深度学习 ——style reconstruction

多层神经网络的实质就是为了找出更复杂，更内在的features...图像的style, how to describe, impossible! 但是人眼却可以分辨。

（参考论文 A Neural algorithm of artistic style）使用卷积神经网络来做  content， style reconstruction. 

• loss= content loss+ style loss

1. content reconstruction: 经过pre-trained 的vgg16 的卷积网络来做content 的重构， 使用前三层的content reconstruction 效果比较好，d,e 丢失了部分细节信息。保留了比较high-level 的信息(我做显著性检测中，使用高层信息，细节信息过少，考虑添加浅层的信息，加上c层的信息检测效果就很perfect)
2. style reconstruction: 在5层cnn上计算，相当于在不同的尺度上匹配图像本身的style

步骤：

1. 将content经过CNN，取relu4_2和relu5_2 的content_feature_maps

2. 将 style 经过CNN， 取5层relu 的style_feature_maps

3. 将最后结果图定义为一个变量tensor，经过CNN，得到feature_map_net最优化loss， loss=(feature_map_net-content_feature_maps)+(feature_map_net-style_feature_maps), 最优化结果图变量

实现：

1. 入口函数：

import os

import numpy as np
import scipy.misc

from stylize import stylize

import math
from argparse import ArgumentParser

from PIL import Image

# default arguments
CONTENT_WEIGHT = 5e0
CONTENT_WEIGHT_BLEND = 1
STYLE_WEIGHT = 5e2
TV_WEIGHT = 1e2
STYLE_LAYER_WEIGHT_EXP = 1
LEARNING_RATE = 1e1
BETA1 = 0.9
BETA2 = 0.999
EPSILON = 1e-08
STYLE_SCALE = 1.0
ITERATIONS = 1000
VGG_PATH = 'imagenet-vgg-verydeep-19.mat'
POOLING = 'max'

# 定义命令行参数
def build_parser():
    parser = ArgumentParser()
    # 要转化的图像
    parser.add_argument('--content',
            dest='content', help='content image',
            metavar='CONTENT', required=True)
    # 想要转化为的style,可以为多种style,nargs='+'
    parser.add_argument('--styles',
            dest='styles',
            nargs='+', help='one or more style images',
            metavar='STYLE', required=True)
   # 输出文件位置
    parser.add_argument('--output',
            dest='output', help='output path',
            metavar='OUTPUT', required=True)
    # 迭代次数
    parser.add_argument('--iterations', type=int,
            dest='iterations', help='iterations (default %(default)s)',
            metavar='ITERATIONS', default=ITERATIONS)
    parser.add_argument('--print-iterations', type=int,
            dest='print_iterations', help='statistics printing frequency',
            metavar='PRINT_ITERATIONS')
    parser.add_argument('--checkpoint-output',
            dest='checkpoint_output', help='checkpoint output format, e.g. output%%s.jpg',
            metavar='OUTPUT')
    parser.add_argument('--checkpoint-iterations', type=int,
            dest='checkpoint_iterations', help='checkpoint frequency',
            metavar='CHECKPOINT_ITERATIONS')
    parser.add_argument('--width', type=int,
            dest='width', help='output width',
            metavar='WIDTH')
    parser.add_argument('--style-scales', type=float,
            dest='style_scales',
            nargs='+', help='one or more style scales',
            metavar='STYLE_SCALE')
    parser.add_argument('--network',
            dest='network', help='path to network parameters (default %(default)s)',
            metavar='VGG_PATH', default=VGG_PATH)
    parser.add_argument('--content-weight-blend', type=float,
            dest='content_weight_blend', help='content weight blend, conv4_2 * blend + conv5_2 * (1-blend) (default %(default)s)',
            metavar='CONTENT_WEIGHT_BLEND', default=CONTENT_WEIGHT_BLEND)
    parser.add_argument('--content-weight', type=float,
            dest='content_weight', help='content weight (default %(default)s)',
            metavar='CONTENT_WEIGHT', default=CONTENT_WEIGHT)
    parser.add_argument('--style-weight', type=float,
            dest='style_weight', help='style weight (default %(default)s)',
            metavar='STYLE_WEIGHT', default=STYLE_WEIGHT)
    parser.add_argument('--style-layer-weight-exp', type=float,
            dest='style_layer_weight_exp', help='style layer weight exponentional increase - weight(layer<n+1>) = weight_exp*weight(layer<n>) (default %(default)s)',
            metavar='STYLE_LAYER_WEIGHT_EXP', default=STYLE_LAYER_WEIGHT_EXP)
    parser.add_argument('--style-blend-weights', type=float,
            dest='style_blend_weights', help='style blending weights',
            nargs='+', metavar='STYLE_BLEND_WEIGHT')
    parser.add_argument('--tv-weight', type=float,
            dest='tv_weight', help='total variation regularization weight (default %(default)s)',
            metavar='TV_WEIGHT', default=TV_WEIGHT)
    parser.add_argument('--learning-rate', type=float,
            dest='learning_rate', help='learning rate (default %(default)s)',
            metavar='LEARNING_RATE', default=LEARNING_RATE)
    parser.add_argument('--beta1', type=float,
            dest='beta1', help='Adam: beta1 parameter (default %(default)s)',
            metavar='BETA1', default=BETA1)
    parser.add_argument('--beta2', type=float,
            dest='beta2', help='Adam: beta2 parameter (default %(default)s)',
            metavar='BETA2', default=BETA2)
    parser.add_argument('--eps', type=float,
            dest='epsilon', help='Adam: epsilon parameter (default %(default)s)',
            metavar='EPSILON', default=EPSILON)
    parser.add_argument('--initial',
            dest='initial', help='initial image',
            metavar='INITIAL')
    parser.add_argument('--initial-noiseblend', type=float,
            dest='initial_noiseblend', help='ratio of blending initial image with normalized noise (if no initial image specified, content image is used) (default %(default)s)',
            metavar='INITIAL_NOISEBLEND')
    parser.add_argument('--preserve-colors', action='store_true',
            dest='preserve_colors', help='style-only transfer (preserving colors) - if color transfer is not needed')
    parser.add_argument('--pooling',
            dest='pooling', help='pooling layer configuration: max or avg (default %(default)s)',
            metavar='POOLING', default=POOLING)
    return parser


def main():
    parser = build_parser()
    options = parser.parse_args()
    # 对图像尺寸的限制
    if not os.path.isfile(options.network):
        parser.error("Network %s does not exist. (Did you forget to download it?)" % options.network)

    content_image = imread(options.content)
    style_images = [imread(style) for style in options.styles]

    width = options.width
    if width is not None:
        new_shape = (int(math.floor(float(content_image.shape[0]) /
                content_image.shape[1] * width)), width)
        content_image = scipy.misc.imresize(content_image, new_shape)
    target_shape = content_image.shape
    for i in range(len(style_images)):
        style_scale = STYLE_SCALE
        if options.style_scales is not None:
            style_scale = options.style_scales[i]
        style_images[i] = scipy.misc.imresize(style_images[i], style_scale *
                target_shape[1] / style_images[i].shape[1])

    # 多种style中，归一化每种style的权重
    style_blend_weights = options.style_blend_weights
    if style_blend_weights is None:
        # default is equal weights
        style_blend_weights = [1.0/len(style_images) for _ in style_images]
    else:
        total_blend_weight = sum(style_blend_weights)
        style_blend_weights = [weight/total_blend_weight
                               for weight in style_blend_weights]

    initial = options.initial
    if initial is not None:
        initial = scipy.misc.imresize(imread(initial), content_image.shape[:2])
        # Initial guess is specified, but not noiseblend - no noise should be blended
        if options.initial_noiseblend is None:
            options.initial_noiseblend = 0.0
    else:
        # Neither inital, nor noiseblend is provided, falling back to random generated initial guess
        if options.initial_noiseblend is None:
            options.initial_noiseblend = 1.0
        if options.initial_noiseblend < 1.0:
            initial = content_image

    if options.checkpoint_output and "%s" not in options.checkpoint_output:
        parser.error("To save intermediate images, the checkpoint output "
                     "parameter must contain `%s` (e.g. `foo%s.jpg`)")

    # 调用 stylize 函数
    for iteration, image in stylize(
        network=options.network,
        initial=initial,
        initial_noiseblend=options.initial_noiseblend,
        content=content_image,
        styles=style_images,
        preserve_colors=options.preserve_colors,
        iterations=options.iterations,
        content_weight=options.content_weight,
        content_weight_blend=options.content_weight_blend,
        style_weight=options.style_weight,
        style_layer_weight_exp=options.style_layer_weight_exp,
        style_blend_weights=style_blend_weights,
        tv_weight=options.tv_weight,
        learning_rate=options.learning_rate,
        beta1=options.beta1,
        beta2=options.beta2,
        epsilon=options.epsilon,
        pooling=options.pooling,
        print_iterations=options.print_iterations,
        checkpoint_iterations=options.checkpoint_iterations
    ):
        output_file = None
        combined_rgb = image
        if iteration is not None:
            if options.checkpoint_output:
                output_file = options.checkpoint_output % iteration
        else:
            output_file = options.output
        if output_file:
            imsave(output_file, combined_rgb)


def imread(path):
    img = scipy.misc.imread(path).astype(np.float)
    if len(img.shape) == 2:
        # grayscale
        img = np.dstack((img,img,img))
    elif img.shape[2] == 4:
        # PNG with alpha channel
        img = img[:,:,:3]
    return img


def imsave(path, img):
    img = np.clip(img, 0, 255).astype(np.uint8)
    Image.fromarray(img).save(path, quality=95)

if __name__ == '__main__':
    main()

　　2. style 生成函数：

import vgg

import tensorflow as tf
import numpy as np

from sys import stderr

from PIL import Image

# content_layers 和style_layers分布对应要计算loss 的层
CONTENT_LAYERS = ('relu4_2', 'relu5_2')
STYLE_LAYERS = ('relu1_1', 'relu2_1', 'relu3_1', 'relu4_1', 'relu5_1')

try:
    reduce
except NameError:
    from functools import reduce


def stylize(network, initial, initial_noiseblend, content, styles, preserve_colors, iterations,
        content_weight, content_weight_blend, style_weight, style_layer_weight_exp, style_blend_weights, tv_weight,
        learning_rate, beta1, beta2, epsilon, pooling,
        print_iterations=None, checkpoint_iterations=None):
    """
    Stylize images.
    This function yields tuples (iteration, image); `iteration` is None
    if this is the final image (the last iteration).  Other tuples are yielded
    every `checkpoint_iterations` iterations.
    :rtype: iterator[tuple[int|None,image]]
    """
    # 将shape从图像的三维变为4维
    shape = (1,) + content.shape
    style_shapes = [(1,) + style.shape for style in styles]
    content_features = {}
    style_features = [{} for _ in styles]

   # 加载网络
    vgg_weights, vgg_mean_pixel = vgg.load_net(network)
    # 计算style loss 时，每层的权重
    layer_weight = 1.0
    style_layers_weights = {}
    for style_layer in STYLE_LAYERS:
        style_layers_weights[style_layer] = layer_weight
        layer_weight *= style_layer_weight_exp

    # normalize style layer weights
    layer_weights_sum = 0
    for style_layer in STYLE_LAYERS:
        layer_weights_sum += style_layers_weights[style_layer]
    for style_layer in STYLE_LAYERS:
        style_layers_weights[style_layer] /= layer_weights_sum

    # compute content features in feedforward mode
    #定义一个graph来计算content feature maps
    g = tf.Graph()
    with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
        image = tf.placeholder('float', shape=shape)
        net = vgg.net_preloaded(vgg_weights, image, pooling)
        content_pre = np.array([vgg.preprocess(content, vgg_mean_pixel)])
        for layer in CONTENT_LAYERS:
            content_features[layer] = net[layer].eval(feed_dict={image: content_pre})

    # 循环定义graph 计算每种style 的feature maps
    # compute style features in feedforward mode
    for i in range(len(styles)):
        g = tf.Graph()
        with g.as_default(), g.device('/cpu:0'), tf.Session() as sess:
            image = tf.placeholder('float', shape=style_shapes[i])
            net = vgg.net_preloaded(vgg_weights, image, pooling)
            style_pre = np.array([vgg.preprocess(styles[i], vgg_mean_pixel)])
            for layer in STYLE_LAYERS:
                features = net[layer].eval(feed_dict={image: style_pre})
                features = np.reshape(features, (-1, features.shape[3]))
                gram = np.matmul(features.T, features) / features.size
                style_features[i][layer] = gram

    initial_content_noise_coeff = 1.0 - initial_noiseblend

    # 计算content loss 和style loss， 对loss 最小化，优化image 变量。
    # make stylized image using backpropogation
    with tf.Graph().as_default():
        if initial is None:
            noise = np.random.normal(size=shape, scale=np.std(content) * 0.1)
            initial = tf.random_normal(shape) * 0.256
        else:
            initial = np.array([vgg.preprocess(initial, vgg_mean_pixel)])
            initial = initial.astype('float32')
            noise = np.random.normal(size=shape, scale=np.std(content) * 0.1)
            initial = (initial) * initial_content_noise_coeff + (tf.random_normal(shape) * 0.256) * (1.0 - initial_content_noise_coeff)
        image = tf.Variable(initial)
        net = vgg.net_preloaded(vgg_weights, image, pooling)

        # content loss
        content_layers_weights = {}
        content_layers_weights['relu4_2'] = content_weight_blend
        content_layers_weights['relu5_2'] = 1.0 - content_weight_blend

        content_loss = 0
        content_losses = []
        for content_layer in CONTENT_LAYERS:
            content_losses.append(content_layers_weights[content_layer] * content_weight * (2 * tf.nn.l2_loss(
                    net[content_layer] - content_features[content_layer]) /
                    content_features[content_layer].size))
        content_loss += reduce(tf.add, content_losses)

        # style loss
        style_loss = 0
        for i in range(len(styles)):
            style_losses = []
            for style_layer in STYLE_LAYERS:
                layer = net[style_layer]
                _, height, width, number = map(lambda i: i.value, layer.get_shape())
                size = height * width * number
                feats = tf.reshape(layer, (-1, number))
                gram = tf.matmul(tf.transpose(feats), feats) / size
                style_gram = style_features[i][style_layer]
                style_losses.append(style_layers_weights[style_layer] * 2 * tf.nn.l2_loss(gram - style_gram) / style_gram.size)
            style_loss += style_weight * style_blend_weights[i] * reduce(tf.add, style_losses)

        # total variation denoising
        tv_y_size = _tensor_size(image[:,1:,:,:])
        tv_x_size = _tensor_size(image[:,:,1:,:])
        tv_loss = tv_weight * 2 * (
                (tf.nn.l2_loss(image[:,1:,:,:] - image[:,:shape[1]-1,:,:]) /
                    tv_y_size) +
                (tf.nn.l2_loss(image[:,:,1:,:] - image[:,:,:shape[2]-1,:]) /
                    tv_x_size))
        # overall loss
        loss = content_loss + style_loss + tv_loss

        # optimizer setup
        train_step = tf.train.AdamOptimizer(learning_rate, beta1, beta2, epsilon).minimize(loss)

        def print_progress():
            stderr.write('  content loss: %g
' % content_loss.eval())
            stderr.write('    style loss: %g
' % style_loss.eval())
            stderr.write('       tv loss: %g
' % tv_loss.eval())
            stderr.write('    total loss: %g
' % loss.eval())

        # optimization
        best_loss = float('inf')
        best = None
        with tf.Session() as sess:
            sess.run(tf.global_variables_initializer())
            stderr.write('Optimization started...
')
            if (print_iterations and print_iterations != 0):
                print_progress()
            for i in range(iterations):
                stderr.write('Iteration %4d/%4d
' % (i + 1, iterations))
                train_step.run()

                last_step = (i == iterations - 1)
                if last_step or (print_iterations and i % print_iterations == 0):
                    print_progress()

                if (checkpoint_iterations and i % checkpoint_iterations == 0) or last_step:
                    this_loss = loss.eval()
                    if this_loss < best_loss:
                        best_loss = this_loss
                        best = image.eval()

                    img_out = vgg.unprocess(best.reshape(shape[1:]), vgg_mean_pixel)

                    if preserve_colors and preserve_colors == True:
                        original_image = np.clip(content, 0, 255)
                        styled_image = np.clip(img_out, 0, 255)

                        # Luminosity transfer steps:
                        # 1. Convert stylized RGB->grayscale accoriding to Rec.601 luma (0.299, 0.587, 0.114)
                        # 2. Convert stylized grayscale into YUV (YCbCr)
                        # 3. Convert original image into YUV (YCbCr)
                        # 4. Recombine (stylizedYUV.Y, originalYUV.U, originalYUV.V)
                        # 5. Convert recombined image from YUV back to RGB

                        # 1
                        styled_grayscale = rgb2gray(styled_image)
                        styled_grayscale_rgb = gray2rgb(styled_grayscale)

                        # 2
                        styled_grayscale_yuv = np.array(Image.fromarray(styled_grayscale_rgb.astype(np.uint8)).convert('YCbCr'))

                        # 3
                        original_yuv = np.array(Image.fromarray(original_image.astype(np.uint8)).convert('YCbCr'))

                        # 4
                        w, h, _ = original_image.shape
                        combined_yuv = np.empty((w, h, 3), dtype=np.uint8)
                        combined_yuv[..., 0] = styled_grayscale_yuv[..., 0]
                        combined_yuv[..., 1] = original_yuv[..., 1]
                        combined_yuv[..., 2] = original_yuv[..., 2]

                        # 5
                        img_out = np.array(Image.fromarray(combined_yuv, 'YCbCr').convert('RGB'))


                    yield (
                        (None if last_step else i),
                        img_out
                    )


def _tensor_size(tensor):
    from operator import mul
    return reduce(mul, (d.value for d in tensor.get_shape()), 1)

def rgb2gray(rgb):
    return np.dot(rgb[...,:3], [0.299, 0.587, 0.114])

def gray2rgb(gray):
    w, h = gray.shape
    rgb = np.empty((w, h, 3), dtype=np.float32)
    rgb[:, :, 2] = rgb[:, :, 1] = rgb[:, :, 0] = gray
    return rgb

　　3. vgg 函数

import tensorflow as tf
import numpy as np
import scipy.io

VGG19_LAYERS = (
    'conv1_1', 'relu1_1', 'conv1_2', 'relu1_2', 'pool1',

    'conv2_1', 'relu2_1', 'conv2_2', 'relu2_2', 'pool2',

    'conv3_1', 'relu3_1', 'conv3_2', 'relu3_2', 'conv3_3',
    'relu3_3', 'conv3_4', 'relu3_4', 'pool3',

    'conv4_1', 'relu4_1', 'conv4_2', 'relu4_2', 'conv4_3',
    'relu4_3', 'conv4_4', 'relu4_4', 'pool4',

    'conv5_1', 'relu5_1', 'conv5_2', 'relu5_2', 'conv5_3',
    'relu5_3', 'conv5_4', 'relu5_4'
)

# 加载mat，得到权重和图像平均值
def load_net(data_path):
    data = scipy.io.loadmat(data_path)
    if not all(i in data for i in ('layers', 'classes', 'normalization')):
        raise ValueError("You're using the wrong VGG19 data. Please follow the instructions in the README to download the correct data.")
    mean = data['normalization'][0][0][0]
    mean_pixel = np.mean(mean, axis=(0, 1))
    weights = data['layers'][0]
    return weights, mean_pixel

# 表示每层的feature map 的计算
def net_preloaded(weights, input_image, pooling):
    net = {}
    current = input_image
    for i, name in enumerate(VGG19_LAYERS):
        kind = name[:4]
        if kind == 'conv':
            kernels, bias = weights[i][0][0][0][0]
            # matconvnet: weights are [width, height, in_channels, out_channels]
            # tensorflow: weights are [height, width, in_channels, out_channels]
            kernels = np.transpose(kernels, (1, 0, 2, 3))
            bias = bias.reshape(-1)
            current = _conv_layer(current, kernels, bias)
        elif kind == 'relu':
            current = tf.nn.relu(current)
        elif kind == 'pool':
            current = _pool_layer(current, pooling)
        net[name] = current

    assert len(net) == len(VGG19_LAYERS)
    return net

def _conv_layer(input, weights, bias):
    conv = tf.nn.conv2d(input, tf.constant(weights), strides=(1, 1, 1, 1),
            padding='SAME')
    return tf.nn.bias_add(conv, bias)


def _pool_layer(input, pooling):
    if pooling == 'avg':
        return tf.nn.avg_pool(input, ksize=(1, 2, 2, 1), strides=(1, 2, 2, 1),
                padding='SAME')
    else:
        return tf.nn.max_pool(input, ksize=(1, 2, 2, 1), strides=(1, 2, 2, 1),
                padding='SAME')

def preprocess(image, mean_pixel):
    return image - mean_pixel


def unprocess(image, mean_pixel):
    return image + mean_pixel