합성곱 GAN

생성자와 판별자의 특성 맵은 다른 특성을 가진다.

생성자에서는 많은 특성 맵 개수에서 시작해서 마지막 층으로 갈 수록 줄어든다.

판별자는 작은 채널 개수로 시작해서 마지막 층으로 갈수록 늘어난다.

(특성맵의 개수와 특성 맵의 공간 차원의 크기가 반대, 특성 맵의 공간 크기가 증가하면, 특성 맵의 개수가 줄어든다.)

배치 정규화 앞의 층에서는 절편 유닛을 사용하지 않는 것이 보통 권장됨

Generator & Discriminator (DCGAN)

import tensorflow as tf

import numpy as np

def make_dcgan_generator(z_size=20, output_size=(28, 28, 1), n_filters=128, n_blocks=2):

size_factor=2**n_blocks

hidden_size=(output_size[0]//size_factor, output_size[1]//size_factor)

model=tf.keras.Sequential([

tf.keras.layers.Input(shape=(z_size, )),

tf.keras.layers.Dense(units=n_filters*np.prod(hidden_size), use_bias=False),

tf.keras.layers.BatchNormalization(),

tf.keras.layers.LeakyReLU(),

tf.keras.layers.Reshape((hidden_size[0], hidden_size[1], n_filters)),

tf.keras.layers.Conv2DTranspose(filters=n_filters, kernel_size=(5, 5), strides=(1, 1), padding='same', use_bias=False),

tf.keras.layers.BatchNormalization(),

tf.keras.layers.LeakyReLU()])

nf=n_filters

for i in range(n_blocks):

nf=nf//2

model.add(tf.keras.layers.Conv2DTranspose(

filters=nf, kernel_size=(5, 5), strides=(2, 2), padding='same', use_bias=False))

model.add(tf.keras.layers.BatchNormalization())

model.add(tf.keras.layers.LeakyReLU())

model.add(tf.keras.layers.Conv2DTranspose(

filters=output_size[2], kernel_size=(5, 5), strides=(1, 1), padding='same', use_bias=False, activation='tanh'))

return model

def make_dcgan_discriminator(input_size=(28, 28, 1), n_filters=64, n_blocks=2):

model=tf.keras.Sequential([

tf.keras.layers.Input(shape=input_size),

tf.keras.layers.Conv2D(filters=n_filters, kernel_size=5, strides=(1, 1), padding='same'),

tf.keras.layers.BatchNormalization(),

tf.keras.layers.LeakyReLU()])

nf=n_filters

for i in range(n_blocks):

nf=nf*2

model.add(tf.keras.layers.Conv2D(

filters=nf, kernel_size=(5, 5), strides=(2, 2), padding='same'))

model.add(tf.keras.layers.BatchNormalization())

model.add(tf.keras.layers.LeakyReLU())

model.add(tf.keras.layers.Dropout(0.3))

model.add(tf.keras.layers.Conv2D(filters=1, kernel_size=(7, 7), padding='valid'))

model.add(tf.keras.layers.Reshape((1, )))

return model

mnist data & build

GAN 모델에서 preprocess는 데이터를 펼친 벡터 형태로 반환했지만, DCGAN에서는 이미지 텐서를 출력

import tensorflow_datasets as tfds

mnist_bldr=tfds.builder('mnist')

mnist_bldr.download_and_prepare()

mnist=mnist_bldr.as_dataset(shuffle_files=False)

def preprocess(ex, mode='uniform'):

image=ex['image']

image=tf.image.convert_image_dtype(image, tf.float32)

image=image*2-1.0

if mode=='uniform':

input_z=tf.random.uniform(shape=(z_size, ), minval=-1.0, maxval=1.0)

elif mode=='normal':

input_z=tf.random.normal(shape=(z_size, ))

return input_z, image

gen_model=make_dcgan_generator()

disc_model=make_dcgan_discriminator()

gen_model.summary()

disc_model.summary()

Model: "sequential"

_________________________________________________________________

Layer (type) Output Shape Param #

=================================================================

dense (Dense) (None, 6272) 125440

_________________________________________________________________

batch_normalization (BatchNo (None, 6272) 25088

_________________________________________________________________

leaky_re_lu (LeakyReLU) (None, 6272) 0

_________________________________________________________________

reshape (Reshape) (None, 7, 7, 128) 0

_________________________________________________________________

conv2d_transpose (Conv2DTran (None, 7, 7, 128) 409600

_________________________________________________________________

batch_normalization_1 (Batch (None, 7, 7, 128) 512

_________________________________________________________________

leaky_re_lu_1 (LeakyReLU) (None, 7, 7, 128) 0

_________________________________________________________________

conv2d_transpose_1 (Conv2DTr (None, 14, 14, 64) 204800

_________________________________________________________________

batch_normalization_2 (Batch (None, 14, 14, 64) 256

_________________________________________________________________

leaky_re_lu_2 (LeakyReLU) (None, 14, 14, 64) 0

_________________________________________________________________

conv2d_transpose_2 (Conv2DTr (None, 28, 28, 32) 51200

_________________________________________________________________

batch_normalization_3 (Batch (None, 28, 28, 32) 128

_________________________________________________________________

leaky_re_lu_3 (LeakyReLU) (None, 28, 28, 32) 0

_________________________________________________________________

conv2d_transpose_3 (Conv2DTr (None, 28, 28, 1) 800

=================================================================

Total params: 817,824

Trainable params: 804,832

Non-trainable params: 12,992

_________________________________________________________________

Model: "sequential_1"

_________________________________________________________________

Layer (type) Output Shape Param #

=================================================================

conv2d (Conv2D) (None, 28, 28, 64) 1664

_________________________________________________________________

batch_normalization_4 (Batch (None, 28, 28, 64) 256

_________________________________________________________________

leaky_re_lu_4 (LeakyReLU) (None, 28, 28, 64) 0

_________________________________________________________________

conv2d_1 (Conv2D) (None, 14, 14, 128) 204928

_________________________________________________________________

batch_normalization_5 (Batch (None, 14, 14, 128) 512

_________________________________________________________________

leaky_re_lu_5 (LeakyReLU) (None, 14, 14, 128) 0

_________________________________________________________________

dropout (Dropout) (None, 14, 14, 128) 0

_________________________________________________________________

conv2d_2 (Conv2D) (None, 7, 7, 256) 819456

_________________________________________________________________

batch_normalization_6 (Batch (None, 7, 7, 256) 1024

_________________________________________________________________

leaky_re_lu_6 (LeakyReLU) (None, 7, 7, 256) 0

_________________________________________________________________

dropout_1 (Dropout) (None, 7, 7, 256) 0

_________________________________________________________________

conv2d_3 (Conv2D) (None, 1, 1, 1) 12545

_________________________________________________________________

reshape_1 (Reshape) (None, 1) 0

=================================================================

Total params: 1,040,385

Trainable params: 1,039,489

Non-trainable params: 896

_________________________________________________________________

DCGAN Generator & Discriminator