简介

使用TensorFlow实现快速图像风格迁移（Fast Neural Style Transfer）

原理

在之前介绍的图像风格迁移中，我们根据内容图片和风格图片优化输入图片，使得内容损失函数和风格损失函数尽可能小

和DeepDream一样，属于网络参数不变，根据损失函数调整输入数据，因此每生成一张图片都相当于训练一个模型，需要很长时间

训练模型需要很长时间，而使用训练好的模型进行推断则很快

使用快速图像风格迁移可大大缩短生成一张迁移图片所需的时间，其模型结构如下，包括转换网络和损失网络

风格图片是固定的，而内容图片是可变的输入，因此以上模型用于将任意图片快速转换为指定风格的图片

转换网络：参数需要训练，将内容图片转换成迁移图片
损失网络：计算迁移图片和风格图片之间的风格损失，以及迁移图片和原始内容图片之间的内容损失

经过训练后，转换网络所生成的迁移图片，在内容上和输入的内容图片相似，在风格上和指定的风格图片相似

进行推断时，仅使用转换网络，输入内容图片，即可得到对应的迁移图片

如果有多个风格图片，对每个风格分别训练一个模型即可

实现

基于以下两个项目进行修改，https://github.com/lengstrom/fast-style-transfer、https://github.com/hzy46/fast-neural-style-tensorflow

依然通过之前用过的imagenet-vgg-verydeep-19.mat计算内容损失函数和风格损失函数

需要一些图片作为输入的内容图片，对图片具体内容没有任何要求，也不需要任何标注，这里选择使用MSCOCO数据集的train2014部分，http://cocodataset.org/#download，共82612张图片

加载库

# -*- coding: utf-8 -*-

import tensorflow as tf
import numpy as np
import cv2
from imageio import imread, imsave
import scipy.io
import os
import glob
from tqdm import tqdm
import matplotlib.pyplot as plt
%matplotlib inline

查看风格图片，共10张

style_images = glob.glob('styles/*.jpg')
print(style_images)

加载内容图片，去掉黑白图片，处理成指定大小，暂时不进行归一化，像素值范围为0至255之间

def resize_and_crop(image, image_size):
    h = image.shape[0]
    w = image.shape[1]
    if h > w:
        image = image[h // 2 - w // 2: h // 2 + w // 2, :, :]
    else:
        image = image[:, w // 2 - h // 2: w // 2 + h // 2, :]    
    image = cv2.resize(image, (image_size, image_size))
    return image

X_data = []
image_size = 256
paths = glob.glob('train2014/*.jpg')
for i in tqdm(range(len(paths))):
    path = paths[i]
    image = imread(path)
    if len(image.shape) < 3:
        continue
    X_data.append(resize_and_crop(image, image_size))
X_data = np.array(X_data)
print(X_data.shape)

加载vgg19模型，并定义一个函数，对于给定的输入，返回vgg19各个层的输出值，就像在GAN中那样，通过variable_scope重用实现网络的重用

vgg = scipy.io.loadmat('imagenet-vgg-verydeep-19.mat')
vgg_layers = vgg['layers']

def vgg_endpoints(inputs, reuse=None):
    with tf.variable_scope('endpoints', reuse=reuse):
        def _weights(layer, expected_layer_name):
            W = vgg_layers[0][layer][0][0][2][0][0]
            b = vgg_layers[0][layer][0][0][2][0][1]
            layer_name = vgg_layers[0][layer][0][0][0][0]
            assert layer_name == expected_layer_name
            return W, b

        def _conv2d_relu(prev_layer, layer, layer_name):
            W, b = _weights(layer, layer_name)
            W = tf.constant(W)
            b = tf.constant(np.reshape(b, (b.size)))
            return tf.nn.relu(tf.nn.conv2d(prev_layer, filter=W, strides=[1, 1, 1, 1], padding='SAME') + b)

        def _avgpool(prev_layer):
            return tf.nn.avg_pool(prev_layer, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')

        graph = {}
        graph['conv1_1']  = _conv2d_relu(inputs, 0, 'conv1_1')
        graph['conv1_2']  = _conv2d_relu(graph['conv1_1'], 2, 'conv1_2')
        graph['avgpool1'] = _avgpool(graph['conv1_2'])
        graph['conv2_1']  = _conv2d_relu(graph['avgpool1'], 5, 'conv2_1')
        graph['conv2_2']  = _conv2d_relu(graph['conv2_1'], 7, 'conv2_2')
        graph['avgpool2'] = _avgpool(graph['conv2_2'])
        graph['conv3_1']  = _conv2d_relu(graph['avgpool2'], 10, 'conv3_1')
        graph['conv3_2']  = _conv2d_relu(graph['conv3_1'], 12, 'conv3_2')
        graph['conv3_3']  = _conv2d_relu(graph['conv3_2'], 14, 'conv3_3')
        graph['conv3_4']  = _conv2d_relu(graph['conv3_3'], 16, 'conv3_4')
        graph['avgpool3'] = _avgpool(graph['conv3_4'])
        graph['conv4_1']  = _conv2d_relu(graph['avgpool3'], 19, 'conv4_1')
        graph['conv4_2']  = _conv2d_relu(graph['conv4_1'], 21, 'conv4_2')
        graph['conv4_3']  = _conv2d_relu(graph['conv4_2'], 23, 'conv4_3')
        graph['conv4_4']  = _conv2d_relu(graph['conv4_3'], 25, 'conv4_4')
        graph['avgpool4'] = _avgpool(graph['conv4_4'])
        graph['conv5_1']  = _conv2d_relu(graph['avgpool4'], 28, 'conv5_1')
        graph['conv5_2']  = _conv2d_relu(graph['conv5_1'], 30, 'conv5_2')
        graph['conv5_3']  = _conv2d_relu(graph['conv5_2'], 32, 'conv5_3')
        graph['conv5_4']  = _conv2d_relu(graph['conv5_3'], 34, 'conv5_4')
        graph['avgpool5'] = _avgpool(graph['conv5_4'])

        return graph

选择一张风格图，减去通道颜色均值后，得到风格图片在vgg19各个层的输出值，计算四个风格层对应的Gram矩阵

style_index = 1
X_style_data = resize_and_crop(imread(style_images[style_index]), image_size)
X_style_data = np.expand_dims(X_style_data, 0)
print(X_style_data.shape)

MEAN_VALUES = np.array([123.68, 116.779, 103.939]).reshape((1, 1, 1, 3))

X_style = tf.placeholder(dtype=tf.float32, shape=X_style_data.shape, name='X_style')
style_endpoints = vgg_endpoints(X_style - MEAN_VALUES)
STYLE_LAYERS = ['conv1_2', 'conv2_2', 'conv3_3', 'conv4_3']
style_features = {}

sess = tf.Session()
for layer_name in STYLE_LAYERS:
    features = sess.run(style_endpoints[layer_name], feed_dict={X_style: X_style_data})
    features = np.reshape(features, (-1, features.shape[3]))
    gram = np.matmul(features.T, features) / features.size
    style_features[layer_name] = gram

定义转换网络，典型的卷积、残差、逆卷积结构，内容图片输入之前也需要减去通道颜色均值

batch_size = 4
X = tf.placeholder(dtype=tf.float32, shape=[None, None, None, 3], name='X')
k_initializer = tf.truncated_normal_initializer(0, 0.1)

def relu(x):
    return tf.nn.relu(x)

def conv2d(inputs, filters, kernel_size, strides):
    p = int(kernel_size / 2)
    h0 = tf.pad(inputs, [[0, 0], [p, p], [p, p], [0, 0]], mode='reflect')
    return tf.layers.conv2d(inputs=h0, filters=filters, kernel_size=kernel_size, strides=strides, padding='valid', kernel_initializer=k_initializer)

def deconv2d(inputs, filters, kernel_size, strides):
    shape = tf.shape(inputs)
    height, width = shape[1], shape[2]
    h0 = tf.image.resize_images(inputs, [height * strides * 2, width * strides * 2], tf.image.ResizeMethod.NEAREST_NEIGHBOR)
    return conv2d(h0, filters, kernel_size, strides)
    
def instance_norm(inputs):
    return tf.contrib.layers.instance_norm(inputs)

def residual(inputs, filters, kernel_size):
    h0 = relu(conv2d(inputs, filters, kernel_size, 1))
    h0 = conv2d(h0, filters, kernel_size, 1)
    return tf.add(inputs, h0)

with tf.variable_scope('transformer', reuse=None):
    h0 = tf.pad(X - MEAN_VALUES, [[0, 0], [10, 10], [10, 10], [0, 0]], mode='reflect')
    h0 = relu(instance_norm(conv2d(h0, 32, 9, 1)))
    h0 = relu(instance_norm(conv2d(h0, 64, 3, 2)))
    h0 = relu(instance_norm(conv2d(h0, 128, 3, 2)))

    for i in range(5):
        h0 = residual(h0, 128, 3)

    h0 = relu(instance_norm(deconv2d(h0, 64, 3, 2)))
    h0 = relu(instance_norm(deconv2d(h0, 32, 3, 2)))
    h0 = tf.nn.tanh(instance_norm(conv2d(h0, 3, 9, 1)))
    h0 = (h0 + 1) / 2 * 255.
    shape = tf.shape(h0)
    g = tf.slice(h0, [0, 10, 10, 0], [-1, shape[1] - 20, shape[2] - 20, -1], name='g')

将转换网络的输出即迁移图片，以及原始内容图片都输入到vgg19，得到各自对应层的输出，计算内容损失函数

CONTENT_LAYER = 'conv3_3'
content_endpoints = vgg_endpoints(X - MEAN_VALUES, True)
g_endpoints = vgg_endpoints(g - MEAN_VALUES, True)

def get_content_loss(endpoints_x, endpoints_y, layer_name):
    x = endpoints_x[layer_name]
    y = endpoints_y[layer_name]
    return 2 * tf.nn.l2_loss(x - y) / tf.to_float(tf.size(x))

content_loss = get_content_loss(content_endpoints, g_endpoints, CONTENT_LAYER)

根据迁移图片和风格图片在指定风格层的输出，计算风格损失函数

style_loss = []
for layer_name in STYLE_LAYERS:
    layer = g_endpoints[layer_name]
    shape = tf.shape(layer)
    bs, height, width, channel = shape[0], shape[1], shape[2], shape[3]
    
    features = tf.reshape(layer, (bs, height * width, channel))
    gram = tf.matmul(tf.transpose(features, (0, 2, 1)), features) / tf.to_float(height * width * channel)
    
    style_gram = style_features[layer_name]
    style_loss.append(2 * tf.nn.l2_loss(gram - style_gram) / tf.to_float(tf.size(layer)))

style_loss = tf.reduce_sum(style_loss)

计算全变差正则，得到总的损失函数

def get_total_variation_loss(inputs):
    h = inputs[:, :-1, :, :] - inputs[:, 1:, :, :]
    w = inputs[:, :, :-1, :] - inputs[:, :, 1:, :]
    return tf.nn.l2_loss(h) / tf.to_float(tf.size(h)) + tf.nn.l2_loss(w) / tf.to_float(tf.size(w)) 

total_variation_loss = get_total_variation_loss(g)

content_weight = 1
style_weight = 250
total_variation_weight = 0.01

loss = content_weight * content_loss + style_weight * style_loss + total_variation_weight * total_variation_loss

定义优化器，通过调整转换网络中的参数降低总损失

vars_t = [var for var in tf.trainable_variables() if var.name.startswith('transformer')]
optimizer = tf.train.AdamOptimizer(learning_rate=0.001).minimize(loss, var_list=vars_t)

训练模型，每轮训练结束后，用一张测试图片进行测试，并且将一些tensor的值写入events文件，便于使用tensorboard查看

style_name = style_images[style_index]
style_name = style_name[style_name.find('/') + 1:].rstrip('.jpg')
OUTPUT_DIR = 'samples_%s' % style_name
if not os.path.exists(OUTPUT_DIR):
    os.mkdir(OUTPUT_DIR)

tf.summary.scalar('losses/content_loss', content_loss)
tf.summary.scalar('losses/style_loss', style_loss)
tf.summary.scalar('losses/total_variation_loss', total_variation_loss)
tf.summary.scalar('losses/loss', loss)
tf.summary.scalar('weighted_losses/weighted_content_loss', content_weight * content_loss)
tf.summary.scalar('weighted_losses/weighted_style_loss', style_weight * style_loss)
tf.summary.scalar('weighted_losses/weighted_total_variation_loss', total_variation_weight * total_variation_loss)
tf.summary.image('transformed', g)
tf.summary.image('origin', X)
summary = tf.summary.merge_all()
writer = tf.summary.FileWriter(OUTPUT_DIR)

sess.run(tf.global_variables_initializer())
losses = []
epochs = 2

X_sample = imread('sjtu.jpg')
h_sample = X_sample.shape[0]
w_sample = X_sample.shape[1]

for e in range(epochs):
    data_index = np.arange(X_data.shape[0])
    np.random.shuffle(data_index)
    X_data = X_data[data_index]
    
    for i in tqdm(range(X_data.shape[0] // batch_size)):
        X_batch = X_data[i * batch_size: i * batch_size + batch_size]
        ls_, _ = sess.run([loss, optimizer], feed_dict={X: X_batch})
        losses.append(ls_)
        
        if i > 0 and i % 20 == 0:
            writer.add_summary(sess.run(summary, feed_dict={X: X_batch}), e * X_data.shape[0] // batch_size + i)
            writer.flush()
        
    print('Epoch %d Loss %f' % (e, np.mean(losses)))
    losses = []

    gen_img = sess.run(g, feed_dict={X: [X_sample]})[0]
    gen_img = np.clip(gen_img, 0, 255)
    result = np.zeros((h_sample, w_sample * 2, 3))
    result[:, :w_sample, :] = X_sample / 255.
    result[:, w_sample:, :] = gen_img[:h_sample, :w_sample, :] / 255.
    plt.axis('off')
    plt.imshow(result)
    plt.show()
    imsave(os.path.join(OUTPUT_DIR, 'sample_%d.jpg' % e), result)

保存模型

saver = tf.train.Saver()
saver.save(sess, os.path.join(OUTPUT_DIR, 'fast_style_transfer'))

测试图片依旧是之前用过的交大庙门

风格迁移结果

训练过程中可以使用tensorboard查看训练过程

tensorboard --logdir=samples_starry

在单机上使用以下代码即可快速完成风格迁移，在CPU上也只需要10秒左右

# -*- coding: utf-8 -*-

import tensorflow as tf
import numpy as np
from imageio import imread, imsave
import os
import time

def the_current_time():
    print(time.strftime("%Y-%m-%d %H:%M:%S", time.localtime(int(time.time()))))

style = 'wave'
model = 'samples_%s' % style
content_image = 'sjtu.jpg'
result_image = 'sjtu_%s.jpg' % style
X_image = imread(content_image)

sess = tf.Session()
sess.run(tf.global_variables_initializer())

saver = tf.train.import_meta_graph(os.path.join(model, 'fast_style_transfer.meta'))
saver.restore(sess, tf.train.latest_checkpoint(model))

graph = tf.get_default_graph()
X = graph.get_tensor_by_name('X:0')
g = graph.get_tensor_by_name('transformer/g:0')

the_current_time()

gen_img = sess.run(g, feed_dict={X: [X_image]})[0]
gen_img = np.clip(gen_img, 0, 255) / 255.
imsave(result_image, gen_img)

the_current_time()

对于其他风格图片，用相同方法训练对应模型即可

参考

Perceptual Losses for Real-Time Style Transfer and Super-Resolution：https://arxiv.org/abs/1603.08155v1
Fast Style Transfer in TensorFlow：https://github.com/lengstrom/fast-style-transfer
A Tensorflow Implementation for Fast Neural Style：https://github.com/hzy46/fast-neural-style-tensorflow

视频讲解课程

深度有趣（一）