728x90
반응형
SMALL
데이터 로드
import numpy as np
import matplotlib.pyplot as plt
from keras.datasets import mnist
from keras.layers import Input, Dense, Lambda
from keras.models import Model
from keras import backend as K
from keras import losses
from scipy.stats import norm
(x_train, _), (x_test, _) = mnist.load_data()
x_train, x_test = x_train.astype('float32')/255., x_test.astype('float32')/255.
x_train, x_test = x_train.reshape(x_train.shape[0], -1), x_test.reshape(x_test.shape[0], -1)
print(x_train.shape, x_test.shape)
모델 생성
# 네트워크 파라미터
batch_size, n_epoch = 100, 100
n_hidden, z_dim = 256, 2
# 인코더
x = Input(shape=(x_train.shape[1:]))
x_encoded = Dense(n_hidden, activation='relu')(x)
x_encoded = Dense(n_hidden//2, activation='relu')(x_encoded)
mu = Dense(z_dim)(x_encoded)
log_var = Dense(z_dim)(x_encoded)
# 샘플링 함수
def sampling(args):
mu, log_var = args
eps = K.random_normal(shape=(batch_size, z_dim), mean=0., stddev=1.0)
return mu + K.exp(log_var) * eps
z = Lambda(sampling, output_shape=(z_dim,))([mu, log_var])
인코더 모델에서는 인코딩 층을 𝑚𝑢층과 log _𝑣𝑎𝑟 층에 연결한다. 영상이 입력되면 mu와 log _var의 값을 출력한다. Lambda 층이 잠재 공산에서 𝑚𝑢와 log _𝑣𝑎𝑟로 정의되는 정규분포로부터 포인트 𝑧를 샘플링한다.
# 디코더
z_decoder1 = Dense(n_hidden//2, activation='relu')
z_decoder2 = Dense(n_hidden, activation='relu')
y_decoder = Dense(x_train.shape[1], activation='sigmoid')
z_decoded = z_decoder1(z)
z_decoded = z_decoder2(z_decoded)
y = y_decoder(z_decoded)
모델 학습
# 손실 함수 설정
reconstruction_loss = losses.binary_crossentropy(x, y) * x_train.shape[1]
kl_loss = 0.5 * K.sum(K.square(mu) + K.exp(log_var) - log_var - 1, axis = -1)
vae_loss = reconstruction_loss + kl_loss
vae = Model(x, y)
vae.add_loss(vae_loss)
vae.compile(optimizer='rmsprop')
vae.summary()
# 학습
vae.fit(x_train,
shuffle=True,
epochs=n_epoch,
batch_size=batch_size,
validation_data=(x_test, None),
verbose=1)Model: "model_1"
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_1 (InputLayer) [(None, 784)] 0 []
dense (Dense) (None, 256) 200960 ['input_2[0][0]']
dense_1 (Dense) (None, 128) 32896 ['dense[0][0]']
dense_2 (Dense) (None, 2) 258 ['dense_1[0][0]']
dense_3 (Dense) (None, 2) 258 ['dense_1[0][0]']
lambda (Lambda) (100, 2) 0 ['dense_2[0][0]',
'dense_3[0][0]']
dense_6 (Dense) (100, 128) 384 ['lambda[0][0]']
dense_7 (Dense) (100, 256) 33024 ['dense_6[0][0]']
dense_8 (Dense) (100, 784) 201488 ['dense_7[0][0]']
tf.math.square_1 (TFOpLambda) (None, 2) 0 ['dense_2[0][0]']
tf.math.exp_1 (TFOpLambda) (None, 2) 0 ['dense_3[0][0]']
tf.__operators__.add_2 (TFOpLa (None, 2) 0 ['tf.math.square_1[0][0]',
mbda) 'tf.math.exp_1[0][0]']
tf.cast_1 (TFOpLambda) (None, 784) 0 ['input_2[0][0]']
tf.convert_to_tensor_3 (TFOpLa (100, 784) 0 ['dense_8[0][0]']
mbda)
tf.math.subtract_2 (TFOpLambda (None, 2) 0 ['tf.__operators__.add_2[0][0]',
) 'dense_3[0][0]']
tf.keras.backend.binary_crosse (100, 784) 0 ['tf.cast_1[0][0]',
ntropy_1 (TFOpLambda) 'tf.convert_to_tensor_3[0][0]']
tf.math.subtract_3 (TFOpLambda (None, 2) 0 ['tf.math.subtract_2[0][0]']
)
tf.math.reduce_mean_1 (TFOpLam (100,) 0 ['tf.keras.backend.binary_crossen
bda) tropy_1[0][0]']
tf.math.reduce_sum_1 (TFOpLamb (None,) 0 ['tf.math.subtract_3[0][0]']
da)
tf.math.multiply_2 (TFOpLambda (100,) 0 ['tf.math.reduce_mean_1[0][0]']
)
tf.math.multiply_3 (TFOpLambda (None,) 0 ['tf.math.reduce_sum_1[0][0]']
)
tf.__operators__.add_3 (TFOpLa (100,) 0 ['tf.math.multiply_2[0][0]',
mbda) 'tf.math.multiply_3[0][0]']
add_loss_1 (AddLoss) (100,) 0 ['tf.__operators__.add_3[0][0]']
==================================================================================================
Total params: 469,268
Trainable params: 469,268
Non-trainable params: 0encoder = Model(x, mu)
encoder.summary()
def display_embeddings(
encoder,
test_x,
test_y,
batch_size
):
# display a 2D plot of the digit classes in the encoding space
x_test_encoded = encoder.predict(test_x, batch_size=batch_size)
plt.figure(figsize=(6, 6))
plt.scatter(x_test_encoded[:, 0], x_test_encoded[:, 1], c=test_y)
plt.colorbar()
plt.show()(X_train, y_train), (X_test, y_test) = mnist.load_data()
display_embeddings(encoder, x_test, y_test, 100)
decoder_input = Input(shape=(z_dim,))
_z_decoded = z_decoder1(decoder_input)
_z_decoded = z_decoder2(_z_decoded)
_y = y_decoder(_z_decoded)
generator = Model(decoder_input, _y)
n = 15
digit_size = 28
figure = np.zeros((digit_size * n, digit_size * n))
grid_x = norm.ppf(np.linspace(0.05, 0.95, n))
grid_y = norm.ppf(np.linspace(0.05, 0.95, n))
for i, yi in enumerate(grid_x):
for j, xi in enumerate(grid_y):
z_sample = np.array([[xi, yi]])
x_decoded = generator.predict(z_sample)
digit = x_decoded[0].reshape(digit_size, digit_size)
figure[i * digit_size: (i + 1) * digit_size, j * digit_size: (j + 1) * digit_size] = digit
plt.figure(figsize=(10, 10))
plt.imshow(figure, cmap='Greys_r')
plt.show
728x90
반응형
LIST
'Visual Intelligence > Generative Model' 카테고리의 다른 글
| [Generative Model] Neural style transfer (0) | 2022.12.09 |
|---|---|
| [Generative Model] 스타일 전송 (Style Transfer) (0) | 2022.12.09 |
| [Generative Model] VAE (Variational AutoEncoder) (0) | 2022.12.08 |
| [Generative Model] 오토인코더 (MNIST) (0) | 2022.12.08 |
| [Generative Model] 오토인코더 (AutoEncoder) (0) | 2022.12.08 |
