[生成對抗網絡GAN入門指南]（10）InfoGAN: Interpretable Representation Learning by Information Maximizing GAN

本篇blog的內容基於原始論文InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets（NPIs2016）和《生成對抗網絡入門指南》第六章。完整代碼及簡析見文章末尾

---

一、爲什麼要使用InfoGAN

InfoGAN採用無監督的方式學習，並嘗試實現可解釋特徵。使用了信息論的原理，通過最大化輸入噪聲和觀察值之間的互信息（Mutual Information，MI）來對網絡模型進行優化。InfoGAN能適用於各種複雜的數據集，可以同時實現離散特徵和連續特徵。

---

二、輸入端數據

InfoGAN在輸入端把隨機輸入分爲兩個部分：

第一部分爲z，代表噪聲；

第二部分爲c，代表隱含編碼；

目標是希望在每個維度上都具備可解釋型特徵。

在同時輸入噪聲z和隱含編碼c後，生成概率 ，爲了應對這個問題，在InfoGAN中需要對隱含編碼c和生成分佈G(z,c)求互信息 ，並使其最大化

 

三、InfoGAN結構

InfoGAN和前面介紹過的GAN區別在於，真實訓練數據不有標籤數據，二輸入數據爲隱含編碼和隨機噪聲的組合，最後通過判別器一端和最大化互信息的方式還原隱含編碼的信息。也就是說，判別器D最終需要同時具備還原隱含編碼和辨別真僞的能力。前者爲了生成圖像能夠很好具備編碼中的特性，也就是說隱含編碼可以對生網絡產生相對顯著地成果；後者是要求生成模型在還原信息的同時保證生成的數據與真實數據非常逼近。

1. 互信息

互信息表示兩個隨機變量之間的依賴程度的度量。對於隨機變量X和Y，互信息爲I(X;Y)，H(X)和H(Y)爲邊緣熵，H(X|Y)和H(Y|X)爲條件熵。

2. 結構

 

3. 目標函數

當X和Y相互獨立時候，互信息爲0.給定任意的輸入，希望生成器的  有一個相對較小的熵，即希望隱含編碼c的信息在生成過程中不會流失。對此我們修改目標函數：

由於概率能以得到，導致互信息難以最大化，實際計算可以定義一個近似概率的輔助分佈來獲取互信息的下界，推導如下：

由此可以得到互信息的下界值：

                                       

 

4. InfoGAN的推導

我們可以重新改寫之前不等式，並重新使蒙特卡洛方法逼近

得到我們最終的目標函數

 

四、實驗效果

1.MNIST數據

我們發現通過控制隱含編碼中的可以調節生成數字是幾，其他參數可以調節生成字符的傾斜程度、字體寬度等。

 

2. 3D人臉數據

 

3. 椅子數據集

4. 門牌號數據集

 

五、實驗代碼

1. 導入相關包及超參數

from __future__ import print_function, division

from keras.datasets import mnist
from keras.layers import Input, Dense, Reshape, Flatten, Dropout, multiply, concatenate
from keras.layers import BatchNormalization, Activation, Embedding, ZeroPadding2D, Lambda
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.utils import to_categorical
import keras.backend as K

import matplotlib.pyplot as plt

import numpy as np

class INFOGAN():
    def __init__(self):
        self.img_rows = 28
        self.img_cols = 28
        self.channels = 1
        self.num_classes = 10
        self.img_shape = (self.img_rows, self.img_cols, self.channels)
        self.latent_dim = 72


        optimizer = Adam(0.0002, 0.5)
        losses = ['binary_crossentropy', self.mutual_info_loss]

        # Build and the discriminator and recognition network
        self.discriminator, self.auxilliary = self.build_disk_and_q_net()

        self.discriminator.compile(loss=['binary_crossentropy'],
            optimizer=optimizer,
            metrics=['accuracy'])

        # Build and compile the recognition network Q
        self.auxilliary.compile(loss=[self.mutual_info_loss],
            optimizer=optimizer,
            metrics=['accuracy'])

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

        # The generator takes noise and the target label as input
        # and generates the corresponding digit of that label
        gen_input = Input(shape=(self.latent_dim,))
        img = self.generator(gen_input)

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

        # The discriminator takes generated image as input and determines validity
        valid = self.discriminator(img)
        # The recognition network produces the label
        target_label = self.auxilliary(img)

        # The combined model  (stacked generator and discriminator)
        self.combined = Model(gen_input, [valid, target_label])
        self.combined.compile(loss=losses,
            optimizer=optimizer)

 

2. 構造生成器和判別器

    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(self.channels, kernel_size=3, padding='same'))
        model.add(Activation("tanh"))

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

        model.summary()

        return Model(gen_input, img)


    def build_disk_and_q_net(self):

        img = Input(shape=self.img_shape)

        # Shared layers between discriminator and recognition network
        model = Sequential()
        model.add(Conv2D(64, 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(128, 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(256, 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(512, 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(Flatten())

        img_embedding = model(img)

        # Discriminator
        validity = Dense(1, activation='sigmoid')(img_embedding)

        # Recognition
        q_net = Dense(128, activation='relu')(img_embedding)
        label = Dense(self.num_classes, activation='softmax')(q_net)

        # Return discriminator and recognition network
        return Model(img, validity), Model(img, label)

 

3. 構造互信息

    def mutual_info_loss(self, c, c_given_x):
        """The mutual information metric we aim to minimize"""
        eps = 1e-8
        conditional_entropy = K.mean(- K.sum(K.log(c_given_x + eps) * c, axis=1))
        entropy = K.mean(- K.sum(K.log(c + eps) * c, axis=1))

        return conditional_entropy + entropy

    def sample_generator_input(self, batch_size):
        # Generator inputs
        sampled_noise = np.random.normal(0, 1, (batch_size, 62))
        sampled_labels = np.random.randint(0, self.num_classes, batch_size).reshape(-1, 1)
        sampled_labels = to_categorical(sampled_labels, num_classes=self.num_classes)

        return sampled_noise, sampled_labels

 

4. 訓練

    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)

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

            # Sample noise and categorical labels
            sampled_noise, sampled_labels = self.sample_generator_input(batch_size)
            gen_input = np.concatenate((sampled_noise, sampled_labels), axis=1)

            # Generate a half batch of new images
            gen_imgs = self.generator.predict(gen_input)

            # Train on real and generated data
            d_loss_real = self.discriminator.train_on_batch(imgs, valid)
            d_loss_fake = self.discriminator.train_on_batch(gen_imgs, fake)

            # Avg. loss
            d_loss = 0.5 * np.add(d_loss_real, d_loss_fake)

            # ---------------------
            #  Train Generator and Q-network
            # ---------------------

            g_loss = self.combined.train_on_batch(gen_input, [valid, sampled_labels])

            # Plot the progress
            print ("%d [D loss: %.2f, acc.: %.2f%%] [Q loss: %.2f] [G loss: %.2f]" % (epoch, d_loss[0], 100*d_loss[1], g_loss[1], g_loss[2]))

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

 

5. 可視化

    def sample_images(self, epoch):
        r, c = 10, 10

        fig, axs = plt.subplots(r, c)
        for i in range(c):
            sampled_noise, _ = self.sample_generator_input(c)
            label = to_categorical(np.full(fill_value=i, shape=(r,1)), num_classes=self.num_classes)
            gen_input = np.concatenate((sampled_noise, label), axis=1)
            gen_imgs = self.generator.predict(gen_input)
            gen_imgs = 0.5 * gen_imgs + 0.5
            for j in range(r):
                axs[j,i].imshow(gen_imgs[j,:,:,0], cmap='gray')
                axs[j,i].axis('off')
        fig.savefig("images/%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, "generator")
        save(self.discriminator, "discriminator")


if __name__ == '__main__':
    infogan = INFOGAN()
    infogan.train(epochs=50000, batch_size=128, sample_interval=50)

 

實驗結果

 

完整代碼

from __future__ import print_function, division

from keras.datasets import mnist
from keras.layers import Input, Dense, Reshape, Flatten, Dropout, multiply, concatenate
from keras.layers import BatchNormalization, Activation, Embedding, ZeroPadding2D, Lambda
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.utils import to_categorical
import keras.backend as K

import matplotlib.pyplot as plt

import numpy as np

class INFOGAN():
    def __init__(self):
        self.img_rows = 28
        self.img_cols = 28
        self.channels = 1
        self.num_classes = 10
        self.img_shape = (self.img_rows, self.img_cols, self.channels)
        self.latent_dim = 72


        optimizer = Adam(0.0002, 0.5)
        losses = ['binary_crossentropy', self.mutual_info_loss]

        # Build and the discriminator and recognition network
        self.discriminator, self.auxilliary = self.build_disk_and_q_net()

        self.discriminator.compile(loss=['binary_crossentropy'],
            optimizer=optimizer,
            metrics=['accuracy'])

        # Build and compile the recognition network Q
        self.auxilliary.compile(loss=[self.mutual_info_loss],
            optimizer=optimizer,
            metrics=['accuracy'])

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

        # The generator takes noise and the target label as input
        # and generates the corresponding digit of that label
        gen_input = Input(shape=(self.latent_dim,))
        img = self.generator(gen_input)

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

        # The discriminator takes generated image as input and determines validity
        valid = self.discriminator(img)
        # The recognition network produces the label
        target_label = self.auxilliary(img)

        # The combined model  (stacked generator and discriminator)
        self.combined = Model(gen_input, [valid, target_label])
        self.combined.compile(loss=losses,
            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(self.channels, kernel_size=3, padding='same'))
        model.add(Activation("tanh"))

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

        model.summary()

        return Model(gen_input, img)


    def build_disk_and_q_net(self):

        img = Input(shape=self.img_shape)

        # Shared layers between discriminator and recognition network
        model = Sequential()
        model.add(Conv2D(64, 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(128, 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(256, 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(512, 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(Flatten())

        img_embedding = model(img)

        # Discriminator
        validity = Dense(1, activation='sigmoid')(img_embedding)

        # Recognition
        q_net = Dense(128, activation='relu')(img_embedding)
        label = Dense(self.num_classes, activation='softmax')(q_net)

        # Return discriminator and recognition network
        return Model(img, validity), Model(img, label)


    def mutual_info_loss(self, c, c_given_x):
        """The mutual information metric we aim to minimize"""
        eps = 1e-8
        conditional_entropy = K.mean(- K.sum(K.log(c_given_x + eps) * c, axis=1))
        entropy = K.mean(- K.sum(K.log(c + eps) * c, axis=1))

        return conditional_entropy + entropy

    def sample_generator_input(self, batch_size):
        # Generator inputs
        sampled_noise = np.random.normal(0, 1, (batch_size, 62))
        sampled_labels = np.random.randint(0, self.num_classes, batch_size).reshape(-1, 1)
        sampled_labels = to_categorical(sampled_labels, num_classes=self.num_classes)

        return sampled_noise, sampled_labels

    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)

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

            # Sample noise and categorical labels
            sampled_noise, sampled_labels = self.sample_generator_input(batch_size)
            gen_input = np.concatenate((sampled_noise, sampled_labels), axis=1)

            # Generate a half batch of new images
            gen_imgs = self.generator.predict(gen_input)

            # Train on real and generated data
            d_loss_real = self.discriminator.train_on_batch(imgs, valid)
            d_loss_fake = self.discriminator.train_on_batch(gen_imgs, fake)

            # Avg. loss
            d_loss = 0.5 * np.add(d_loss_real, d_loss_fake)

            # ---------------------
            #  Train Generator and Q-network
            # ---------------------

            g_loss = self.combined.train_on_batch(gen_input, [valid, sampled_labels])

            # Plot the progress
            print ("%d [D loss: %.2f, acc.: %.2f%%] [Q loss: %.2f] [G loss: %.2f]" % (epoch, d_loss[0], 100*d_loss[1], g_loss[1], g_loss[2]))

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

    def sample_images(self, epoch):
        r, c = 10, 10

        fig, axs = plt.subplots(r, c)
        for i in range(c):
            sampled_noise, _ = self.sample_generator_input(c)
            label = to_categorical(np.full(fill_value=i, shape=(r,1)), num_classes=self.num_classes)
            gen_input = np.concatenate((sampled_noise, label), axis=1)
            gen_imgs = self.generator.predict(gen_input)
            gen_imgs = 0.5 * gen_imgs + 0.5
            for j in range(r):
                axs[j,i].imshow(gen_imgs[j,:,:,0], cmap='gray')
                axs[j,i].axis('off')
        fig.savefig("images/%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, "generator")
        save(self.discriminator, "discriminator")


if __name__ == '__main__':
    infogan = INFOGAN()
    infogan.train(epochs=50000, batch_size=128, sample_interval=50)