cogan.py

from __future__ import print_function, division
import scipy

from keras.datasets import mnist
from keras.layers import Input, Dense, Reshape, Flatten, Dropout
from keras.layers import BatchNormalization, Activation, 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

import matplotlib.pyplot as plt

import sys

import numpy as np

class COGAN():
    """Reference: https://wiseodd.github.io/techblog/2017/02/18/coupled_gan/"""
    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)

        optimizer = Adam(0.0002, 0.5)

        # Build and compile the discriminator
        self.d1, self.d2 = self.build_discriminators()
        self.d1.compile(loss='binary_crossentropy',
            optimizer=optimizer,
            metrics=['accuracy'])
        self.d2.compile(loss='binary_crossentropy',
            optimizer=optimizer,
            metrics=['accuracy'])

        # Build and compile the generator
        self.g1, self.g2 = self.build_generators()
        self.g1.compile(loss='binary_crossentropy', optimizer=optimizer)
        self.g2.compile(loss='binary_crossentropy', optimizer=optimizer)

        # The generator takes noise as input and generated imgs
        z = Input(shape=(100,))
        img1 = self.g1(z)
        img2 = self.g2(z)

        # For the combined model we will only train the generators
        self.d1.trainable = False
        self.d2.trainable = False

        # The valid takes generated images as input and determines validity
        valid1 = self.d1(img1)
        valid2 = self.d2(img2)

        # The combined model  (stacked generators and discriminators) takes
        # noise as input => generates images => determines validity
        self.combined = Model(z, [valid1, valid2])
        self.combined.compile(loss=['binary_crossentropy', 'binary_crossentropy'],
                                    optimizer=optimizer)

    def build_generators(self):

        noise_shape = (100,)
        noise = Input(shape=noise_shape)

        # Shared weights between generators
        model = Sequential()
        model.add(Dense(256, input_shape=noise_shape))
        model.add(LeakyReLU(alpha=0.2))
        model.add(BatchNormalization(momentum=0.8))
        model.add(Dense(512))
        model.add(LeakyReLU(alpha=0.2))
        model.add(BatchNormalization(momentum=0.8))

        latent = model(noise)

        # Generator 1
        g1 = Dense(1024)(latent)
        g1 = LeakyReLU(alpha=0.2)(g1)
        g1 = BatchNormalization(momentum=0.8)(g1)
        g1 = Dense(np.prod(self.img_shape), activation='tanh')(g1)
        img1 = Reshape(self.img_shape)(g1)

        # Generator 2
        g2 = Dense(1024)(latent)
        g2 = LeakyReLU(alpha=0.2)(g2)
        g2 = BatchNormalization(momentum=0.8)(g2)
        g2 = Dense(np.prod(self.img_shape), activation='tanh')(g2)
        img2 = Reshape(self.img_shape)(g2)

        model.summary()

        return Model(noise, img1), Model(noise, img2)

    def build_discriminators(self):

        img_shape = (self.img_rows, self.img_cols, self.channels)
        img1 = Input(shape=img_shape)
        img2 = Input(shape=img_shape)

        # Shared discriminator layers
        model = Sequential()
        model.add(Flatten(input_shape=img_shape))
        model.add(Dense(512))
        model.add(LeakyReLU(alpha=0.2))
        model.add(Dense(256))
        model.add(LeakyReLU(alpha=0.2))

        img1_embedding = model(img1)
        img2_embedding = model(img2)

        # Discriminator 1
        validity1 = Dense(1, activation='sigmoid')(img1_embedding)
        # Discriminator 2
        validity2 = Dense(1, activation='sigmoid')(img2_embedding)

        return Model(img1, validity1), Model(img2, validity2)

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

        # Load the dataset
        (X_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)

        # Images in domain A and B (rotated)
        X1 = X_train[:int(X_train.shape[0]/2)]
        X2 = X_train[int(X_train.shape[0]/2):]
        X2 = scipy.ndimage.interpolation.rotate(X2, 90, axes=(1, 2))

        half_batch = int(batch_size / 2)

        for epoch in range(epochs):

            # ----------------------
            #  Train Discriminators
            # ----------------------

            # Select a random half batch of images
            idx = np.random.randint(0, X1.shape[0], half_batch)
            imgs1 = X1[idx]
            imgs2 = X2[idx]

            noise = np.random.normal(0, 1, (half_batch, 100))

            # Generate a half batch of new images
            gen_imgs1 = self.g1.predict(noise)
            gen_imgs2 = self.g2.predict(noise)

            # Train the discriminators
            d1_loss_real = self.d1.train_on_batch(imgs1, np.ones((half_batch, 1)))
            d2_loss_real = self.d2.train_on_batch(imgs2, np.ones((half_batch, 1)))
            d1_loss_fake = self.d1.train_on_batch(gen_imgs1, np.zeros((half_batch, 1)))
            d2_loss_fake = self.d2.train_on_batch(gen_imgs2, np.zeros((half_batch, 1)))
            d1_loss = 0.5 * np.add(d1_loss_real, d1_loss_fake)
            d2_loss = 0.5 * np.add(d2_loss_real, d2_loss_fake)


            # ------------------
            #  Train Generators
            # ------------------

            noise = np.random.normal(0, 1, (batch_size, 100))

            # The generators wants the discriminators to label the generated samples
            # as valid (ones)
            valid = np.array([1] * batch_size)

            # Train the generators
            g_loss = self.combined.train_on_batch(noise, [valid, valid])

            # Plot the progress
            print ("%d [D1 loss: %f, acc.: %.2f%%] [D2 loss: %f, acc.: %.2f%%] [G loss: %f]" \
                % (epoch, d1_loss[0], 100*d1_loss[1], d2_loss[0], 100*d2_loss[1], g_loss[0]))

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

    def sample_images(self, epoch):
        r, c = 4, 4
        noise = np.random.normal(0, 1, (r * int(c/2), 100))
        gen_imgs1 = self.g1.predict(noise)
        gen_imgs2 = self.g2.predict(noise)

        gen_imgs = np.concatenate([gen_imgs1, gen_imgs2])

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

        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()


if __name__ == '__main__':
    gan = COGAN()
    gan.train(epochs=30000, batch_size=32, sample_interval=200)