np.uint8(image * 255)
  return image


def save_jpeg(jpeg_file, image):
  pil_image = PIL.Image.fromarray(image)
  pil_image.save(jpeg_file)
  print('Saved to file: ', jpeg_file)


def main(unused_argv):
  # Maybe download and uncompress pretrained Inception network.
  maybe_download(FLAGS.data_dir)

  model_fn = os.path.join(FLAGS.data_dir, 'tensorflow_inception_graph.pb')

  # Load the pretrained Inception model as a GraphDef.
  with tf.gfile.FastGFile(model_fn, 'rb') as f:
    graph_def = tf.GraphDef()
    graph_def.ParseFromString(f.read())

  with tf.Graph().as_default():
    # Input for the network.
    input_image = tf.placeholder(np.float32, name='input')
    pixel_mean = 117.0
    input_preprocessed = tf.expand_dims(input_image - pixel_mean, 0)
    tf.import_graph_def(graph_def, {'input': input_preprocessed})

    # Grab a list of the names of Tensor's that are the output of convolutions.
    graph = tf.get_default_graph()
    layers = [op.name for op in graph.get_operations()
              if op.type == 'Conv2D' and 'import/' in op.name]
    feature_nums = [int(graph.get_tensor_by_name(name+':0').get_shape()[-1])
                    for name in layers]
    # print('Layers available: %s' % ','.join(layers))
    print('Number of layers', len(layers))
    print('Number of features:', sum(feature_nums))

    # Pick an internal layer and node to visualize.
    # Note that we use outputs before applying the ReLU nonlinearity to
    # have non-zero gradients for features with negative initial activations.
    layer = 'mixed4d_3x3_bottleneck_pre_relu'
    channel = 139
    layer_channel = get_layer(layer)[:, :, :, channel]
    print('layer %s, channel %d: %s' % (layer, channel, layer_channel))

    # Define the optimization as the average across all spatial locations.
    score = tf.reduce_mean(layer_channel)

    # Automatic differentiation with TensorFlow. Magic!
    input_gradient = tf.gradients(score, input_image)[0]

    # Employ random noise as a image.
    noise_image = np.random.uniform(size=(224, 224, 3)) + 100.0
    image = noise_image.copy()
    
    ################################################################
    ### BEGIN SOLUTION #####
    ################################################################
    step_scale = 1.0
    num_iter = 20
    with tf.Session() as sess:
      for i in xrange(num_iter):
        image_gradient, score_value = sess.run([input_gradient, score], {input_image:image})
        # Normalize the gradient, so the same step size should work 
        image_gradient /= image_gradient.std() + 1e-8 
        image += image_gradient * step_scale
        print('At step = %d, score = %.3f' % (i, score_value))

  # Save the image.
  stddev = 0.1
  image = (image - image.mean()) / max(image.std(), 1e-4) * stddev + 0.5
  image = normalize_image(image)
  save_jpeg(FLAGS.jpeg_file, image)
  ##################################################################
  ### END SOLUTION #####
  ##################################################################


if __name__ == '__main__':
  tf.app.run()