neural_style.py

import tensorflow as tf
import numpy as np 
import scipy.io  
import argparse 
import struct
import errno
import time                       
import cv2
import os

'''
  parsing and configuration
'''
def parse_args():

  desc = "TensorFlow implementation of 'A Neural Algorithm for Artisitc Style'"  
  parser = argparse.ArgumentParser(description=desc)

  # options for single image
  parser.add_argument('--verbose', action='store_true',
    help='Boolean flag indicating if statements should be printed to the console.')

  parser.add_argument('--img_name', type=str, 
    default='result',
    help='Filename of the output image.')

  parser.add_argument('--style_imgs', nargs='+', type=str,
    help='Filenames of the style images (example: starry-night.jpg)', 
    required=True)
  
  parser.add_argument('--style_imgs_weights', nargs='+', type=float,
    default=[1.0],
    help='Interpolation weights of each of the style images. (example: 0.5 0.5)')
  
  parser.add_argument('--content_img', type=str,
    help='Filename of the content image (example: lion.jpg)')

  parser.add_argument('--style_imgs_dir', type=str,
    default='./styles',
    help='Directory path to the style images. (default: %(default)s)')

  parser.add_argument('--content_img_dir', type=str,
    default='./image_input',
    help='Directory path to the content image. (default: %(default)s)')
  
  parser.add_argument('--init_img_type', type=str, 
    default='content',
    choices=['random', 'content', 'style'], 
    help='Image used to initialize the network. (default: %(default)s)')
  
  parser.add_argument('--max_size', type=int, 
    default=512,
    help='Maximum width or height of the input images. (default: %(default)s)')
  
  parser.add_argument('--content_weight', type=float, 
    default=5e0,
    help='Weight for the content loss function. (default: %(default)s)')
  
  parser.add_argument('--style_weight', type=float, 
    default=1e4,
    help='Weight for the style loss function. (default: %(default)s)')
  
  parser.add_argument('--tv_weight', type=float, 
    default=1e-3,
    help='Weight for the total variational loss function. Set small (e.g. 1e-3). (default: %(default)s)')

  parser.add_argument('--temporal_weight', type=float, 
    default=2e2,
    help='Weight for the temporal loss function. (default: %(default)s)')

  parser.add_argument('--content_loss_function', type=int,
    default=1,
    choices=[1, 2, 3],
    help='Different constants for the content layer loss function. (default: %(default)s)')
  
  parser.add_argument('--content_layers', nargs='+', type=str, 
    default=['conv4_2'],
    help='VGG19 layers used for the content image. (default: %(default)s)')
  
  parser.add_argument('--style_layers', nargs='+', type=str,
    default=['relu1_1', 'relu2_1', 'relu3_1', 'relu4_1', 'relu5_1'],
    help='VGG19 layers used for the style image. (default: %(default)s)')
  
  parser.add_argument('--content_layer_weights', nargs='+', type=float, 
    default=[1.0], 
    help='Contributions (weights) of each content layer to loss. (default: %(default)s)')
  
  parser.add_argument('--style_layer_weights', nargs='+', type=float, 
    default=[0.2, 0.2, 0.2, 0.2, 0.2],
    help='Contributions (weights) of each style layer to loss. (default: %(default)s)')
    
  parser.add_argument('--original_colors', action='store_true',
    help='Transfer the style but not the colors.')

  parser.add_argument('--color_convert_type', type=str,
    default='yuv',
    choices=['yuv', 'ycrcb', 'luv', 'lab'],
    help='Color space for conversion to original colors (default: %(default)s)')

  parser.add_argument('--color_convert_time', type=str,
    default='after',
    choices=['after', 'before'],
    help='Time (before or after) to convert to original colors (default: %(default)s)')

  parser.add_argument('--style_mask', action='store_true',
    help='Transfer the style to masked regions.')

  parser.add_argument('--style_mask_imgs', nargs='+', type=str, 
    default=None,
    help='Filenames of the style mask images (example: face_mask.png) (default: %(default)s)')
  
  parser.add_argument('--noise_ratio', type=float, 
    default=1.0, 
    help="Interpolation value between the content image and noise image if the network is initialized with 'random'.")

  parser.add_argument('--seed', type=int, 
    default=0,
    help='Seed for the random number generator. (default: %(default)s)')
  
  parser.add_argument('--model_weights', type=str, 
    default='imagenet-vgg-verydeep-19.mat',
    help='Weights and biases of the VGG-19 network.')
  
  parser.add_argument('--pooling_type', type=str,
    default='avg',
    choices=['avg', 'max'],
    help='Type of pooling in convolutional neural network. (default: %(default)s)')
  
  parser.add_argument('--device', type=str, 
    default='/gpu:0',
    choices=['/gpu:0', '/cpu:0'],
    help='GPU or CPU mode.  GPU mode requires NVIDIA CUDA. (default|recommended: %(default)s)')
  
  parser.add_argument('--img_output_dir', type=str, 
    default='./image_output',
    help='Relative or absolute directory path to output image and data.')
  
  # optimizations
  parser.add_argument('--optimizer', type=str, 
    default='lbfgs',
    choices=['lbfgs', 'adam'],
    help='Loss minimization optimizer.  L-BFGS gives better results.  Adam uses less memory. (default|recommended: %(default)s)')
  
  parser.add_argument('--learning_rate', type=float, 
    default=1e0, 
    help='Learning rate parameter for the Adam optimizer. (default: %(default)s)')
  
  parser.add_argument('--max_iterations', type=int, 
    default=1000,
    help='Max number of iterations for the Adam or L-BFGS optimizer. (default: %(default)s)')

  parser.add_argument('--print_iterations', type=int, 
    default=50,
    help='Number of iterations between optimizer print statements. (default: %(default)s)')
  
  # options for video frames
  parser.add_argument('--video', action='store_true', 
    help='Boolean flag indicating if the user is generating a video.')

  parser.add_argument('--start_frame', type=int, 
    default=1,
    help='First frame number.')
  
  parser.add_argument('--end_frame', type=int, 
    default=1,
    help='Last frame number.')
  
  parser.add_argument('--first_frame_type', type=str,
    choices=['random', 'content', 'style'], 
    default='content',
    help='Image used to initialize the network during the rendering of the first frame.')
  
  parser.add_argument('--init_frame_type', type=str, 
    choices=['prev_warped', 'prev', 'random', 'content', 'style'], 
    default='prev_warped',
    help='Image used to initialize the network during the every rendering after the first frame.')
  
  parser.add_argument('--video_input_dir', type=str, 
    default='./video_input',
    help='Relative or absolute directory path to input frames.')
  
  parser.add_argument('--video_output_dir', type=str, 
    default='./video_output',
    help='Relative or absolute directory path to output frames.')
  
  parser.add_argument('--content_frame_frmt', type=str, 
    default='frame_{}.ppm',
    help='Filename format of the input content frames.')
  
  parser.add_argument('--backward_optical_flow_frmt', type=str, 
    default='backward_{}_{}.flo',
    help='Filename format of the backward optical flow files.')
  
  parser.add_argument('--forward_optical_flow_frmt', type=str, 
    default='forward_{}_{}.flo',
    help='Filename format of the forward optical flow files')
  
  parser.add_argument('--content_weights_frmt', type=str, 
    default='reliable_{}_{}.txt',
    help='Filename format of the optical flow consistency files.')
  
  parser.add_argument('--prev_frame_indices', nargs='+', type=int, 
    default=[1],
    help='Previous frames to consider for longterm temporal consistency.')

  parser.add_argument('--first_frame_iterations', type=int, 
    default=2000,
    help='Maximum number of optimizer iterations of the first frame. (default: %(default)s)')
  
  parser.add_argument('--frame_iterations', type=int, 
    default=800,
    help='Maximum number of optimizer iterations for each frame after the first frame. (default: %(default)s)')

  args = parser.parse_args()

  # normalize weights
  args.style_layer_weights   = normalize(args.style_layer_weights)
  args.content_layer_weights = normalize(args.content_layer_weights)
  args.style_imgs_weights    = normalize(args.style_imgs_weights)

  # create directories for output
  if args.video:
    maybe_make_directory(args.video_output_dir)
  else:
    maybe_make_directory(args.img_output_dir)

  return args

'''
  pre-trained vgg19 convolutional neural network
  remark: layers are manually initialized for clarity.
'''

def build_model(input_img):
  if args.verbose: print('\nBUILDING VGG-19 NETWORK')
  net = {}
  _, h, w, d     = input_img.shape
  
  if args.verbose: print('loading model weights...')
  vgg_rawnet     = scipy.io.loadmat(args.model_weights)
  vgg_layers     = vgg_rawnet['layers'][0]
  if args.verbose: print('constructing layers...')
  net['input']   = tf.Variable(np.zeros((1, h, w, d), dtype=np.float32))

  if args.verbose: print('LAYER GROUP 1')
  net['conv1_1'] = conv_layer('conv1_1', net['input'], W=get_weights(vgg_layers, 0))
  net['relu1_1'] = relu_layer('relu1_1', net['conv1_1'], b=get_bias(vgg_layers, 0))

  net['conv1_2'] = conv_layer('conv1_2', net['relu1_1'], W=get_weights(vgg_layers, 2))
  net['relu1_2'] = relu_layer('relu1_2', net['conv1_2'], b=get_bias(vgg_layers, 2))
  
  net['pool1']   = pool_layer('pool1', net['relu1_2'])

  if args.verbose: print('LAYER GROUP 2')  
  net['conv2_1'] = conv_layer('conv2_1', net['pool1'], W=get_weights(vgg_layers, 5))
  net['relu2_1'] = relu_layer('relu2_1', net['conv2_1'], b=get_bias(vgg_layers, 5))
  
  net['conv2_2'] = conv_layer('conv2_2', net['relu2_1'], W=get_weights(vgg_layers, 7))
  net['relu2_2'] = relu_layer('relu2_2', net['conv2_2'], b=get_bias(vgg_layers, 7))
  
  net['pool2']   = pool_layer('pool2', net['relu2_2'])
  
  if args.verbose: print('LAYER GROUP 3')
  net['conv3_1'] = conv_layer('conv3_1', net['pool2'], W=get_weights(vgg_layers, 10))
  net['relu3_1'] = relu_layer('relu3_1', net['conv3_1'], b=get_bias(vgg_layers, 10))

  net['conv3_2'] = conv_layer('conv3_2', net['relu3_1'], W=get_weights(vgg_layers, 12))
  net['relu3_2'] = relu_layer('relu3_2', net['conv3_2'], b=get_bias(vgg_layers, 12))

  net['conv3_3'] = conv_layer('conv3_3', net['relu3_2'], W=get_weights(vgg_layers, 14))
  net['relu3_3'] = relu_layer('relu3_3', net['conv3_3'], b=get_bias(vgg_layers, 14))

  net['conv3_4'] = conv_layer('conv3_4', net['relu3_3'], W=get_weights(vgg_layers, 16))
  net['relu3_4'] = relu_layer('relu3_4', net['conv3_4'], b=get_bias(vgg_layers, 16))

  net['pool3']   = pool_layer('pool3', net['relu3_4'])

  if args.verbose: print('LAYER GROUP 4')
  net['conv4_1'] = conv_layer('conv4_1', net['pool3'], W=get_weights(vgg_layers, 19))
  net['relu4_1'] = relu_layer('relu4_1', net['conv4_1'], b=get_bias(vgg_layers, 19))

  net['conv4_2'] = conv_layer('conv4_2', net['relu4_1'], W=get_weights(vgg_layers, 21))
  net['relu4_2'] = relu_layer('relu4_2', net['conv4_2'], b=get_bias(vgg_layers, 21))

  net['conv4_3'] = conv_layer('conv4_3', net['relu4_2'], W=get_weights(vgg_layers, 23))
  net['relu4_3'] = relu_layer('relu4_3', net['conv4_3'], b=get_bias(vgg_layers, 23))

  net['conv4_4'] = conv_layer('conv4_4', net['relu4_3'], W=get_weights(vgg_layers, 25))
  net['relu4_4'] = relu_layer('relu4_4', net['conv4_4'], b=get_bias(vgg_layers, 25))

  net['pool4']   = pool_layer('pool4', net['relu4_4'])

  if args.verbose: print('LAYER GROUP 5')
  net['conv5_1'] = conv_layer('conv5_1', net['pool4'], W=get_weights(vgg_layers, 28))
  net['relu5_1'] = relu_layer('relu5_1', net['conv5_1'], b=get_bias(vgg_layers, 28))

  net['conv5_2'] = conv_layer('conv5_2', net['relu5_1'], W=get_weights(vgg_layers, 30))
  net['relu5_2'] = relu_layer('relu5_2', net['conv5_2'], b=get_bias(vgg_layers, 30))

  net['conv5_3'] = conv_layer('conv5_3', net['relu5_2'], W=get_weights(vgg_layers, 32))
  net['relu5_3'] = relu_layer('relu5_3', net['conv5_3'], b=get_bias(vgg_layers, 32))

  net['conv5_4'] = conv_layer('conv5_4', net['relu5_3'], W=get_weights(vgg_layers, 34))
  net['relu5_4'] = relu_layer('relu5_4', net['conv5_4'], b=get_bias(vgg_layers, 34))

  net['pool5']   = pool_layer('pool5', net['relu5_4'])

  return net

def conv_layer(layer_name, layer_input, W):
  conv = tf.nn.conv2d(layer_input, W, strides=[1, 1, 1, 1], padding='SAME')
  if args.verbose: print('--{} | shape={} | weights_shape={}'.format(layer_name, 
    conv.get_shape(), W.get_shape()))
  return conv

def relu_layer(layer_name, layer_input, b):
  relu = tf.nn.relu(layer_input + b)
  if args.verbose: 
    print('--{} | shape={} | bias_shape={}'.format(layer_name, relu.get_shape(), 
      b.get_shape()))
  return relu

def pool_layer(layer_name, layer_input):
  if args.pooling_type == 'avg':
    pool = tf.nn.avg_pool(layer_input, ksize=[1, 2, 2, 1], 
      strides=[1, 2, 2, 1], padding='SAME')
  elif args.pooling_type == 'max':
    pool = tf.nn.max_pool(layer_input, ksize=[1, 2, 2, 1], 
      strides=[1, 2, 2, 1], padding='SAME')
  if args.verbose: 
    print('--{}   | shape={}'.format(layer_name, pool.get_shape()))
  return pool

def get_weights(vgg_layers, i):
  weights = vgg_layers[i][0][0][2][0][0]
  W = tf.constant(weights)
  return W

def get_bias(vgg_layers, i):
  bias = vgg_layers[i][0][0][2][0][1]
  b = tf.constant(np.reshape(bias, (bias.size)))
  return b

'''
  'a neural algorithm for artistic style' loss functions
'''
def content_layer_loss(p, x):
  _, h, w, d = p.get_shape()
  M = h.value * w.value
  N = d.value
  if args.content_loss_function   == 1:
    K = 1. / (2. * N**0.5 * M**0.5)
  elif args.content_loss_function == 2:
    K = 1. / (N * M)
  elif args.content_loss_function == 3:  
    K = 1. / 2.
  loss = K * tf.reduce_sum(tf.pow((x - p), 2))
  return loss

def style_layer_loss(a, x):
  _, h, w, d = a.get_shape()
  M = h.value * w.value
  N = d.value
  A = gram_matrix(a, M, N)
  G = gram_matrix(x, M, N)
  loss = (1./(4 * N**2 * M**2)) * tf.reduce_sum(tf.pow((G - A), 2))
  return loss

def gram_matrix(x, area, depth):
  F = tf.reshape(x, (area, depth))
  G = tf.matmul(tf.transpose(F), F)
  return G

def mask_style_layer(a, x, mask_img):
  _, h, w, d = a.get_shape()
  mask = get_mask_image(mask_img, w.value, h.value)
  mask = tf.convert_to_tensor(mask)
  tensors = []
  for _ in range(d.value): 
    tensors.append(mask)
  mask = tf.stack(tensors, axis=2)
  mask = tf.stack(mask, axis=0)
  mask = tf.expand_dims(mask, 0)
  a = tf.multiply(a, mask)
  x = tf.multiply(x, mask)
  return a, x

def sum_masked_style_losses(sess, net, style_imgs):
  total_style_loss = 0.
  weights = args.style_imgs_weights
  masks = args.style_mask_imgs
  for img, img_weight, img_mask in zip(style_imgs, weights, masks):
    sess.run(net['input'].assign(img))
    style_loss = 0.
    for layer, weight in zip(args.style_layers, args.style_layer_weights):
      a = sess.run(net[layer])
      x = net[layer]
      a = tf.convert_to_tensor(a)
      a, x = mask_style_layer(a, x, img_mask)
      style_loss += style_layer_loss(a, x) * weight
    style_loss /= float(len(args.style_layers))
    total_style_loss += (style_loss * img_weight)
  total_style_loss /= float(len(style_imgs))
  return total_style_loss

def sum_style_losses(sess, net, style_imgs):
  total_style_loss = 0.
  weights = args.style_imgs_weights
  for img, img_weight in zip(style_imgs, weights):
    sess.run(net['input'].assign(img))
    style_loss = 0.
    for layer, weight in zip(args.style_layers, args.style_layer_weights):
      a = sess.run(net[layer])
      x = net[layer]
      a = tf.convert_to_tensor(a)
      style_loss += style_layer_loss(a, x) * weight
    style_loss /= float(len(args.style_layers))
    total_style_loss += (style_loss * img_weight)
  total_style_loss /= float(len(style_imgs))
  return total_style_loss

def sum_content_losses(sess, net, content_img):
  sess.run(net['input'].assign(content_img))
  content_loss = 0.
  for layer, weight in zip(args.content_layers, args.content_layer_weights):
    p = sess.run(net[layer])
    x = net[layer]
    p = tf.convert_to_tensor(p)
    content_loss += content_layer_loss(p, x) * weight
  content_loss /= float(len(args.content_layers))
  return content_loss

'''
  'artistic style transfer for videos' loss functions
'''
def temporal_loss(x, w, c):
  c = c[np.newaxis,:,:,:]
  D = float(x.size)
  loss = (1. / D) * tf.reduce_sum(c * tf.nn.l2_loss(x - w))
  loss = tf.cast(loss, tf.float32)
  return loss

def get_longterm_weights(i, j):
  c_sum = 0.
  for k in range(args.prev_frame_indices):
    if i - k > i - j:
      c_sum += get_content_weights(i, i - k)
  c = get_content_weights(i, i - j)
  c_max = tf.maximum(c - c_sum, 0.)
  return c_max

def sum_longterm_temporal_losses(sess, net, frame, input_img):
  x = sess.run(net['input'].assign(input_img))
  loss = 0.
  for j in range(args.prev_frame_indices):
    prev_frame = frame - j
    w = get_prev_warped_frame(frame)
    c = get_longterm_weights(frame, prev_frame)
    loss += temporal_loss(x, w, c)
  return loss

def sum_shortterm_temporal_losses(sess, net, frame, input_img):
  x = sess.run(net['input'].assign(input_img))
  prev_frame = frame - 1
  w = get_prev_warped_frame(frame)
  c = get_content_weights(frame, prev_frame)
  loss = temporal_loss(x, w, c)
  return loss

'''
  utilities and i/o
'''
def read_image(path):
  # bgr image
  img = cv2.imread(path, cv2.IMREAD_COLOR)
  check_image(img, path)
  img = img.astype(np.float32)
  img = preprocess(img)
  return img

def write_image(path, img):
  img = postprocess(img)
  cv2.imwrite(path, img)

def preprocess(img):
  # bgr to rgb
  img = img[...,::-1]
  # shape (h, w, d) to (1, h, w, d)
  img = img[np.newaxis,:,:,:]
  img -= np.array([123.68, 116.779, 103.939]).reshape((1,1,1,3))
  return img

def postprocess(img):
  img += np.array([123.68, 116.779, 103.939]).reshape((1,1,1,3))
  # shape (1, h, w, d) to (h, w, d)
  img = img[0]
  img = np.clip(img, 0, 255).astype('uint8')
  # rgb to bgr
  img = img[...,::-1]
  return img

def read_flow_file(path):
  with open(path, 'rb') as f:
    # 4 bytes header
    header = struct.unpack('4s', f.read(4))[0]
    # 4 bytes width, height    
    w = struct.unpack('i', f.read(4))[0]
    h = struct.unpack('i', f.read(4))[0]   
    flow = np.ndarray((2, h, w), dtype=np.float32)
    for y in range(h):
      for x in range(w):
        flow[0,y,x] = struct.unpack('f', f.read(4))[0]
        flow[1,y,x] = struct.unpack('f', f.read(4))[0]
  return flow

def read_weights_file(path):
  lines = open(path).readlines()
  header = list(map(int, lines[0].split(' ')))
  w = header[0]
  h = header[1]
  vals = np.zeros((h, w), dtype=np.float32)
  for i in range(1, len(lines)):
    line = lines[i].rstrip().split(' ')
    vals[i-1] = np.array(list(map(np.float32, line)))
    vals[i-1] = list(map(lambda x: 0. if x < 255. else 1., vals[i-1]))
  # expand to 3 channels
  weights = np.dstack([vals.astype(np.float32)] * 3)
  return weights

def normalize(weights):
  denom = sum(weights)
  if denom > 0.:
    return [float(i) / denom for i in weights]
  else: return [0.] * len(weights)

def maybe_make_directory(dir_path):
  if not os.path.exists(dir_path):  
    os.makedirs(dir_path)

def check_image(img, path):
  if img is None:
    raise OSError(errno.ENOENT, "No such file", path)

'''
  rendering -- where the magic happens
'''
def stylize(content_img, style_imgs, init_img, frame=None):
  with tf.device(args.device), tf.Session() as sess:
    # setup network
    net = build_model(content_img)
    
    # style loss
    if args.style_mask:
      L_style = sum_masked_style_losses(sess, net, style_imgs)
    else:
      L_style = sum_style_losses(sess, net, style_imgs)
    
    # content loss
    L_content = sum_content_losses(sess, net, content_img)
    
    # denoising loss
    L_tv = tf.image.total_variation(net['input'])
    
    # loss weights
    alpha = args.content_weight
    beta  = args.style_weight
    theta = args.tv_weight
    
    # total loss
    L_total  = alpha * L_content
    L_total += beta  * L_style
    L_total += theta * L_tv
    
    # video temporal loss
    if args.video and frame > 1:
      gamma      = args.temporal_weight
      L_temporal = sum_shortterm_temporal_losses(sess, net, frame, init_img)
      L_total   += gamma * L_temporal

    # optimization algorithm
    optimizer = get_optimizer(L_total)

    if args.optimizer == 'adam':
      minimize_with_adam(sess, net, optimizer, init_img, L_total)
    elif args.optimizer == 'lbfgs':
      minimize_with_lbfgs(sess, net, optimizer, init_img)
    
    output_img = sess.run(net['input'])
    
    if args.original_colors:
      output_img = convert_to_original_colors(np.copy(content_img), output_img)

    if args.video:
      write_video_output(frame, output_img)
    else:
      write_image_output(output_img, content_img, style_imgs, init_img)

def minimize_with_lbfgs(sess, net, optimizer, init_img):
  if args.verbose: print('\nMINIMIZING LOSS USING: L-BFGS OPTIMIZER')
  init_op = tf.global_variables_initializer()
  sess.run(init_op)
  sess.run(net['input'].assign(init_img))
  optimizer.minimize(sess)

def minimize_with_adam(sess, net, optimizer, init_img, loss):
  if args.verbose: print('\nMINIMIZING LOSS USING: ADAM OPTIMIZER')
  train_op = optimizer.minimize(loss)
  init_op = tf.global_variables_initializer()
  sess.run(init_op)
  sess.run(net['input'].assign(init_img))
  iterations = 0
  while (iterations < args.max_iterations):
    sess.run(train_op)
    if iterations % args.print_iterations == 0 and args.verbose:
      curr_loss = loss.eval()
      print("At iterate {}\tf=  {:.5E}".format(iterations, curr_loss))
    iterations += 1

def get_optimizer(loss):
  print_iterations = args.print_iterations if args.verbose else 0
  if args.optimizer == 'lbfgs':
    optimizer = tf.contrib.opt.ScipyOptimizerInterface(
      loss, method='L-BFGS-B',
      options={'maxiter': args.max_iterations,
                  'disp': print_iterations})
  elif args.optimizer == 'adam':
    optimizer = tf.train.AdamOptimizer(args.learning_rate)
  return optimizer

def write_video_output(frame, output_img):
  fn = args.content_frame_frmt.format(str(frame).zfill(4))
  path = os.path.join(args.video_output_dir, fn)
  write_image(path, output_img)

def write_image_output(output_img, content_img, style_imgs, init_img):
  out_dir = os.path.join(args.img_output_dir, args.img_name)
  maybe_make_directory(out_dir)
  img_path = os.path.join(out_dir, args.img_name+'.png')
  content_path = os.path.join(out_dir, 'content.png')
  init_path = os.path.join(out_dir, 'init.png')

  write_image(img_path, output_img)
  write_image(content_path, content_img)
  write_image(init_path, init_img)
  index = 0
  for style_img in style_imgs:
    path = os.path.join(out_dir, 'style_'+str(index)+'.png')
    write_image(path, style_img)
    index += 1
  
  # save the configuration settings
  out_file = os.path.join(out_dir, 'meta_data.txt')
  f = open(out_file, 'w')
  f.write('image_name: {}\n'.format(args.img_name))
  f.write('content: {}\n'.format(args.content_img))
  index = 0
  for style_img, weight in zip(args.style_imgs, args.style_imgs_weights):
    f.write('styles['+str(index)+']: {} * {}\n'.format(weight, style_img))
    index += 1
  index = 0
  if args.style_mask_imgs is not None:
    for mask in args.style_mask_imgs:
      f.write('style_masks['+str(index)+']: {}\n'.format(mask))
      index += 1
  f.write('init_type: {}\n'.format(args.init_img_type))
  f.write('content_weight: {}\n'.format(args.content_weight))
  f.write('style_weight: {}\n'.format(args.style_weight))
  f.write('tv_weight: {}\n'.format(args.tv_weight))
  f.write('content_layers: {}\n'.format(args.content_layers))
  f.write('style_layers: {}\n'.format(args.style_layers))
  f.write('optimizer_type: {}\n'.format(args.optimizer))
  f.write('max_iterations: {}\n'.format(args.max_iterations))
  f.write('max_image_size: {}\n'.format(args.max_size))
  f.close()

'''
  image loading and processing
'''
def get_init_image(init_type, content_img, style_imgs, frame=None):
  if init_type == 'content':
    return content_img
  elif init_type == 'style':
    return style_imgs[0]
  elif init_type == 'random':
    init_img = get_noise_image(args.noise_ratio, content_img)
    return init_img
  # only for video frames
  elif init_type == 'prev':
    init_img = get_prev_frame(frame)
    return init_img
  elif init_type == 'prev_warped':
    init_img = get_prev_warped_frame(frame)
    return init_img

def get_content_frame(frame):
  fn = args.content_frame_frmt.format(str(frame).zfill(4))
  path = os.path.join(args.video_input_dir, fn)
  img = read_image(path)
  return img

def get_content_image(content_img):
  path = os.path.join(args.content_img_dir, content_img)
   # bgr image
  img = cv2.imread(path, cv2.IMREAD_COLOR)
  check_image(img, path)
  img = img.astype(np.float32)
  h, w, d = img.shape
  mx = args.max_size
  # resize if > max size
  if h > w and h > mx:
    w = (float(mx) / float(h)) * w
    img = cv2.resize(img, dsize=(int(w), mx), interpolation=cv2.INTER_AREA)
  if w > mx:
    h = (float(mx) / float(w)) * h
    img = cv2.resize(img, dsize=(mx, int(h)), interpolation=cv2.INTER_AREA)
  img = preprocess(img)
  return img

def get_style_images(content_img):
  _, ch, cw, cd = content_img.shape
  style_imgs = []
  for style_fn in args.style_imgs:
    path = os.path.join(args.style_imgs_dir, style_fn)
    # bgr image
    img = cv2.imread(path, cv2.IMREAD_COLOR)
    check_image(img, path)
    img = img.astype(np.float32)
    img = cv2.resize(img, dsize=(cw, ch), interpolation=cv2.INTER_AREA)
    img = preprocess(img)
    style_imgs.append(img)
  return style_imgs

def get_noise_image(noise_ratio, content_img):
  np.random.seed(args.seed)
  noise_img = np.random.uniform(-20., 20., content_img.shape).astype(np.float32)
  img = noise_ratio * noise_img + (1.-noise_ratio) * content_img
  return img

def get_mask_image(mask_img, width, height):
  path = os.path.join(args.content_img_dir, mask_img)
  img = cv2.imread(path, cv2.IMREAD_GRAYSCALE)
  check_image(img, path)
  img = cv2.resize(img, dsize=(width, height), interpolation=cv2.INTER_AREA)
  img = img.astype(np.float32)
  mx = np.amax(img)
  img /= mx
  return img

def get_prev_frame(frame):
  # previously stylized frame
  prev_frame = frame - 1
  fn = args.content_frame_frmt.format(str(prev_frame).zfill(4))
  path = os.path.join(args.video_output_dir, fn)
  img = cv2.imread(path, cv2.IMREAD_COLOR)
  check_image(img, path)
  return img

def get_prev_warped_frame(frame):
  prev_img = get_prev_frame(frame)
  prev_frame = frame - 1
  # backwards flow: current frame -> previous frame
  fn = args.backward_optical_flow_frmt.format(str(frame), str(prev_frame))
  path = os.path.join(args.video_input_dir, fn)
  flow = read_flow_file(path)
  warped_img = warp_image(prev_img, flow).astype(np.float32)
  img = preprocess(warped_img)
  return img

def get_content_weights(frame, prev_frame):
  forward_fn = args.content_weights_frmt.format(str(prev_frame), str(frame))
  backward_fn = args.content_weights_frmt.format(str(frame), str(prev_frame))
  forward_path = os.path.join(args.video_input_dir, forward_fn)
  backward_path = os.path.join(args.video_input_dir, backward_fn)
  forward_weights = read_weights_file(forward_path)
  backward_weights = read_weights_file(backward_path)
  return forward_weights #, backward_weights

def warp_image(src, flow):
  _, h, w = flow.shape
  flow_map = np.zeros(flow.shape, dtype=np.float32)
  for y in range(h):
    flow_map[1,y,:] = float(y) + flow[1,y,:]
  for x in range(w):
    flow_map[0,:,x] = float(x) + flow[0,:,x]
  # remap pixels to optical flow
  dst = cv2.remap(
    src, flow_map[0], flow_map[1], 
    interpolation=cv2.INTER_CUBIC, borderMode=cv2.BORDER_TRANSPARENT)
  return dst

def convert_to_original_colors(content_img, stylized_img):
  content_img  = postprocess(content_img)
  stylized_img = postprocess(stylized_img)
  if args.color_convert_type == 'yuv':
    cvt_type = cv2.COLOR_BGR2YUV
    inv_cvt_type = cv2.COLOR_YUV2BGR
  elif args.color_convert_type == 'ycrcb':
    cvt_type = cv2.COLOR_BGR2YCR_CB
    inv_cvt_type = cv2.COLOR_YCR_CB2BGR
  elif args.color_convert_type == 'luv':
    cvt_type = cv2.COLOR_BGR2LUV
    inv_cvt_type = cv2.COLOR_LUV2BGR
  elif args.color_convert_type == 'lab':
    cvt_type = cv2.COLOR_BGR2LAB
    inv_cvt_type = cv2.COLOR_LAB2BGR
  content_cvt = cv2.cvtColor(content_img, cvt_type)
  stylized_cvt = cv2.cvtColor(stylized_img, cvt_type)
  c1, _, _ = cv2.split(stylized_cvt)
  _, c2, c3 = cv2.split(content_cvt)
  merged = cv2.merge((c1, c2, c3))
  dst = cv2.cvtColor(merged, inv_cvt_type).astype(np.float32)
  dst = preprocess(dst)
  return dst

def render_single_image():
  content_img = get_content_image(args.content_img)
  style_imgs = get_style_images(content_img)
  with tf.Graph().as_default():
    print('\n---- RENDERING SINGLE IMAGE ----\n')
    init_img = get_init_image(args.init_img_type, content_img, style_imgs)
    tick = time.time()
    stylize(content_img, style_imgs, init_img)
    tock = time.time()
    print('Single image elapsed time: {}'.format(tock - tick))

def render_video():
  for frame in range(args.start_frame, args.end_frame+1):
    with tf.Graph().as_default():
      print('\n---- RENDERING VIDEO FRAME: {}/{} ----\n'.format(frame, args.end_frame))
      if frame == 1:
        content_frame = get_content_frame(frame)
        style_imgs = get_style_images(content_frame)
        init_img = get_init_image(args.first_frame_type, content_frame, style_imgs, frame)
        args.max_iterations = args.first_frame_iterations
        tick = time.time()
        stylize(content_frame, style_imgs, init_img, frame)
        tock = time.time()
        print('Frame {} elapsed time: {}'.format(frame, tock - tick))
      else:
        content_frame = get_content_frame(frame)
        style_imgs = get_style_images(content_frame)
        init_img = get_init_image(args.init_frame_type, content_frame, style_imgs, frame)
        args.max_iterations = args.frame_iterations
        tick = time.time()
        stylize(content_frame, style_imgs, init_img, frame)
        tock = time.time()
        print('Frame {} elapsed time: {}'.format(frame, tock - tick))

def main():
  global args
  args = parse_args()
  if args.video: render_video()
  else: render_single_image()

if __name__ == '__main__':
  main()