cifar_pca_tf_gan.py

import os, time, itertools, imageio, pickle
import numpy as np
import matplotlib.pyplot as plt
import tensorflow as tf
from tflearn.datasets import cifar10
from tflearn.data_utils import shuffle, to_categorical
from sklearn.decomposition import PCA
from utils import *
import pdb

global fixed_sample_batch

# training parameters
batch_size = 100
lr = 0.0002
train_epoch = 400

(train_set, Y), _ = cifar10.load_data()
train_set, Y = shuffle(train_set, Y)
# Mix 'er up
np.random.shuffle(train_set)
# Samples to reconstruct in pca visualization
fixed_sample_batch = train_set[:25]

train_set = np.reshape(train_set, (-1, 32*32*3))

pca = PCA(n_components=300)
pca.fit(train_set)

train_set = np.reshape(train_set, (-1, 32, 32, 3))


# G(z)
def generator(x, isTrain=True, reuse=False):
    with tf.variable_scope('generator', reuse=reuse):

        # 1st hidden layer
        conv1 = tf.layers.conv2d_transpose(x, 128, [5, 5], strides=(2, 2), padding='same')
        lrelu1 = lrelu(tf.layers.batch_normalization(conv1, training=isTrain), 0.2)

        # 2nd hidden layer
        conv2 = tf.layers.conv2d_transpose(lrelu1, 64, [5, 5], strides=(2, 2), padding='same')
        lrelu2 = lrelu(tf.layers.batch_normalization(conv2, training=isTrain), 0.2)

        # 3rd hidden layer
        conv3 = tf.layers.conv2d_transpose(lrelu2, 32, [5, 5], strides=(2, 2), padding='same')
        lrelu3 = lrelu(tf.layers.batch_normalization(conv3, training=isTrain), 0.2)

        # 4th hidden layer
        conv4 = tf.layers.conv2d_transpose(lrelu3, 16, [5, 5], strides=(2, 2), padding='same')
        lrelu4 = lrelu(tf.layers.batch_normalization(conv4, training=isTrain), 0.2)

        # output layer
        conv5 = tf.layers.conv2d_transpose(lrelu4, 3, [5, 5], strides=(2, 2), padding='same')
        o = tf.nn.tanh(conv5)

        return o

# D(x)
def discriminator(x, isTrain=True, reuse=False):
    with tf.variable_scope('discriminator', reuse=reuse):
        # 1st hidden layer
        conv1 = tf.layers.conv2d(x, 16, [5, 5], strides=(2, 2), padding='same')
        lrelu1 = lrelu(conv1, 0.2)

        # 2nd hidden layer
        conv2 = tf.layers.conv2d(lrelu1, 32, [5, 5], strides=(2, 2), padding='same')
        lrelu2 = lrelu(tf.layers.batch_normalization(conv2, training=isTrain), 0.2)

        # 3rd hidden layer
        conv3 = tf.layers.conv2d(lrelu2, 64, [5, 5], strides=(2, 2), padding='same')
        lrelu3 = lrelu(tf.layers.batch_normalization(conv3, training=isTrain), 0.2)

        conv4 = tf.layers.conv2d(lrelu3, 128, [5, 5], strides=(2, 2), padding='same')
        lrelu4 = lrelu(tf.layers.batch_normalization(conv4, training=isTrain), 0.2)

        # output layer
        conv5 = tf.layers.conv2d(lrelu4, 1, [5, 5], strides=(2, 2), padding='same')
        o = tf.nn.sigmoid(conv5)

        return o, conv5

fixed_z_ = np.random.normal(0, 1, (25, 1, 1, 100))
def show_result(num_epoch, show = False, save = False, path = 'result.png'):
    test_images = sess.run(G_z, {z: fixed_z_, isTrain: False})
    # test_images = test_images + abs(np.min(test_images))
    test_images =  (test_images - np.max(test_images))/-np.ptp(test_images)

    size_figure_grid = 5
    fig, ax = plt.subplots(size_figure_grid, size_figure_grid, figsize=(5, 5))
    for i, j in itertools.product(range(size_figure_grid), range(size_figure_grid)):
        ax[i, j].get_xaxis().set_visible(False)
        ax[i, j].get_yaxis().set_visible(False)

    for k in range(size_figure_grid*size_figure_grid):
        i = k // size_figure_grid
        j = k % size_figure_grid
        ax[i, j].cla()
        ax[i, j].imshow(np.reshape(test_images[k], (32, 32, 3)))#, cmap='gray')

    label = 'Epoch {0}'.format(num_epoch)
    fig.text(0.5, 0.04, label, ha='center')

    if save:
        plt.savefig(path)

    if show:
        plt.show()
    else:
        plt.close()

def show_pca(num_pcs, path = 'pca.png', fixed_sample_batch=train_set[:25]):
    size_figure_grid = 5
    fig, ax = plt.subplots(size_figure_grid, size_figure_grid, figsize=(5, 5))

    fixed_sample_batch = np.reshape(fixed_sample_batch, (-1, 32*32*3))

    pca_samples = pca_reconstruct(pca, fixed_sample_batch, num_pcs)

    fixed_sample_batch = np.reshape(fixed_sample_batch, (-1, 32, 32, 3))
    # fixed_sample_batch = fixed_sample_batch + abs(np.min(fixed_sample_batch))
    fixed_sample_batch =  (fixed_sample_batch - np.max(fixed_sample_batch))/-np.ptp(fixed_sample_batch)
    # Remove tiks
    for i, j in itertools.product(range(size_figure_grid), range(size_figure_grid)):
        ax[i, j].get_xaxis().set_visible(False)
        ax[i, j].get_yaxis().set_visible(False)

    # Fill images
    for k in range(5*5):
        i = k // 5
        j = k % 5
        ax[i, j].cla()
        ax[i, j].imshow(np.reshape(pca_samples[k], (32, 32, 3)))

    if num_pcs < 300:
        label = 'Principle Components: {0}'.format(num_pcs)
    else:
        label = 'Target Dataset Sample'
    fig.text(0.5, 0.004, label, ha='center')
    plt.savefig(path)


def show_train_hist(hist, show = False, save = False, path = 'Train_hist.png'):
    x = range(len(hist['D_losses']))

    y1 = hist['D_losses']
    y2 = hist['G_losses']

    plt.plot(x, y1, label='D_loss')
    plt.plot(x, y2, label='G_loss')

    plt.xlabel('Epoch')
    plt.ylabel('Loss')

    plt.legend(loc=4)
    plt.grid(True)
    plt.tight_layout()

    if save:
        plt.savefig(path)

    if show:
        plt.show()
    else:
        plt.close()

# variables : input
x = tf.placeholder(tf.float32, shape=(None, 32, 32, 3))
z = tf.placeholder(tf.float32, shape=(None, 1, 1, 100))
isTrain = tf.placeholder(dtype=tf.bool)

# networks : generator
G_z = generator(z, isTrain)

# networks : discriminator
D_real, D_real_logits = discriminator(x, isTrain)
D_fake, D_fake_logits = discriminator(G_z, isTrain, reuse=True)

# loss for each network
D_loss_real = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=D_real_logits, labels=tf.ones([batch_size, 1, 1, 1])))
D_loss_fake = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=D_fake_logits, labels=tf.zeros([batch_size, 1, 1, 1])))
D_loss = D_loss_real + D_loss_fake
G_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=D_fake_logits, labels=tf.ones([batch_size, 1, 1, 1])))

# trainable variables for each network
T_vars = tf.trainable_variables()
D_vars = [var for var in T_vars if var.name.startswith('discriminator')]
G_vars = [var for var in T_vars if var.name.startswith('generator')]

# optimizer for each network
with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
    D_optim = tf.train.AdamOptimizer(lr, beta1=0.5).minimize(D_loss, var_list=D_vars)
    G_optim = tf.train.AdamOptimizer(lr, beta1=0.5).minimize(G_loss, var_list=G_vars)

# open session and initialize all variables
sess = tf.InteractiveSession()
tf.global_variables_initializer().run()


# results save folder
root = 'CIFAR_PCA_DCGAN_results4/'
model = 'CIFAR_PCA_DCGAN4_'
if not os.path.isdir(root):
    os.mkdir(root)
if not os.path.isdir(root + 'Fixed_results'):
    os.mkdir(root + 'Fixed_results')
if not os.path.isdir(root + 'PCA_IMG'):
    os.mkdir(root + 'PCA_IMG')

train_hist = {}
train_hist['D_losses'] = []
train_hist['G_losses'] = []
train_hist['per_epoch_ptimes'] = []
train_hist['total_ptime'] = []

# training-loop
np.random.seed(int(time.time()))
print('training start!')
start_time = time.time()
for epoch in range(train_epoch):
    G_losses = []
    D_losses = []
    epoch_start_time = time.time()
    for iter in range(train_set.shape[0] // batch_size):
        # update discriminator
        x_ = train_set[iter*batch_size:(iter+1)*batch_size]
        x_ = np.reshape(x_, (-1, 32*32*3))
        x_ = pca_reconstruct(pca, x_, int(epoch))
        x_ = np.reshape(x_, (-1, 32, 32, 3))

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

        loss_d_, _ = sess.run([D_loss, D_optim], {x: x_, z: z_, isTrain: True})
        D_losses.append(loss_d_)

        # update generator
        z_ = np.random.normal(0, 1, (batch_size, 1, 1, 100))
        loss_g_, _ = sess.run([G_loss, G_optim], {z: z_, x: x_, isTrain: True})
        G_losses.append(loss_g_)

    epoch_end_time = time.time()
    per_epoch_ptime = epoch_end_time - epoch_start_time
    print('[%d/%d] - ptime: %.2f loss_d: %.3f, loss_g: %.3f' % ((epoch + 1), train_epoch, per_epoch_ptime, np.mean(D_losses), np.mean(G_losses)))
    pca_path = root + 'PCA_IMG/' + str(epoch + 1) + '.png'
    fixed_p = root + 'Fixed_results/' + model + str(epoch + 1) + '.png'
    show_result((epoch + 1), save=True, path=fixed_p)
    show_pca(epoch, pca_path, train_set[:25])
    train_hist['D_losses'].append(np.mean(D_losses))
    train_hist['G_losses'].append(np.mean(G_losses))
    train_hist['per_epoch_ptimes'].append(per_epoch_ptime)

end_time = time.time()
total_ptime = end_time - start_time
train_hist['total_ptime'].append(total_ptime)

print('Avg per epoch ptime: %.2f, total %d epochs ptime: %.2f' % (np.mean(train_hist['per_epoch_ptimes']), train_epoch, total_ptime))
print("Training finish!... save training results")
with open(root + model + 'train_hist.pkl', 'wb') as f:
    pickle.dump(train_hist, f)

show_train_hist(train_hist, save=True, path=root + model + 'train_hist.png')

images = []
for e in range(train_epoch):
    img_name = root + 'Fixed_results/' + model + str(e + 1) + '.png'
    images.append(imageio.imread(img_name))
imageio.mimsave(root + model + 'generation_animation.gif', images, fps=5)

sess.close()