new_neural_style_transfer.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "import os\n",
    "import numpy as np\n",
    "from scipy.misc import imsave\n",
    "import keras.backend as K\n",
    "from scipy.optimize import fmin_l_bfgs_b\n",
    "\n",
    "def limit_mem():\n",
    "    cfg = K.tf.ConfigProto()\n",
    "    cfg.gpu_options.allow_growth = True\n",
    "    K.set_session(K.tf.Session(config=cfg))\n",
    "limit_mem()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "data_path = '/home/bfortuner/workplace/data/'\n",
    "images_path = data_path+'imagenet_sample/'\n",
    "styles_path = data_path+'styles/'\n",
    "results_path = data_path+'neural_style_results/'\n",
    "style_img_path = styles_path+\"starrynight.jpg\""
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Get Images"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "from PIL import Image\n",
    "from keras.preprocessing.image import load_img, img_to_array\n",
    "import random\n",
    "\n",
    "def get_paths_to_files(dir_path):\n",
    "    filepaths = []\n",
    "    for (dirpath, dirnames, filenames) in os.walk(dir_path):\n",
    "        filepaths.extend(os.path.join(dirpath, f) for f in filenames)\n",
    "    return filepaths\n",
    "\n",
    "def get_random_image_path(dir_path):\n",
    "    filepaths = get_paths_to_files(dir_path) \n",
    "    return filepaths[random.randrange(len(filepaths))] \n",
    "\n",
    "def get_random_image(dir_path):\n",
    "    filepaths = get_paths_to_files(dir_path) \n",
    "    return load_img(filepaths[random.randrange(len(filepaths))]) \n",
    "\n",
    "def get_volume_dims_as_integers(volume):\n",
    "    integer_dims = []\n",
    "    for dim in volume.shape.dims:\n",
    "        integer_dims.append(dim.value)\n",
    "    return integer_dims\n",
    "\n",
    "get_random_image(images_path)\n",
    "load_img(style_img_path)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Initialize VGG"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "from keras.applications import vgg16\n",
    "\n",
    "def get_vgg(input_tensor):\n",
    "    return vgg16.VGG16(input_tensor=input_tensor,\n",
    "            weights='imagenet', include_top=False)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Hyperparameters"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "content_loss_weight = .5\n",
    "style_loss_weight = .5\n",
    "#total_variation_weight?"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Preprocess Images\n",
    "\n",
    "VGG authors preprocessed their images. We subtract the mean of each channel of the imagenet data and reverse the order of RGB->BGR since those are the preprocessing steps that the VGG authors did - so their model won't work unless we do the same thing.\n",
    "\n",
    "Keras VGG has a nice preprocess step for us, but we need to implement the deprocess."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "content_img_path = get_random_image_path(images_path)\n",
    "style_img = load_img(style_img_path)\n",
    "\n",
    "print (content_img_path)\n",
    "print(style_img.size)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "\n",
    "def get_output_image_dims(image_path):\n",
    "    width, height = load_img(image_path).size\n",
    "    img_nrows = 400\n",
    "    img_ncols = int(width * img_nrows / height)\n",
    "    return img_nrows, img_ncols\n",
    "\n",
    "def preprocess_image(image_path, target_rows, target_cols):\n",
    "    img = load_img(image_path, target_size=(target_rows, target_cols))\n",
    "    img = img_to_array(img)\n",
    "    img = np.expand_dims(img, axis=0)\n",
    "    img = vgg16.preprocess_input(img)\n",
    "    return img\n",
    "\n",
    "def deprocess_image(img_vol, target_rows, target_cols):\n",
    "    img_vol = img_vol.reshape((target_rows, target_cols, 3))\n",
    "    # Remove zero-center by mean pixel\n",
    "    img_vol[:, :, 0] += 103.939\n",
    "    img_vol[:, :, 1] += 116.779\n",
    "    img_vol[:, :, 2] += 123.68\n",
    "    # 'BGR'->'RGB'\n",
    "    img_vol = img_vol[:, :, ::-1]\n",
    "    img_vol = np.clip(img_vol, 0, 255).astype('uint8')\n",
    "    return img_vol\n",
    "\n",
    "def get_image_volume(image_path, target_rows, target_cols):\n",
    "    img_arr = preprocess_image(image_path, target_rows, target_cols)\n",
    "    return K.variable(img_arr) #Tensor\n",
    "\n",
    "    \n",
    "# def plot_image_arr(arr):\n",
    "#     plt.imshow(deprocess_image(arr,arr.shape[0],arr.shape[1]))\n",
    "\n",
    "\n",
    "target_rows,target_cols = get_output_image_dims(content_img_path)\n",
    "print(target_rows,target_cols)\n",
    "\n",
    "img = load_img(content_img_path, target_size=(target_rows, target_cols))\n",
    "img_a = img_to_array(img)\n",
    "print (img_a)\n",
    "\n",
    "content_img_vol = preprocess_image(content_img_path, target_rows, target_cols)\n",
    "print (content_img_vol)\n",
    "\n",
    "output_image = deprocess_image(content_img_vol, target_rows, target_cols)\n",
    "print (output_image)\n",
    "print (output_image.shape)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Get Image, Style, Noise Volume"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "content_img_path = get_random_image_path(images_path)\n",
    "style_img = load_img(style_img_path)\n",
    "\n",
    "img_nrows,img_ncols = get_output_image_dims(content_img_path)\n",
    "\n",
    "content_img_vol = get_image_volume(content_img_path, img_nrows, img_ncols)\n",
    "style_img_vol = get_image_volume(style_img_path, img_nrows, img_ncols)\n",
    "output_img_vol = K.placeholder((1, img_nrows, img_ncols, 3))\n",
    "print (output_img_vol.shape)\n",
    "dimensions = output_img_vol.shape.dims\n",
    "type(dimensions[1])\n",
    "\n",
    "\n",
    "from tensorflow import Dimension #tensorflow.python.framework.tensor_shape.Dimension\n",
    "d = Dimension(2)\n",
    "d.value"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "# combine the 3 images into a single Keras tensor\n",
    "input_tensor = K.concatenate([content_img_vol,\n",
    "                              style_img_vol,\n",
    "                              output_img_vol], axis=0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "vgg_model = get_vgg(input_tensor)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "# get the symbolic outputs of each \"key\" layer (we gave them unique names).\n",
    "layers_dict = dict([(layer.name, layer.output) for layer in vgg_model.layers])\n",
    "layers_dict"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Gram Matrix "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "def gram_matrix(x):\n",
    "    assert K.ndim(x) == 3\n",
    "    features = K.batch_flatten(K.permute_dimensions(x, (2, 0, 1)))\n",
    "    gram = K.dot(features, K.transpose(features))\n",
    "    return gram"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Style Loss"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "conv_layer_names = ['block1_conv1', 'block2_conv1',\n",
    "              'block3_conv1', 'block4_conv1',\n",
    "              'block5_conv1']\n",
    "    \n",
    "def style_loss(style_vol, output_vol, img_nrows, img_ncols):\n",
    "    assert K.ndim(style_vol) == 3\n",
    "    assert K.ndim(output_vol) == 3\n",
    "    S = gram_matrix(style_vol)\n",
    "    C = gram_matrix(output_vol)\n",
    "    channels = 3\n",
    "    img_dims = get_volume_dims_as_integers(output_vol)\n",
    "    img_nrows,img_ncols = img_dims[1],img_dims[2]\n",
    "    size = img_nrows * img_ncols\n",
    "    return K.sum(K.square(S - C)) / (4. * (channels ** 2) * (size ** 2))\n",
    "\n",
    "def get_style_loss(conv_layer_names, img_nrows, img_ncols, style_weight=1.0):\n",
    "    loss = K.variable(0.)\n",
    "    for layer_name in conv_layer_names:\n",
    "        layer = layers_dict[layer_name]\n",
    "        style_activation = layer[1, :, :, :]\n",
    "        output_activation = layer[2, :, :, :]\n",
    "        sl = style_loss(style_activation, output_activation, img_nrows, img_ncols)\n",
    "        loss += (style_weight / len(conv_layer_names)) * sl\n",
    "    return loss\n",
    "\n",
    "get_style_loss(conv_layer_names, 400, 400)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Content Loss"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "def content_loss(content_vol, output_vol):\n",
    "    return K.sum(K.square(output_vol - content_vol))\n",
    "\n",
    "def get_content_loss(layer_activations, content_weight=.025):\n",
    "    content_activation = layer_activations[0, :, :, :]\n",
    "    output_activation = layer_activations[2, :, :, :]\n",
    "    return content_weight * content_loss(content_activation,\n",
    "                                      output_activation)\n",
    "\n",
    "loss = K.variable(0.)\n",
    "layer_activations = layers_dict['block4_conv2']\n",
    "get_content_loss(layer_activations)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Total Variation Loss"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "def total_variation_loss(x, img_nrows, img_ncols, variation_weight=1.0):\n",
    "    assert K.ndim(x) == 4\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)) * variation_weight"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Total Loss (Combined)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "img_nrows, img_ncols = get_output_image_dims(content_img_path)\n",
    "content_img_vol = get_image_volume(content_img_path, img_nrows, img_ncols)\n",
    "style_img_vol = get_image_volume(style_img_path, img_nrows, img_ncols)\n",
    "output_img_vol = K.placeholder((1, img_nrows, img_ncols, 3))\n",
    "\n",
    "output_layer_activations = layers_dict['block4_conv2']\n",
    "content_loss = get_content_loss(output_layer_activations)\n",
    "style_loss = get_style_loss(conv_layer_names, img_nrows, img_ncols)\n",
    "variation_loss = total_variation_loss(output_img_vol, img_nrows, img_ncols)\n",
    "\n",
    "total_loss = style_loss + content_loss + variation_loss\n",
    "print (total_loss.value)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Gradient Descent"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "grads = K.gradients(total_loss, output_img_vol)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "grads"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "outputs = [loss]\n",
    "if isinstance(grads, (list, tuple)):\n",
    "    outputs += grads\n",
    "else:\n",
    "    outputs.append(grads)\n",
    "\n",
    "f_outputs = K.function([output_img_vol], outputs)\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "def eval_loss_and_grads(x):\n",
    "    x = x.reshape((1, img_nrows, img_ncols, 3))\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"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Scipy Evaluator "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "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",
    "\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\n",
    "\n",
    "evaluator = Evaluator()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Train"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "import time\n",
    "# run scipy-based optimization (L-BFGS) over the pixels of the generated image\n",
    "# so as to minimize the neural style loss\n",
    "x = np.random.uniform(0, 255, (1, img_nrows, img_ncols, 3)) - 128.\n",
    "iterations = 1\n",
    "for i in range(iterations):\n",
    "    print('Start of iteration', i)\n",
    "    start_time = time.time()\n",
    "    x, min_val, info = fmin_l_bfgs_b(evaluator.loss, x.flatten(),\n",
    "                                     fprime=evaluator.grads, maxfun=20)\n",
    "    print('Current loss value:', min_val)\n",
    "    # save current generated image\n",
    "    img = deprocess_image(x.copy())\n",
    "    fname = result_prefix + '_at_iteration_%d.png' % i\n",
    "    imsave(fname, img)\n",
    "    end_time = time.time()\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": {
    "collapsed": true
   },
   "outputs": [],
   "source": []
  }
 ],
 "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": "3.6.0"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}