[生成对抗网络GAN入门指南]（8）SGAN：Semi-Supervised Learning with Generative Adversarial Networks

本篇blog的内容基于原始论文SGAN：Semi-Supervised Learning with Generative Adversarial Networks和《生成对抗网络入门指南》第六章。（论文比较短只有3页，其中还包含一页citation）

---

为什么研究SGAN？

• 有大量的数据是不带标签的，带标签的数据只占一小部分；

• 在DCGAN的研究中我们看到使用生成模型特征抽取后形成的判别器可以达到分类的效果。由判别器D学习到的特征可以提升分类器C的效果，两者相辅相成；

• 现有的训练方式无法同时训练分类器C和生成器G；

• 提升判别器后，生成器G的效果也会变好；

三者交替会趋向于一个理想的平衡点。

 

一、SGAN

Semi-Supervised Learning with Generative Adversarial Networks

• 传统的GAN在判别器网络的输出端会使用二分类模式，代表真和假。

• 在SGAN中，就是把这个二分类（sigmoid）转化为多分类（softmax），类型数量为N+1，指代N个标签的数据和“一个假数据”，表示为 。

 

二、伪代码

                                                         SGAN伪代码

---

输入：I:总迭代次数

       

              从生成器前置随机分布  取出 m 个随机样本 ；

              从真实数据分布 取出 m 个真实样本  ；

              最小化 NLL，更新  的参数；

              从生成器前置随机分布  取出 m 个随机样本 ；

              最大化 NLL ，更新  的参数；

       

---

跟之前的cGAN对比，差别在于：

• 对于生成器的输入端我们没有将标签信息作为输入，所以判别器产生的生成数据是随机分布的，不受网络输入的控制。

• 对于判别器的输出而言cGAN仅仅是一个真或者假的二分类，而SGAN则是一个分类器与判别器的结合体。

 

三、实验代码

以MNIST做实验

1. 导入相关包

from __future__ import print_function, division

from keras.datasets import mnist
from keras.layers import Input, Dense, Reshape, Flatten, Dropout, multiply, GaussianNoise
from keras.layers import BatchNormalization, Activation, Embedding, ZeroPadding2D
from keras.layers.advanced_activations import LeakyReLU
from keras.layers.convolutional import UpSampling2D, Conv2D
from keras.models import Sequential, Model
from keras.optimizers import Adam
from keras import losses
from keras.utils import to_categorical
import keras.backend as K

import matplotlib.pyplot as plt

import numpy as np

2. 初始化超参数

class SGAN():
    def __init__(self):
        self.img_rows = 28
        self.img_cols = 28
        self.channels = 1
        self.img_shape = (self.img_rows, self.img_cols, self.channels)
        self.num_classes = 10
        self.latent_dim = 100

        optimizer = Adam(0.0002, 0.5)

3.声明生成器和分类器，并编译

        # Build and compile the discriminator
        self.discriminator = self.build_discriminator()
        self.discriminator.compile(loss=['binary_crossentropy', 'categorical_crossentropy'],
            loss_weights=[0.5, 0.5],
            optimizer=optimizer,
            metrics=['accuracy'])

        # Build the generator
        self.generator = self.build_generator()

        # The generator takes noise as input and generates imgs
        noise = Input(shape=(100,))
        img = self.generator(noise)

        # For the combined model we will only train the generator
        self.discriminator.trainable = False

        # The valid takes generated images as input and determines valid
        valid, _ = self.discriminator(img)

        # The combined model  (stacked generator and discriminator)
        # Trains generator to fool discriminator
        self.combined = Model(noise , valid)
        self.combined.compile(loss=['binary_crossentropy'],
            optimizer=optimizer)

4. 构造生成器和分类器

这里的分类器在最后一步使用softmax，并且label作为条件输入。

        features = model(img)
        valid = Dense(1, activation="sigmoid")(features)
        label = Dense(self.num_classes+1, activation="softmax")(features)

        return Model(img, [valid, label])

    def build_generator(self):

        model = Sequential()

        model.add(Dense(128 * 7 * 7, activation="relu", input_dim=self.latent_dim))
        model.add(Reshape((7, 7, 128)))
        model.add(BatchNormalization(momentum=0.8))
        model.add(UpSampling2D())
        model.add(Conv2D(128, kernel_size=3, padding="same"))
        model.add(Activation("relu"))
        model.add(BatchNormalization(momentum=0.8))
        model.add(UpSampling2D())
        model.add(Conv2D(64, kernel_size=3, padding="same"))
        model.add(Activation("relu"))
        model.add(BatchNormalization(momentum=0.8))
        model.add(Conv2D(1, kernel_size=3, padding="same"))
        model.add(Activation("tanh"))

        model.summary()

        noise = Input(shape=(self.latent_dim,))
        img = model(noise)

        return Model(noise, img)

    def build_discriminator(self):

        model = Sequential()

        model.add(Conv2D(32, kernel_size=3, strides=2, input_shape=self.img_shape, padding="same"))
        model.add(LeakyReLU(alpha=0.2))
        model.add(Dropout(0.25))
        model.add(Conv2D(64, kernel_size=3, strides=2, padding="same"))
        model.add(ZeroPadding2D(padding=((0,1),(0,1))))
        model.add(LeakyReLU(alpha=0.2))
        model.add(Dropout(0.25))
        model.add(BatchNormalization(momentum=0.8))
        model.add(Conv2D(128, kernel_size=3, strides=2, padding="same"))
        model.add(LeakyReLU(alpha=0.2))
        model.add(Dropout(0.25))
        model.add(BatchNormalization(momentum=0.8))
        model.add(Conv2D(256, kernel_size=3, strides=1, padding="same"))
        model.add(LeakyReLU(alpha=0.2))
        model.add(Dropout(0.25))
        model.add(Flatten())

        model.summary()

        img = Input(shape=self.img_shape)

        features = model(img)
        valid = Dense(1, activation="sigmoid")(features)
        label = Dense(self.num_classes+1, activation="softmax")(features)

        return Model(img, [valid, label])

 

5. 训练

    def train(self, epochs, batch_size=128, sample_interval=50):

        # Load the dataset
        (X_train, y_train), (_, _) = mnist.load_data()

        # Rescale -1 to 1
        X_train = (X_train.astype(np.float32) - 127.5) / 127.5
        X_train = np.expand_dims(X_train, axis=3)
        y_train = y_train.reshape(-1, 1)

        # Class weights:
        # To balance the difference in occurences of digit class labels.
        # 50% of labels that the discriminator trains on are 'fake'.
        # Weight = 1 / frequency
        cw1 = {0: 1, 1: 1}
        cw2 = {i: self.num_classes / 64 for i in range(self.num_classes)}
        cw2[self.num_classes] = 1 / 64

        # Adversarial ground truths
        valid = np.ones((batch_size, 1))
        fake = np.zeros((batch_size, 1))

        for epoch in range(epochs):

            # ---------------------
            #  Train Discriminator
            # ---------------------

            # Select a random batch of images
            idx = np.random.randint(0, X_train.shape[0], batch_size)
            imgs = X_train[idx]

            # Sample noise and generate a batch of new images
            noise = np.random.normal(0, 1, (batch_size, self.latent_dim))
            gen_imgs = self.generator.predict(noise)

            # One-hot encoding of labels
            labels = to_categorical(y_train[idx], num_classes=self.num_classes+1)
            fake_labels = to_categorical(np.full((batch_size, 1), self.num_classes), num_classes=self.num_classes+1)

            # Train the discriminator
            d_loss_real = self.discriminator.train_on_batch(imgs, [valid, labels], class_weight=[cw1, cw2])
            d_loss_fake = self.discriminator.train_on_batch(gen_imgs, [fake, fake_labels], class_weight=[cw1, cw2])
            d_loss = 0.5 * np.add(d_loss_real, d_loss_fake)


            # ---------------------
            #  Train Generator
            # ---------------------

            g_loss = self.combined.train_on_batch(noise, valid, class_weight=[cw1, cw2])

            # Plot the progress
            print ("%d [D loss: %f, acc: %.2f%%, op_acc: %.2f%%] [G loss: %f]" % (epoch, d_loss[0], 100*d_loss[3], 100*d_loss[4], g_loss))

            # If at save interval => save generated image samples
            if epoch % sample_interval == 0:
                self.sample_images(epoch)

 

6. 可视化 并保存model

    def sample_images(self, epoch):
        r, c = 5, 5
        noise = np.random.normal(0, 1, (r * c, self.latent_dim))
        gen_imgs = self.generator.predict(noise)

        # Rescale images 0 - 1
        gen_imgs = 0.5 * gen_imgs + 1

        fig, axs = plt.subplots(r, c)
        cnt = 0
        for i in range(r):
            for j in range(c):
                axs[i,j].imshow(gen_imgs[cnt, :,:,0], cmap='gray')
                axs[i,j].axis('off')
                cnt += 1
        fig.savefig("images/mnist_%d.png" % epoch)
        plt.close()

    def save_model(self):

        def save(model, model_name):
            model_path = "saved_model/%s.json" % model_name
            weights_path = "saved_model/%s_weights.hdf5" % model_name
            options = {"file_arch": model_path,
                        "file_weight": weights_path}
            json_string = model.to_json()
            open(options['file_arch'], 'w').write(json_string)
            model.save_weights(options['file_weight'])

        save(self.generator, "mnist_sgan_generator")
        save(self.discriminator, "mnist_sgan_discriminator")
        save(self.combined, "mnist_sgan_adversarial")


if __name__ == '__main__':
    sgan = SGAN()
    sgan.train(epochs=20000, batch_size=32, sample_interval=50)

 

四、实验结果

0 [D loss: 0.579261, acc: 39.06%, op_acc: 7.81%] [G loss: 0.713937]

1000 [D loss: 0.316801, acc: 71.88%, op_acc: 54.69%] [G loss: 1.037289]

2000 [D loss: 0.366537, acc: 59.38%, op_acc: 54.69%] [G loss: 0.987430]

3000 [D loss: 0.328151, acc: 67.19%, op_acc: 57.81%] [G loss: 1.385562]

4000 [D loss: 0.343711, acc: 68.75%, op_acc: 57.81%] [G loss: 1.151738]

5000 [D loss: 0.348945, acc: 65.62%, op_acc: 56.25%] [G loss: 1.219509]

6000 [D loss: 0.318236, acc: 67.19%, op_acc: 54.69%] [G loss: 1.235397]

7000 [D loss: 0.368775, acc: 59.38%, op_acc: 51.56%] [G loss: 1.175400]

8000 [D loss: 0.340480, acc: 67.19%, op_acc: 59.38%] [G loss: 1.065718]

 

完整代码

from __future__ import print_function, division

from keras.datasets import mnist
from keras.layers import Input, Dense, Reshape, Flatten, Dropout, multiply, GaussianNoise
from keras.layers import BatchNormalization, Activation, Embedding, ZeroPadding2D
from keras.layers.advanced_activations import LeakyReLU
from keras.layers.convolutional import UpSampling2D, Conv2D
from keras.models import Sequential, Model
from keras.optimizers import Adam
from keras import losses
from keras.utils import to_categorical
import keras.backend as K

import matplotlib.pyplot as plt

import numpy as np

class SGAN():
    def __init__(self):
        self.img_rows = 28
        self.img_cols = 28
        self.channels = 1
        self.img_shape = (self.img_rows, self.img_cols, self.channels)
        self.num_classes = 10
        self.latent_dim = 100

        optimizer = Adam(0.0002, 0.5)

        # Build and compile the discriminator
        self.discriminator = self.build_discriminator()
        self.discriminator.compile(loss=['binary_crossentropy', 'categorical_crossentropy'],
            loss_weights=[0.5, 0.5],
            optimizer=optimizer,
            metrics=['accuracy'])

        # Build the generator
        self.generator = self.build_generator()

        # The generator takes noise as input and generates imgs
        noise = Input(shape=(100,))
        img = self.generator(noise)

        # For the combined model we will only train the generator
        self.discriminator.trainable = False

        # The valid takes generated images as input and determines valid
        valid, _ = self.discriminator(img)

        # The combined model  (stacked generator and discriminator)
        # Trains generator to fool discriminator
        self.combined = Model(noise , valid)
        self.combined.compile(loss=['binary_crossentropy'],
            optimizer=optimizer)


    def build_generator(self):

        model = Sequential()

        model.add(Dense(128 * 7 * 7, activation="relu", input_dim=self.latent_dim))
        model.add(Reshape((7, 7, 128)))
        model.add(BatchNormalization(momentum=0.8))
        model.add(UpSampling2D())
        model.add(Conv2D(128, kernel_size=3, padding="same"))
        model.add(Activation("relu"))
        model.add(BatchNormalization(momentum=0.8))
        model.add(UpSampling2D())
        model.add(Conv2D(64, kernel_size=3, padding="same"))
        model.add(Activation("relu"))
        model.add(BatchNormalization(momentum=0.8))
        model.add(Conv2D(1, kernel_size=3, padding="same"))
        model.add(Activation("tanh"))

        model.summary()

        noise = Input(shape=(self.latent_dim,))
        img = model(noise)

        return Model(noise, img)

    def build_discriminator(self):

        model = Sequential()

        model.add(Conv2D(32, kernel_size=3, strides=2, input_shape=self.img_shape, padding="same"))
        model.add(LeakyReLU(alpha=0.2))
        model.add(Dropout(0.25))
        model.add(Conv2D(64, kernel_size=3, strides=2, padding="same"))
        model.add(ZeroPadding2D(padding=((0,1),(0,1))))
        model.add(LeakyReLU(alpha=0.2))
        model.add(Dropout(0.25))
        model.add(BatchNormalization(momentum=0.8))
        model.add(Conv2D(128, kernel_size=3, strides=2, padding="same"))
        model.add(LeakyReLU(alpha=0.2))
        model.add(Dropout(0.25))
        model.add(BatchNormalization(momentum=0.8))
        model.add(Conv2D(256, kernel_size=3, strides=1, padding="same"))
        model.add(LeakyReLU(alpha=0.2))
        model.add(Dropout(0.25))
        model.add(Flatten())

        model.summary()

        img = Input(shape=self.img_shape)

        features = model(img)
        valid = Dense(1, activation="sigmoid")(features)
        label = Dense(self.num_classes+1, activation="softmax")(features)

        return Model(img, [valid, label])

    def train(self, epochs, batch_size=128, sample_interval=50):

        # Load the dataset
        (X_train, y_train), (_, _) = mnist.load_data()

        # Rescale -1 to 1
        X_train = (X_train.astype(np.float32) - 127.5) / 127.5
        X_train = np.expand_dims(X_train, axis=3)
        y_train = y_train.reshape(-1, 1)

        # Class weights:
        # To balance the difference in occurences of digit class labels.
        # 50% of labels that the discriminator trains on are 'fake'.
        # Weight = 1 / frequency
        cw1 = {0: 1, 1: 1}
        cw2 = {i: self.num_classes / 64 for i in range(self.num_classes)}
        cw2[self.num_classes] = 1 / 64

        # Adversarial ground truths
        valid = np.ones((batch_size, 1))
        fake = np.zeros((batch_size, 1))

        for epoch in range(epochs):

            # ---------------------
            #  Train Discriminator
            # ---------------------

            # Select a random batch of images
            idx = np.random.randint(0, X_train.shape[0], batch_size)
            imgs = X_train[idx]

            # Sample noise and generate a batch of new images
            noise = np.random.normal(0, 1, (batch_size, self.latent_dim))
            gen_imgs = self.generator.predict(noise)

            # One-hot encoding of labels
            labels = to_categorical(y_train[idx], num_classes=self.num_classes+1)
            fake_labels = to_categorical(np.full((batch_size, 1), self.num_classes), num_classes=self.num_classes+1)

            # Train the discriminator
            d_loss_real = self.discriminator.train_on_batch(imgs, [valid, labels], class_weight=[cw1, cw2])
            d_loss_fake = self.discriminator.train_on_batch(gen_imgs, [fake, fake_labels], class_weight=[cw1, cw2])
            d_loss = 0.5 * np.add(d_loss_real, d_loss_fake)


            # ---------------------
            #  Train Generator
            # ---------------------

            g_loss = self.combined.train_on_batch(noise, valid, class_weight=[cw1, cw2])

            # Plot the progress
            print ("%d [D loss: %f, acc: %.2f%%, op_acc: %.2f%%] [G loss: %f]" % (epoch, d_loss[0], 100*d_loss[3], 100*d_loss[4], g_loss))

            # If at save interval => save generated image samples
            if epoch % sample_interval == 0:
                self.sample_images(epoch)

    def sample_images(self, epoch):
        r, c = 5, 5
        noise = np.random.normal(0, 1, (r * c, self.latent_dim))
        gen_imgs = self.generator.predict(noise)

        # Rescale images 0 - 1
        gen_imgs = 0.5 * gen_imgs + 1

        fig, axs = plt.subplots(r, c)
        cnt = 0
        for i in range(r):
            for j in range(c):
                axs[i,j].imshow(gen_imgs[cnt, :,:,0], cmap='gray')
                axs[i,j].axis('off')
                cnt += 1
        fig.savefig("images/mnist_%d.png" % epoch)
        plt.close()

    def save_model(self):

        def save(model, model_name):
            model_path = "saved_model/%s.json" % model_name
            weights_path = "saved_model/%s_weights.hdf5" % model_name
            options = {"file_arch": model_path,
                        "file_weight": weights_path}
            json_string = model.to_json()
            open(options['file_arch'], 'w').write(json_string)
            model.save_weights(options['file_weight'])

        save(self.generator, "mnist_sgan_generator")
        save(self.discriminator, "mnist_sgan_discriminator")
        save(self.combined, "mnist_sgan_adversarial")


if __name__ == '__main__':
    sgan = SGAN()
    sgan.train(epochs=20000, batch_size=32, sample_interval=50)