生死看淡，不服就GAN(八)----WGAN的改进版本WGAN-GP

WGAN-GP是针对WGAN的存在的问题提出来的，WGAN在真实的实验过程中依旧存在着训练困难、收敛速度慢的 问题，相比较传统GAN在实验上提升不是很明显。WGAN-GP在文章中指出了WGAN存在问题的原因，那就是WGAN在处理Lipschitz限制条件时直接采用了 weight clipping。相关讲解请参考

WGAN-GP的介绍

同往期一样，依然以生成cifar数据集中马的彩色图片为例，关于cifar数据集的读取和生成器模型的验证请参考第6期：

用DCGAN生成马的彩色图片

下面给出WGAN-GP框架

"""
-------------------------------------------------------生死看淡，不服就GAN-------------------------------------------------------------------------
PROJECT: CIFAR10_WGAN-GP
Author: Ephemeroptera
Date:2019-3-19
QQ:605686962

"""
"""
WGAN说明：相比较WGAN，WGAN-GP提出以下改进：
                                        （1）用对判别器梯度惩罚取代WGAN的判决器权值区间截断
                                        （2）判别器取消BN操作
                                        （3）优化器使用ADAM
"""
# 导入包
import numpy as np
import tensorflow as tf
import pickle
import TFRecordTools
import time

############################################### 设置参数 ####################################################################################

real_shape = [-1,32,32,3] # 真实样本尺寸
data_total = 5000 # 真实样本个数
batch_size = 64 # 批大小
noise_size = 128 # 噪声维度
max_iters = 50000 #的最大迭代次数
learning_rate = 5e-5 # 学习率
beta1 = 0.5# ADAM参数1
beta2 = 0.9# ADAM参数2
CRITIC_NUM = 5 # 每次迭代判别器训练次数
lam = 10 #梯度惩罚权重

############################################# 定义生成器和判别器 #############################################################################

# 定义生成器（32x32图片）
def Generator_DC_32x32(z, channel, is_train=True):
    """
    :param z: 噪声信号，tensor类型
    :param channnel: 生成图片的通道数
    :param is_train: 是否为训练状态，该参数主要用于作为batch_normalization方法中的参数使用(训练时候开启)
    """
    # 训练时生成器不允许复用
    with tf.variable_scope("generator", reuse=(not is_train)):

        # layer1: noise_dim --> 4*4*512 --> 4x4x512 -->BN+relu
        layer1 = tf.layers.dense(z, 4 * 4 * 512)
        layer1 = tf.reshape(layer1, [-1, 4, 4, 512])
        layer1 = tf.layers.batch_normalization(layer1, training=is_train,)
        layer1 = tf.nn.relu(layer1)
        # layer1 = tf.nn.dropout(layer1, keep_prob=0.8)# dropout

        # layer2: deconv(ks=3x3,s=2,padding=same):4x4x512 --> 8x8x256 --> BN+relu
        layer2 = tf.layers.conv2d_transpose(layer1, 256, 3, strides=2, padding='same',
                                            kernel_initializer=tf.random_normal_initializer(0, 0.02),
                                            bias_initializer=tf.random_normal_initializer(0, 0.02))
        layer2 = tf.layers.batch_normalization(layer2, training=is_train)
        layer2 = tf.nn.relu(layer2)
        # layer2 = tf.nn.dropout(layer2, keep_prob=0.8)# dropout

        # layer3: deconv(ks=3x3,s=2,padding=same):8x8x256 --> 16x16x128 --> BN+relu
        layer3 = tf.layers.conv2d_transpose(layer2, 128, 3, strides=2, padding='same',
                                            kernel_initializer=tf.random_normal_initializer(0, 0.02),
                                            bias_initializer=tf.random_normal_initializer(0, 0.02))
        layer3 = tf.layers.batch_normalization(layer3, training=is_train)
        layer3 = tf.nn.relu(layer3)
        # layer3 = tf.nn.dropout(layer3, keep_prob=0.8)# dropout

        # layer4: deconv(ks=3x3,s=2,padding=same):16x16x128 --> 32x32x64--> BN+relu
        layer4 = tf.layers.conv2d_transpose(layer3, 64, 3, strides=2, padding='same',
                                            kernel_initializer=tf.random_normal_initializer(0, 0.02),
                                            bias_initializer=tf.random_normal_initializer(0, 0.02))
        layer4 = tf.layers.batch_normalization(layer4, training=is_train)
        layer4 = tf.nn.relu(layer4)
        # layer4 = tf.nn.dropout(layer3, keep_prob=0.8)# dropout

        # logits: deconv(ks=3x3,s=2,padding=same):32x32x64 --> 32x32x3
        logits = tf.layers.conv2d_transpose(layer4, channel, 3, strides=1, padding='same',
                                            kernel_initializer=tf.random_normal_initializer(0, 0.02),
                                            bias_initializer=tf.random_normal_initializer(0, 0.02))
        # outputs
        outputs = tf.tanh(logits)

        return logits,outputs

# 定义判别器（32x32）
def Discriminator_DC_32x32(inputs_img, reuse=False, GAN = False,GP= False,alpha=0.2):
    """
    @param inputs_img: 输入图片，tensor类型
    @param reuse:判别器复用
    @param GP: 使用WGAN-GP时关闭BN
    @param alpha: Leaky ReLU系数
    """

    with tf.variable_scope("discriminator", reuse=reuse):

        # layer1: conv(ks=3x3,s=2,padding=same)+lrelu -->32x32x3 to 16x16x128
        layer1 = tf.layers.conv2d(inputs_img, 128, 3, strides=2, padding='same')
        if GP is False:
            layer1 = tf.layers.batch_normalization(layer1, training=True)
        layer1 = tf.nn.leaky_relu(layer1,alpha=alpha)
        # layer1 = tf.nn.dropout(layer1, keep_prob=0.8)

        # layer2: conv(ks=3x3,s=2,padding=same)+BN+lrelu -->16x16x128 to 8x8x256
        layer2 = tf.layers.conv2d(layer1, 256, 3, strides=2, padding='same')
        if GP is False:
            layer2 = tf.layers.batch_normalization(layer2, training=True)
        layer2 = tf.nn.leaky_relu(layer2, alpha=alpha)
        # layer2 = tf.nn.dropout(layer2, keep_prob=0.8)

        # layer3: conv(ks=3x3,s=2,padding=same)+BN+lrelu -->8x8x256 to 4x4x512
        layer3 = tf.layers.conv2d(layer2, 512, 3, strides=2, padding='same')
        if GP is False:
            layer3 = tf.layers.batch_normalization(layer3, training=True)
        layer3 = tf.nn.leaky_relu(layer3, alpha=alpha)
        layer3 = tf.reshape(layer3, [-1, 4*4* 512])
        # layer3 = tf.nn.dropout(layer2, keep_prob=0.8)

        # logits,output:
        logits = tf.layers.dense(layer3, 1)
        "WGAN:去除sigmoid"
        if GAN:
            outputs = None
        else:
            outputs = tf.sigmoid(logits)

        return logits, outputs

############################################## 定义计算图（网络） #######################################################

#----------------------输入----------------

inputs_real = tf.placeholder(tf.float32, [None, real_shape[1], real_shape[2], real_shape[3]], name='inputs_real') # 真实样本输入
inputs_noise = tf.placeholder(tf.float32, [None, noise_size], name='inputs_noise') # 生成样本输入

#-------------------生成和判别--------------
# 生成样本
_,g_outputs = Generator_DC_32x32(inputs_noise, real_shape[3], is_train=True) # 训练生成器
_,g_test = Generator_DC_32x32(inputs_noise, real_shape[3], is_train=False) # 测试生成器
# 判别样本
'WGAN-GP:判别器废除批归一化'
d_logits_real, _ = Discriminator_DC_32x32(inputs_real,GAN=True,GP=True) #识别真样本
d_logits_fake, _ = Discriminator_DC_32x32(g_outputs,GAN=True,GP=True,reuse=True) #识别假样本

#------------定义原始GAN的损失函数--------------
"WGAN:损失函数去log，采用Wasserstein距离形式"
# 生成器
g_loss = tf.reduce_mean(-d_logits_fake)
# 判别器
d_loss = tf.reduce_mean(d_logits_fake - d_logits_real)
'WGAN-GP:加入判别器梯度惩罚项'
# 判别器梯度惩罚项
alpha_dist = tf.contrib.distributions.Uniform(low=0., high=1.)#获取[0,1]之间正态分布
alpha = alpha_dist.sample((batch_size, 1, 1, 1))
interpolated = inputs_real + alpha*(g_outputs-inputs_real)# 对真实样本和生成样本之间插值
inte_logit,_ = Discriminator_DC_32x32(interpolated, GAN=True,GP=True,reuse=True)
gradients = tf.gradients(inte_logit, [interpolated,])[0]# 求得判别器梯度
grad_l2 = tf.sqrt(tf.reduce_sum(tf.square(gradients), axis=[1,2,3]))
gradient_penalty = tf.reduce_mean((grad_l2-1)**2) # 定义惩罚项
# 加入d_loss
d_loss+=gradient_penalty*lam
#-------------------训练模型-----------------
# 分别获取生成器和判别器的变量空间
train_vars = tf.trainable_variables()
g_vars = [var for var in train_vars if var.name.startswith("generator")]
d_vars = [var for var in train_vars if var.name.startswith("discriminator")]

# Optimizer
with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):# 保证分布白化先完成
    g_train_opt = tf.train.AdamOptimizer(learning_rate=learning_rate, beta1=beta1, beta2=beta2).minimize(g_loss, var_list=g_vars)
    d_train_opt = tf.train.AdamOptimizer(learning_rate=learning_rate, beta1=beta1, beta2=beta2).minimize(d_loss, var_list=d_vars)
############################################# 调用TFRecord读取数据 #####################################################

# 读取TFR,不打乱文件顺序，指定数据类型，开启多线程
[data,label] = TFRecordTools.ReadFromTFRecord(sameName= r'.\TFR\class7-*',isShuffle= False,datatype= tf.float64,
                                labeltype= tf.int32,isMultithreading= True)
# 批量处理，送入队列数据，指定数据大小，打乱数据项，设置批次大小64
[data_batch,label_batch] = TFRecordTools.DataBatch(data,label,dataSize= 32*32*3,labelSize= 1,
                                                   isShuffle= True,batchSize= 64)

############################################### 迭代 ###################################################################

# 存储训练过程中生成日志
GenLog = []
# 存储loss
losses = []
# 保存生成器变量(仅保存生成器模型，保存最近5个)
saver = tf.train.Saver(var_list=[var for var in tf.trainable_variables()
                                 if var.name.startswith("generator")],max_to_keep=5)
# 定义批预处理
def batch_preprocess(data_batch):
    # 提取批数据
    batch = sess.run(data_batch)
    # 整理成RGB（Nx32x32x3）
    batch_images = np.reshape(batch, [-1, 3, 32, 32]).transpose((0, 2, 3, 1))  # (-1,32,32,3)
    # scale to -1, 1
    batch_images = batch_images * 2 - 1
    return  batch_images

# 生成相关目录保存生成信息
def GEN_DIR():
    import os
    if not os.path.isdir('ckpt'):
        print('文件夹ckpt未创建，现在在当前目录下创建..')
        os.mkdir('ckpt')
    if not os.path.isdir('trainLog'):
        print('文件夹ckpt未创建，现在在当前目录下创建..')
        os.mkdir('trainLog')

# 开启会话
with tf.Session() as sess:
    # 生成相关目录
    GEN_DIR()

    # 初始化变量
    init = (tf.global_variables_initializer(), tf.local_variables_initializer())
    sess.run(init)

    # 开启协调器
    coord = tf.train.Coordinator()
    # 启动线程
    threads = tf.train.start_queue_runners(sess=sess, coord=coord)

    time_start = time.time() # 开始计时
    for steps in range(max_iters):
        steps += 1

        # 判别器重复训练设置
        if steps < 25 or steps % 500 == 0:
            critic_num = 100
        else:
            critic_num = CRITIC_NUM

        # 重复训练判别器
        for i in range(CRITIC_NUM):
            batch_images = batch_preprocess(data_batch)  # images
            batch_noise = np.random.normal(size=(batch_size, noise_size))  # noise(normal)
            _ = sess.run(d_train_opt, feed_dict={inputs_real: batch_images,
                                                 inputs_noise: batch_noise})

        # 训练生成器
        batch_images = batch_preprocess(data_batch)  # images
        batch_noise = np.random.normal(size=(batch_size, noise_size))  # noise(normal)
        _ = sess.run(g_train_opt, feed_dict={inputs_real: batch_images,
                                             inputs_noise: batch_noise})

        #  记录训练信息
        if steps % 5 == 1:
            # （1）记录损失函数
            train_loss_d = d_loss.eval({inputs_real: batch_images,
                                        inputs_noise: batch_noise})
            train_loss_g = g_loss.eval({inputs_real: batch_images,
                                        inputs_noise: batch_noise})
            losses.append([train_loss_d, train_loss_g,steps])

            # （2）记录生成样本
            batch_noise = np.random.normal(size=(batch_size, noise_size))
            gen_samples = sess.run(g_test, feed_dict={inputs_noise: batch_noise})
            genLog = (gen_samples[0:11] + 1) / 2  # 恢复颜色空间(取10张)
            GenLog.append(genLog)

            # (3)打印信息
            print('step {}...'.format(steps),
                  "Discriminator Loss: {:.4f}...".format(train_loss_d),
                  "Generator Loss: {:.4f}...".format(train_loss_g))

        # （4）保存生成模型
        if steps % 300 ==0:
            saver.save(sess, './ckpt/generator.ckpt', global_step=steps)

    # 关闭线程
    coord.request_stop()
    coord.join(threads)

#计时结束：
time_end = time.time()
print('迭代结束，耗时：%.2f秒'%(time_end-time_start))

# 保存信息
#  保存loss记录
with open('./trainLog/loss_variation.loss', 'wb') as l:
    losses = np.array(losses)
    pickle.dump(losses,l)
    print('保存loss信息..')

# 保存生成日志
with open('./trainLog/GenLog.log', 'wb') as g:
    pickle.dump(GenLog, g)
    print('保存GenLog信息..')

结果展示

最后一次生成样本

训练期间生成日志

损失函数

验证生成器