examples/style_transfer_tensorflow/style_transfer_keras.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "DSC160 Data Science and the Arts - Twomey - Spring 2020 - [dsc160.roberttwomey.com](http://dsc160.roberttwomey.com)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Style Transfer in Keras\n",
    "\n",
    "Using CNNs and Gram Matrix.\n",
    "\n",
    "Paper: https://www.cv-foundation.org/openaccess/content_cvpr_2016/papers/Gatys_Image_Style_Transfer_CVPR_2016_paper.pdf"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import scipy.io\n",
    "import scipy.misc\n",
    "import tensorflow as tf\n",
    "import numpy as np\n",
    "import time\n",
    "import imageio\n",
    "from IPython.display import Image\n",
    "\n",
    "from tensorflow.keras.preprocessing.image import load_img, img_to_array\n",
    "from scipy.misc import imsave #to be able to save as an image\n",
    "from scipy.optimize import fmin_l_bfgs_b #to optimize our cost function\n",
    "\n",
    "from tensorflow.keras.applications import vgg19 #to load vgg19 network\n",
    "from tensorflow.keras import backend as K"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import tensorflow.keras\n",
    "print(tf.__version__)\n",
    "print(tensorflow.keras.__version__)\n",
    "\n",
    "from tensorflow.keras.utils import plot_model #to be able to visualize the network\n",
    "%matplotlib inline\n",
    "import matplotlib.image as mpimg\n",
    "import matplotlib.pyplot as plt"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "content_file_name = '1-content.jpg'#content image\n",
    "style_file_name = '1-style.jpg'#style image"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "iterations =20 # we will run 20 iterations during the optimization. \n",
    "# each iteration you get better value. this one is about 15 min. \n",
    "\n",
    "# these are the weights of the different loss components\n",
    "total_variation_weight = 1.0 \n",
    "style_weight = 1.0 #the best number authors got from their trials \n",
    "content_weight = 0.025 #the best number authors got from their trials "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# let's load images\n",
    "\n",
    "base_image_path = './images/'+content_file_name\n",
    "style_reference_image_path = './images/'+style_file_name\n",
    "\n",
    "# get dimensions (width, height) of the generated picture\n",
    "width, height = load_img(base_image_path).size\n",
    "\n",
    "img_nrows = 400 #recale the image to 400 pixel rows \n",
    "img_ncols = int(width * img_nrows / height)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# pre-process the image: rescaling, running it through VGG19\n",
    "\n",
    "def preprocess_image(image_path):\n",
    "\n",
    "    img = load_img(image_path, target_size=(img_nrows, img_ncols))\n",
    "    img = img_to_array(img)\n",
    "    img = np.expand_dims(img, axis=0)\n",
    "    img = vgg19.preprocess_input(img)\n",
    "    return img\n",
    "\n",
    "\n",
    "# utility function to convert a tensor into a valid image\n",
    "\n",
    "def deprocess_image(x):\n",
    "\n",
    "    if K.image_data_format() == 'channels_first':\n",
    "        x = x.reshape((3, img_nrows, img_ncols))\n",
    "        x = x.transpose((1, 2, 0))\n",
    "\n",
    "    else:\n",
    "\n",
    "        x = x.reshape((img_nrows, img_ncols, 3))\n",
    "\n",
    "    # Remove zero-center by mean pixel\n",
    "\n",
    "    x[:, :, 0] += 103.939\n",
    "    x[:, :, 1] += 116.779\n",
    "    x[:, :, 2] += 123.68\n",
    "\n",
    "    # 'BGR'->'RGB'\n",
    "\n",
    "    x = x[:, :, ::-1]\n",
    "    x = np.clip(x, 0, 255).astype('uint8')\n",
    "    return x\n",
    "\n",
    "#calculating the mean pixel values of the VGG network. We will use these values for normalization.\n",
    "#mean = data['meta'][0][0][2][0][0][2]\n",
    "#mean_pixel = np.mean(mean, axis=(0, 1))\n",
    "#print(mean_pixel)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "image_content = imageio.imread(base_image_path)\n",
    "#image_content= scipy.misc.imresize(image_content,(224,224))\n",
    "plt.figure(figsize=(11,11))\n",
    "plt.imshow(image_content)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "image_style = imageio.imread(style_reference_image_path)\n",
    "#image_content= scipy.misc.imresize(image_content,(224,224))\n",
    "plt.figure(figsize=(11,11))\n",
    "plt.imshow(image_style)\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# get tensor representations of our images\n",
    "\n",
    "base_image = K.variable(preprocess_image(base_image_path))\n",
    "style_reference_image = K.variable(preprocess_image(style_reference_image_path))\n",
    "\n",
    "# this will contain our generated image\n",
    "if K.image_data_format() == 'channels_first':\n",
    "    combination_image = K.placeholder((1, 3, img_nrows, img_ncols))\n",
    "else:\n",
    "    combination_image = K.placeholder((1, img_nrows, img_ncols, 3))\n",
    "\n",
    "\n",
    "# combine the 3 images (style, content, result image that starts from the white noise) into a single Keras tensor\n",
    "\n",
    "input_tensor = K.concatenate([base_image,\n",
    "                              style_reference_image,\n",
    "                              combination_image], axis=0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# build the VGG19 network with our 3 images as input\n",
    "# the model will be loaded with pre-trained ImageNet weights\n",
    "\n",
    "model = vgg19.VGG19(input_tensor=input_tensor,weights='imagenet', include_top=False)\n",
    "print('Model loaded.')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# plot_model(model, show_shapes=True, to_file='./images/model.png')\n",
    "# Image(\"./images/model.png\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# get the symbolic outputs of each \"key\" layer (we gave them unique names).\n",
    "outputs_dict = dict([(layer.name, layer.output) for layer in model.layers])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "for l, layer in enumerate(model.layers):\n",
    "    print(layer.name)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# compute the neural style loss\n",
    "# first we need to define 4 utility functions\n",
    "# the gram matrix of an image tensor (feature-wise outer product)\n",
    "\n",
    "def gram_matrix(x):\n",
    "\n",
    "    assert K.ndim(x) == 3\n",
    "    if K.image_data_format() == 'channels_first':\n",
    "        features = K.batch_flatten(x)\n",
    "    else:\n",
    "        features = K.batch_flatten(K.permute_dimensions(x, (2, 0, 1)))\n",
    "\n",
    "    gram = K.dot(features, K.transpose(features))\n",
    "    return gram"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# the \"style loss\" is designed to maintain\n",
    "# the style of the reference image in the generated image.\n",
    "# It is based on the gram matrices (which capture style) of\n",
    "# feature maps from the style reference image\n",
    "# and from the generated image\n",
    "\n",
    "def style_loss(style, combination):\n",
    "\n",
    "    assert K.ndim(style) == 3\n",
    "    assert K.ndim(combination) == 3\n",
    "\n",
    "    S = gram_matrix(style)\n",
    "    C = gram_matrix(combination)\n",
    "    channels = 3\n",
    "    size = img_nrows * img_ncols\n",
    "    return K.sum(K.square(S - C)) / (4. * (channels ** 2) * (size ** 2))\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# an auxiliary loss function\n",
    "# designed to maintain the \"content\" of the\n",
    "# base image in the generated image\n",
    "\n",
    "def content_loss(base, combination):\n",
    "    return K.sum(K.square(combination - base))\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# the 3rd loss function, total variation loss,\n",
    "# designed to keep the generated image locally coherent (no big changes)\n",
    "\n",
    "def total_variation_loss(x):\n",
    "    assert K.ndim(x) == 4\n",
    "    if K.image_data_format() == 'channels_first':\n",
    "\n",
    "        a = K.square(x[:, :, :img_nrows - 1, :img_ncols - 1] - x[:, :, 1:, :img_ncols - 1])\n",
    "        b = K.square(x[:, :, :img_nrows - 1, :img_ncols - 1] - x[:, :, :img_nrows - 1, 1:])\n",
    "    else:\n",
    "        a = K.square(x[:, :img_nrows - 1, :img_ncols - 1, :] - x[:, 1:, :img_ncols - 1, :])\n",
    "        b = K.square(x[:, :img_nrows - 1, :img_ncols - 1, :] - x[:, :img_nrows - 1, 1:, :])\n",
    "    return K.sum(K.pow(a + b, 1.25))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# combine these loss functions into a single scalar\n",
    "\n",
    "loss = K.variable(0.)\n",
    "layer_features = outputs_dict['block5_conv2'] #content features\n",
    "base_image_features = layer_features[0, :, :, :]\n",
    "combination_features = layer_features[2, :, :, :]\n",
    "loss = loss + content_weight * content_loss(base_image_features,\n",
    "                                      combination_features)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# style_features\n",
    "feature_layers = ['block1_conv1', 'block2_conv1',\n",
    "                  'block3_conv1', 'block4_conv1',\n",
    "                  'block5_conv1']"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "for layer_name in feature_layers:\n",
    "    layer_features = outputs_dict[layer_name]\n",
    "    style_reference_features = layer_features[1, :, :, :]\n",
    "    combination_features = layer_features[2, :, :, :]\n",
    "    sl = style_loss(style_reference_features, combination_features)\n",
    "    loss += (style_weight / len(feature_layers)) * sl\n",
    "\n",
    "loss += total_variation_weight * total_variation_loss(combination_image)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# get the gradients of the generated image wrt the loss\n",
    "\n",
    "grads = K.gradients(loss, combination_image)\n",
    "outputs = [loss]\n",
    "\n",
    "if isinstance(grads, (list, tuple)):\n",
    "    outputs += grads\n",
    "else:\n",
    "    outputs.append(grads)\n",
    "\n",
    "f_outputs = K.function([combination_image], outputs)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def eval_loss_and_grads(x):\n",
    "\n",
    "    if K.image_data_format() == 'channels_first':\n",
    "        x = x.reshape((1, 3, img_nrows, img_ncols))\n",
    "    else:\n",
    "        x = x.reshape((1, img_nrows, img_ncols, 3))\n",
    "\n",
    "    outs = f_outputs([x])\n",
    "    loss_value = outs[0]\n",
    "    if len(outs[1:]) == 1:\n",
    "        grad_values = outs[1].flatten().astype('float64')\n",
    "    else:\n",
    "        grad_values = np.array(outs[1:]).flatten().astype('float64')\n",
    "    return loss_value, grad_values\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# this Evaluator class makes it possible\n",
    "# to compute loss and gradients in one pass\n",
    "# while retrieving them via two separate functions,\n",
    "# \"loss\" and \"grads\". This is done because scipy.optimize\n",
    "# requires separate functions for loss and gradients,\n",
    "# but computing them separately would be inefficient.\n",
    "\n",
    "class Evaluator(object):\n",
    "\n",
    "    def __init__(self):\n",
    "        self.loss_value = None\n",
    "        self.grads_values = None\n",
    "\n",
    "    def loss(self, x):\n",
    "        assert self.loss_value is None\n",
    "        loss_value, grad_values = eval_loss_and_grads(x)\n",
    "        self.loss_value = loss_value\n",
    "        self.grad_values = grad_values\n",
    "        return self.loss_value\n",
    "\n",
    "    def grads(self, x):\n",
    "        assert self.loss_value is not None\n",
    "        grad_values = np.copy(self.grad_values)\n",
    "        self.loss_value = None\n",
    "        self.grad_values = None\n",
    "        return grad_values"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "evaluator = Evaluator()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# run scipy-based optimization (L-BFGS) over the pixels of the generated image\n",
    "# so as to minimize the neural style loss\n",
    "x = preprocess_image(base_image_path)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "result_prefix ='./images/results_'"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# minimise the loss function\n",
    "\n",
    "for i in range(iterations):\n",
    "\n",
    "    print('Start of iteration', i)\n",
    "    start_time = time.time()\n",
    "\n",
    "    x, min_val, info = fmin_l_bfgs_b(evaluator.loss, x.flatten(),\n",
    "                                     fprime=evaluator.grads, maxfun=20)\n",
    "\n",
    "    print('Current loss value:', min_val)\n",
    "    # save current generated image\n",
    "\n",
    "    img = deprocess_image(x.copy())\n",
    "    fname = result_prefix + '_at_iteration_%d.png' % i\n",
    "    imageio.imwrite(fname, img)\n",
    "    end_time = time.time()\n",
    "\n",
    "    print('Image saved as', fname)\n",
    "    print('Iteration %d completed in %ds' % (i, end_time - start_time))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "#image_content= scipy.misc.imresize(image_content,(224,224))\n",
    "plt.figure(figsize=(11,11))\n",
    "plt.imshow(img)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": true
   },
   "source": [
    "## References\n",
    "- Adapted from https://github.com/kangeunsu/ArtML/tree/master/Style_Transfer_Tensorflow\n",
    "- Additional style transfer reading: https://medium.com/artists-and-machine-intelligence/neural-artistic-style-transfer-a-comprehensive-look-f54d8649c199"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.7.3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 1
}