[TensorFlow] DCGAN으로 MNIST 이미지 생성하기 (최종)

지식 부족으로 계속 실패했었는데 그럭저럭 완성됬습니다.

인터넷 블로그의 글들과 GitHub의 소스코드들을 참고해서 만들었습니다.

코드:

#-*- coding: utf-8 -*-
 
import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
 
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("./mnist/data/", one_hot=True)
 
total_epoch = 100
batch_size = 100
n_noise = 100
 
D_global_step = tf.Variable(0, trainable=False, name='D_global_step')
G_global_step = tf.Variable(0, trainable=False, name='G_global_step')
 
X = tf.placeholder(tf.float32, [None, 28, 28, 1])
Z = tf.placeholder(tf.float32, [None, n_noise])
is_training = tf.placeholder(tf.bool)
 
def leaky_relu(x, leak=0.2):
    return tf.maximum(x, x * leak)
 
def generator(noise):
    with tf.variable_scope('generator'):
        output = tf.layers.dense(noise, 128*7*7)
        output = tf.reshape(output, [-1, 7, 7, 128])
        output = tf.nn.relu(tf.layers.batch_normalization(output, training=is_training))
        output = tf.layers.conv2d_transpose(output, 64, [5, 5], strides=(2, 2), padding='SAME')
        output = tf.nn.relu(tf.layers.batch_normalization(output, training=is_training))
        output = tf.layers.conv2d_transpose(output, 32, [5, 5], strides=(2, 2), padding='SAME')
        output = tf.nn.relu(tf.layers.batch_normalization(output, training=is_training))
        output = tf.layers.conv2d_transpose(output, 1, [5, 5], strides=(1, 1), padding='SAME')
        output = tf.tanh(output)
    return output
 
def discriminator(inputs, reuse=None):
    with tf.variable_scope('discriminator') as scope:
        if reuse:
            scope.reuse_variables()
        output = tf.layers.conv2d(inputs, 32, [5, 5], strides=(2, 2), padding='SAME')
        output = leaky_relu(output)
        output = tf.layers.conv2d(output, 64, [5, 5], strides=(2, 2), padding='SAME')
        output = leaky_relu(tf.layers.batch_normalization(output, training=is_training))
        output = tf.layers.conv2d(output, 128, [5, 5], strides=(2, 2), padding='SAME')
        output = leaky_relu(tf.layers.batch_normalization(output, training=is_training))
        flat = tf.contrib.layers.flatten(output)
        output = tf.layers.dense(flat, 1, activation=None)
    return output
 
def get_noise(batch_size, n_noise):
    return np.random.uniform(-1.0, 1.0, size=[batch_size, n_noise])
 
def get_moving_noise(batch_size, n_noise):
    assert batch_size > 0
 
    noise_list = []
    base_noise = np.random.uniform(-1.0, 1.0, size=[n_noise])
    end_noise = np.random.uniform(-1.0, 1.0, size=[n_noise])
 
    step = (end_noise - base_noise) / batch_size
    noise = np.copy(base_noise)
    for _ in range(batch_size - 1):
        noise_list.append(noise)
        noise = noise + step
    noise_list.append(end_noise)
    
    return noise_list
 
G = generator(Z)
D_real = discriminator(X)
D_gene = discriminator(G, True)
 
loss_D_real = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(
    logits=D_real, labels=tf.ones_like(D_real)
))
loss_D_gene = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(
    logits=D_gene, labels=tf.zeros_like(D_gene)
))
 
loss_D = loss_D_real + loss_D_gene
loss_G = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(
    logits=D_gene, labels=tf.ones_like(D_gene)
))
 
vars_D = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
    scope='discriminator')
vars_G = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
    scope='generator')
 
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
    train_D = tf.train.AdamOptimizer().minimize(loss_D,
        var_list=vars_D, global_step=D_global_step)
    train_G = tf.train.AdamOptimizer().minimize(loss_G,
        var_list=vars_G, global_step=G_global_step)
 
tf.summary.scalar('loss_D', loss_D)
tf.summary.scalar('loss_G', loss_G)
 
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
 
    merged = tf.summary.merge_all()
    writer = tf.summary.FileWriter('./logs', sess.graph)
 
    total_batch = int(mnist.train.num_examples / batch_size)
 
    for epoch in range(total_epoch):
        loss_val_D, loss_val_G = 0, 0
 
        batch_xs, batch_ys = None, None
        noise = None
 
        for i in range(total_batch):
            batch_xs, batch_ys = mnist.train.next_batch(batch_size)
            batch_xs = batch_xs.reshape(-1, 28, 28, 1)
            noise = get_noise(batch_size, n_noise)
 
            _, loss_val_D = sess.run([train_D, loss_D],
                feed_dict={X: batch_xs, Z: noise, is_training: True})
            _, loss_val_G = sess.run([train_G, loss_G],
                feed_dict={X: batch_xs, Z: noise, is_training: True})
 
        summary = sess.run(merged,
            feed_dict={X: batch_xs, Z: noise, is_training: True})
        writer.add_summary(summary, global_step=sess.run(G_global_step))
 
        print('Epoch:', '%04d' % epoch,
            'D loss: {:.4}'.format(loss_val_D),
            'G loss: {:.4}'.format(loss_val_G))
 
        if epoch == 0 or (epoch + 1) % 5 == 0:
            sample_size = 10
            noise = get_noise(sample_size, n_noise)
            samples = sess.run(G, feed_dict={Z: noise, is_training: False})
            test_noise = get_moving_noise(sample_size, n_noise)
            test_samples = sess.run(G, feed_dict={Z: test_noise, is_training: False})
 
            fig, ax = plt.subplots(2, sample_size, figsize=(sample_size, 2))
 
            for i in range(sample_size):
                ax[0][i].set_axis_off()
                ax[1][i].set_axis_off()
                ax[0][i].imshow(np.reshape(samples[i], (28, 28)))
                ax[1][i].imshow(np.reshape(test_samples[i], (28, 28)))
 
            plt.savefig('{}.png'.format(str(epoch).zfill(3)),
                bbox_inches='tight')
            
            plt.close(fig)
 

결과:

아래는 생성한 결과 이미지 입니다.

첫째줄은 단순하게 랜덤 노이즈로 생성한 결과물이고

둘째줄은 노이즈를 두개(a, b) 만들고 자연스럽게 a에서 b로 노이즈 값을 변화시키면서 만든 결과물 입니다.

DCGAN의 장점 중 하나가 입력에 대해 오버피팅되지 않는다는 거라서

학습이 잘 되었다면 노이즈를 변화시키면 결과물도 자연스럽게 바뀐다고 하니 그걸 보려고 두번째 줄을 만들었습니다.

학습 초기의 결과는 뺐습니다.

GPU는 GTX 960M을 사용했는데 100 epochs에 4시간 걸렸네요. ㅠㅠ

왜 고성능 GPU가 딸린 ML 전용 클라우드 서비스를 쓰는지 알 것 같습니다....

아무튼 만족스러운 결과입니다!

딥러닝을 배우고는 있지만 정작 실용적인 무언가를 아직 못만들어봐서 이렇게라도 만족하고 있습니다.

언젠가 노하우가 늘고 창의적인 아이디어가 떠오른다면 열정을 불태워서 진행해보고 싶습니다.