基于DCGAN的动漫头像生成神经网络实现

本文主要是介绍基于DCGAN的动漫头像生成神经网络实现，希望对大家解决编程问题提供一定的参考价值，需要的开发者们随着小编来一起学习吧！

一、前言

1、什么是DCGAN？

2、DCGAN的TensorFlow实现

3、什么是转置卷积？

4、转置卷积的Tensorflow实现

5、Batch Normalization解读

本文假设读者已经了解GAN及CNN的基本原理实现，如不清楚可参考以下文章：

基于GAN的的mnist训练集图片生成神经网络实现

基于CNN的验证码识别神经网络实现

二、实战

1、训练数据处理

(1)数据源：百度云盘提取码：g5qa

(2)创建一个生成器

class Avatar:def __init__(self):self.data_name = 'faces'self.source_shape = (96, 96, 3)self.resize_shape = (48, 48, 3)self.crop = Trueself.img_shape = self.source_shape if not self.crop else self.resize_shapeself.img_list = self._get_img_list()self.batch_size = 64self.batch_shape = (self.batch_size, ) + self.img_shapeself.chunk_size = len(self.img_list) // self.batch_sizedef _get_img_list(self):path = os.path.join(os.getcwd(), self.data_name, '*.jpg')return glob(path)def _get_img(self, name):assert name in self.img_listimg = scipy.misc.imread(name).astype(np.float32)assert img.shape == self.source_shapereturn self._resize(img) if self.crop else imgdef _resize(self, img):h, w = img.shape[:2]resize_h, resize_w = self.resize_shape[:2]crop_h, crop_w = self.source_shape[:2]j = int(round((h - crop_h) / 2.))i = int(round((w - crop_w) / 2.))cropped_image = scipy.misc.imresize(img[j:j + crop_h, i:i + crop_w], [resize_h, resize_w])return np.array(cropped_image) / 127.5 - 1.@staticmethoddef save_img(image, path):scipy.misc.imsave(path, image)return Truedef batches(self):start = 0end = self.batch_sizefor _ in range(self.chunk_size):name_list = self.img_list[start:end]imgs = [self._get_img(name) for name in name_list]batches = np.zeros(self.batch_shape)batches[::] = imgsyield batchesstart += self.batch_sizeend += self.batch_size

读取本地图片数据并创建一个生成器，作为后续模型数据源

2.模型参数定义

    def __init__(self):self.avatar = Avatar()# 真实图片shape (height, width, depth)self.img_shape = self.avatar.img_shape# 一个batch的图片向量shape (batch, height, width, depth)self.batch_shape = self.avatar.batch_shape# 一个batch包含图片数量self.batch_size = self.avatar.batch_size# batch数量self.chunk_size = self.avatar.chunk_size# 噪音图片sizeself.noise_img_size = 100# 卷积转置输出通道数量self.gf_size = 64# 卷积输出通道数量self.df_size = 64# 训练循环次数self.epoch_size = 50# 学习率self.learning_rate = 0.0002# 优化指数衰减率self.beta1 = 0.5# 生成图片数量self.sample_size = 64

3、输入定义

        # 真实图片real_imgs = tf.placeholder(tf.float32, self.batch_shape, name='real_images')# 噪声图片noise_imgs = tf.placeholder(tf.float32, [None, self.noise_img_size], name='noise_images')

我们利用随机的噪音输入来生成图片

4、生成器

    def generator(self, noise_imgs, train=True):with tf.variable_scope('generator'):# 分别对应每个layer的height, widths_h, s_w, _ = self.img_shapes_h2, s_w2 = self.conv_out_size_same(s_h, 2), self.conv_out_size_same(s_w, 2)s_h4, s_w4 = self.conv_out_size_same(s_h2, 2), self.conv_out_size_same(s_w2, 2)s_h8, s_w8 = self.conv_out_size_same(s_h4, 2), self.conv_out_size_same(s_w4, 2)s_h16, s_w16 = self.conv_out_size_same(s_h8, 2), self.conv_out_size_same(s_w8, 2)# layer 0# 对输入噪音图片进行线性变换z, h0_w, h0_b = self.linear(noise_imgs, self.gf_size*8*s_h16*s_w16)# reshape为合适的输入层格式h0 = tf.reshape(z, [-1, s_h16, s_w16, self.gf_size * 8])# 对数据进行归一化处理 加快收敛速度h0 = self.batch_normalizer(h0, train=train, name='g_bn0')# 激活函数h0 = tf.nn.relu(h0)# layer 1# 卷积转置进行上采样h1, h1_w, h1_b = self.deconv2d(h0, [self.batch_size, s_h8, s_w8, self.gf_size*4], name='g_h1')h1 = self.batch_normalizer(h1, train=train, name='g_bn1')h1 = tf.nn.relu(h1)# layer 2h2, h2_w, h2_b = self.deconv2d(h1, [self.batch_size, s_h4, s_w4, self.gf_size*2], name='g_h2')h2 = self.batch_normalizer(h2, train=train, name='g_bn2')h2 = tf.nn.relu(h2)# layer 3h3, h3_w, h3_b = self.deconv2d(h2, [self.batch_size, s_h2, s_w2, self.gf_size*1], name='g_h3')h3 = self.batch_normalizer(h3, train=train, name='g_bn3')h3 = tf.nn.relu(h3)# layer 4h4, h4_w, h4_b = self.deconv2d(h3, self.batch_shape, name='g_h4')return tf.nn.tanh(h4)

DCGAN的生成器为卷积网络，使用转置卷积进行上采样，去除pooling层，利用batch normalization加快收敛速度。

5、判别器

    def discriminator(self, real_imgs, reuse=False):with tf.variable_scope("discriminator", reuse=reuse):# layer 0# 卷积操作h0 = self.conv2d(real_imgs, self.df_size, name='d_h0_conv')# 激活函数h0 = self.lrelu(h0)# layer 1h1 = self.conv2d(h0, self.df_size*2, name='d_h1_conv')h1 = self.batch_normalizer(h1, name='d_bn1')h1 = self.lrelu(h1)# layer 2h2 = self.conv2d(h1, self.df_size*4, name='d_h2_conv')h2 = self.batch_normalizer(h2, name='d_bn2')h2 = self.lrelu(h2)# layer 3h3 = self.conv2d(h2, self.df_size*8, name='d_h3_conv')h3 = self.batch_normalizer(h3, name='d_bn3')h3 = self.lrelu(h3)# layer 4h4, _, _ = self.linear(tf.reshape(h3, [self.batch_size, -1]), 1, name='d_h4_lin')return tf.nn.sigmoid(h4), h4

DCGAN的判别器为卷积网络，这里使用卷积操作对图像进行特征提取识别。

6、损失和优化

    @staticmethoddef loss_graph(real_logits, fake_logits):# 生成器图片loss# 生成器希望判别器判断出来的标签为1gen_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=fake_logits, labels=tf.ones_like(fake_logits)))# 判别器识别生成器图片loss# 判别器希望识别出来的标签为0fake_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=fake_logits, labels=tf.zeros_like(fake_logits)))# 判别器识别真实图片loss# 判别器希望识别出来的标签为1real_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=real_logits, labels=tf.ones_like(real_logits)))# 判别器总lossdis_loss = tf.add(fake_loss, real_loss)return gen_loss, fake_loss, real_loss, dis_loss@staticmethoddef optimizer_graph(gen_loss, dis_loss, learning_rate, beta1):# 所有定义变量train_vars = tf.trainable_variables()# 生成器变量gen_vars = [var for var in train_vars if var.name.startswith('generator')]# 判别器变量dis_vars = [var for var in train_vars if var.name.startswith('discriminator')]# optimizer# 生成器与判别器作为两个网络需要分别优化gen_optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate, beta1=beta1).minimize(gen_loss, var_list=gen_vars)dis_optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate, beta1=beta1).minimize(dis_loss, var_list=dis_vars)return gen_optimizer, dis_optimizer

7、开始训练

        # 开始训练saver = tf.train.Saver()step = 0# 指定占用GPU比例# tensorflow默认占用全部GPU显存 防止在机器显存被其他程序占用过多时可能在启动时报错gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.8)with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:sess.run(tf.global_variables_initializer())for epoch in range(self.epoch_size):batches = self.avatar.batches()for batch_imgs in batches:# generator的输入噪声noises = np.random.uniform(-1, 1, size=(self.batch_size, self.noise_img_size)).astype(np.float32)# 优化_ = sess.run(dis_optimizer, feed_dict={real_imgs: batch_imgs, noise_imgs: noises})_ = sess.run(gen_optimizer, feed_dict={noise_imgs: noises})_ = sess.run(gen_optimizer, feed_dict={noise_imgs: noises})step += 1print(datetime.now().strftime('%c'), epoch, step)

8、结果