Autoencoder
Autoencoder
차원축소 목적, feature extractor, generative model. 데이터를 효율적으로 representation하는 방법을 학습
- Stacked Autoencoder
- 여러개의 hidden layer를 가진경우. 더 복잡한 표현을 학습. Hidden layer를 기준으로 Encoder와 Decoder가 대칭
- 완벽하게 대칭을 이룰경우, Encoder와 Decoder의 가중치를 묶어 훈련속도를 높이고 Overfitting 위험을 줄일 수 있다.
- Denoising Autoencoder
- Stacked Autoencoder의 경우 hidden layer가 많으면 overfitting 가능성.
- Input에 noise를 주어 모델이 데이터 표현을 학습하기 힘들게 함.
-
LSTM Encoder
- Basic
- Encoder
- 입력 x를 저차원 잠재 벡터 $z=f(x)$로 변환. 저차원이기 때문에 Encoder가 Input data에서 가장 중요한 특징만 학습함 * Decoder
- 잠재 벡터 $g(z) = \tilde x$로 부터 입력 데이터를 복원.
- Decoder는 $\tilde x$를 가능한 $x$에 가깝게 만드는 것을 목표로 함.
Imports
import tensorflow as tf
from tensorflow.keras.layers import Dense, Input, Conv2D, Flatten, Reshape, Conv2DTranspose
from tensorflow.keras.models import Model
from tensorflow.keras.datasets import mnist
from tensorflow.keras.utils import plot_model
from tensorflow.keras import backend as K
import tensorboard
import numpy as np
import matplotlib.pyplot as plt
from datetime import datetime
import os
Dataset, Preprocessing
(x_train, _), (x_test, _) = mnist.load_data()
image_size = x_train.shape[1]
print(x_train.shape)
x_train = np.reshape(x_train, [-1, image_size, image_size, 1])
x_test = np.reshape(x_test, [-1, image_size, image_size, 1])
x_train = x_train.astype('float32') / 255
x_test = x_test.astype('float32') / 255
print(x_train.shape)
(60000, 28, 28)
(60000, 28, 28, 1)
Hyperparameters
input_shape = (image_size, image_size, 1)
batch_size = 32
kernel_size = 3
latent_dim = 16
Model
Encoder
inputs = Input(shape = input_shape, name='encoder_input')
x = inputs
x = Conv2D(32, kernel_size = kernel_size, activation='relu', strides = 2, padding = 'same')(x)
x = Conv2D(64, kernel_size= kernel_size, activation='relu', strides=2, padding = 'same')(x)
shape = K.int_shape(x)
x = Flatten()(x)
latent = Dense(latent_dim, name = 'latent_vector')(x)
shape
(None, 7, 7, 64)
encoder = Model(inputs, latent, name='encoder')
encoder.summary()
Model: "encoder"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
encoder_input (InputLayer) [(None, 28, 28, 1)] 0
_________________________________________________________________
conv2d_2 (Conv2D) (None, 14, 14, 32) 320
_________________________________________________________________
conv2d_3 (Conv2D) (None, 7, 7, 64) 18496
_________________________________________________________________
flatten (Flatten) (None, 3136) 0
_________________________________________________________________
latent_vector (Dense) (None, 16) 50192
=================================================================
Total params: 69,008
Trainable params: 69,008
Non-trainable params: 0
_________________________________________________________________
plot_model(encoder, show_shapes=True)
Decoder
latent_inputs = Input(shape=(latent_dim, ), name='decoder_input')
x = Dense(shape[1] * shape[2] * shape[3])(latent_inputs)
x = Reshape((shape[1], shape[2], shape[3]))(x)
x = Conv2DTranspose(filters=64, kernel_size=kernel_size, activation='relu', strides=2, padding='same')(x)
x = Conv2DTranspose(filters=32, kernel_size=kernel_size, activation='relu', strides=2, padding='same')(x)
outputs = Conv2DTranspose(filters=1, kernel_size=kernel_size, activation='sigmoid', padding='same', name='decoder_output')(x)
decoder = Model(latent_inputs, outputs, name='decoder')
decoder.summary()
Model: "decoder"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
decoder_input (InputLayer) [(None, 16)] 0
_________________________________________________________________
dense_1 (Dense) (None, 3136) 53312
_________________________________________________________________
reshape (Reshape) (None, 7, 7, 64) 0
_________________________________________________________________
conv2d_transpose (Conv2DTran (None, 14, 14, 64) 36928
_________________________________________________________________
conv2d_transpose_1 (Conv2DTr (None, 28, 28, 32) 18464
_________________________________________________________________
decoder_output (Conv2DTransp (None, 28, 28, 1) 289
=================================================================
Total params: 108,993
Trainable params: 108,993
Non-trainable params: 0
_________________________________________________________________
plot_model(decoder, show_shapes=True)
Autoencoder
autoencoder = Model(inputs, decoder(encoder(inputs)), name='autoencoder')
autoencoder.summary()
Model: "autoencoder"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
encoder_input (InputLayer) [(None, 28, 28, 1)] 0
_________________________________________________________________
encoder (Functional) (None, 16) 69008
_________________________________________________________________
decoder (Functional) (None, 28, 28, 1) 108993
=================================================================
Total params: 178,001
Trainable params: 178,001
Non-trainable params: 0
_________________________________________________________________
plot_model(autoencoder, show_shapes=True)
autoencoder.compile(loss='mse', optimizer='adam')
logdir = 'log/'+datetime.now().strftime("%Y%m%d-%H%M%S")
tensorboard_callback = tf.keras.callbacks.TensorBoard(log_dir = logdir)
autoencoder.fit(x_train, x_train, validation_data = (x_test, x_test), epochs=20, batch_size = batch_size, callbacks=[tensorboard_callback])
Epoch 1/20
2/1875 [..............................] - ETA: 1:10 - loss: 0.0123WARNING:tensorflow:Callbacks method `on_train_batch_end` is slow compared to the batch time (batch time: 0.0020s vs `on_train_batch_end` time: 0.0710s). Check your callbacks.
1875/1875 [==============================] - 5s 3ms/step - loss: 0.0098 - val_loss: 0.0091
Epoch 2/20
1875/1875 [==============================] - 5s 3ms/step - loss: 0.0087 - val_loss: 0.0083
Epoch 3/20
1875/1875 [==============================] - 5s 3ms/step - loss: 0.0081 - val_loss: 0.0079
Epoch 4/20
1875/1875 [==============================] - 5s 3ms/step - loss: 0.0077 - val_loss: 0.0076
Epoch 5/20
1875/1875 [==============================] - 5s 3ms/step - loss: 0.0075 - val_loss: 0.0074
Epoch 6/20
1875/1875 [==============================] - 5s 3ms/step - loss: 0.0073 - val_loss: 0.0073
Epoch 7/20
1875/1875 [==============================] - 5s 3ms/step - loss: 0.0071 - val_loss: 0.0072
Epoch 8/20
1875/1875 [==============================] - 5s 3ms/step - loss: 0.0070 - val_loss: 0.0071
Epoch 9/20
1875/1875 [==============================] - 5s 3ms/step - loss: 0.0069 - val_loss: 0.0070
Epoch 10/20
1875/1875 [==============================] - 5s 3ms/step - loss: 0.0068 - val_loss: 0.0070
Epoch 11/20
1875/1875 [==============================] - 5s 3ms/step - loss: 0.0067 - val_loss: 0.0069
Epoch 12/20
1875/1875 [==============================] - 5s 3ms/step - loss: 0.0066 - val_loss: 0.0068
Epoch 13/20
1875/1875 [==============================] - 5s 3ms/step - loss: 0.0065 - val_loss: 0.0068
Epoch 14/20
1875/1875 [==============================] - 5s 3ms/step - loss: 0.0065 - val_loss: 0.0067
Epoch 15/20
1875/1875 [==============================] - 5s 3ms/step - loss: 0.0064 - val_loss: 0.0068
Epoch 16/20
1875/1875 [==============================] - 5s 3ms/step - loss: 0.0064 - val_loss: 0.0067
Epoch 17/20
1875/1875 [==============================] - 5s 3ms/step - loss: 0.0063 - val_loss: 0.0067
Epoch 18/20
1875/1875 [==============================] - 5s 3ms/step - loss: 0.0063 - val_loss: 0.0068
Epoch 19/20
1875/1875 [==============================] - 5s 3ms/step - loss: 0.0062 - val_loss: 0.0066
Epoch 20/20
1875/1875 [==============================] - 5s 3ms/step - loss: 0.0062 - val_loss: 0.0066
<tensorflow.python.keras.callbacks.History at 0x248003e7548>
x_decoded = autoencoder.predict(x_test)
imgs = np.concatenate([x_test[:8], x_decoded[:8]])
imgs = imgs.reshape((4, 4, image_size, image_size))
imgs = np.vstack([np.hstack(i) for i in imgs])
plt.figure()
plt.axis('off')
plt.imshow(imgs, interpolation='none', cmap='gray')
plt.show()
