AE3. Denoising AE - 노이즈제거 오토인코더란?

학습목적

이 포스팅에선 오터인코더의 한 사례인 노이즈가 추가된 데이터에서 깨끗한 이미지를 뽑아낼 수 있는 Denoising AutoEncoder에 대해서 알아보겠습니다.

Denosing AutoEncoder

Denosing AutoEncoder는 이미지 데이터에 노이즈가 있을 때 제거할 수 있도록 학습되는 AutoEncoder 중에 하나입니다. 크게 네트워크 모양이 바뀌지는 않지만 기존 AE와 다르게 X, Y값이 다르다는 점이 있습니다.

왜 노이즈 제거에 AE를 사용할까?

AutoEncoder는 차원을 축소시키면서 데이터를 함축적으로 표현하게 되어있습니다. 즉, 핵심적인 데이터만 저장하도록 되어있습니다. 이러한 원리를 이용하여 노이즈가 추가된 데이터에서 핵심적인 데이터를 추출할 수 있는 AutoEncoder를 활용하게 됬습니다.
- 자세한 내용은 Manifold Learning에 대해서 한번 포스팅하겠습니다.

오토인코더 실습

import os
import sys
import warnings

import random

import numpy as np
import pandas as pd

import scipy
import scipy.stats
from scipy.spatial import ConvexHull, convex_hull_plot_2d
from sklearn.datasets import make_regression
from sklearn.ensemble import GradientBoostingRegressor
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score, precision_score, recall_score

import matplotlib as mpl
import matplotlib.pyplot as plt
from plotnine import * 

import tensorflow as tf
from tensorflow.keras.layers import Input, Conv2D, Flatten, Dense, Conv2DTranspose, Reshape, Lambda, Activation, BatchNormalization, LeakyReLU, Dropout
from tensorflow.keras.models import Model, Sequential
from tensorflow.keras import backend as K
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.callbacks import ModelCheckpoint 
from tensorflow.keras import layers, losses
from tensorflow.keras.datasets import mnist, fashion_mnist

%matplotlib inline
warnings.filterwarnings(action='ignore')

(x_train, y_train), (x_test, y_test) = mnist.load_data()

데이터에 임의로 노이즈를 추가해보겠습니다.

x_train = x_train.astype('float32') / 255.
x_test = x_test.astype('float32') / 255.
x_train = np.reshape(x_train, (len(x_train), 28, 28, 1))  # adapt this if using `channels_first` image data format
x_test = np.reshape(x_test, (len(x_test), 28, 28, 1))  # adapt this if using `channels_first` image data format

noise_factor = 0.5
x_train_noisy = x_train + noise_factor * np.random.normal(loc=0.0, scale=1.0, size=x_train.shape) 
x_test_noisy = x_test + noise_factor * np.random.normal(loc=0.0, scale=1.0, size=x_test.shape) 

x_train_noisy = np.clip(x_train_noisy, 0., 1.)
x_test_noisy = np.clip(x_test_noisy, 0., 1.)

n = 10
plt.figure(figsize=(20, 4))
for i in range(n):
    # display original
    ax = plt.subplot(2, n, i + 1)
    plt.imshow(np.reshape(x_train[i], (28, 28)))
    plt.title(y_train[i])
    plt.gray()
    ax.get_xaxis().set_visible(False)
    ax.get_yaxis().set_visible(False)
plt.show()

png

Oops

n = 10
plt.figure(figsize=(20, 4))
for i in range(n):
    # display original
    ax = plt.subplot(2, n, i + 1)
    plt.imshow(np.reshape(x_train_noisy[i], (28, 28)))
    plt.title(y_train[i])
    plt.gray()
    ax.get_xaxis().set_visible(False)
    ax.get_yaxis().set_visible(False)
plt.show()

png

Oops

Denoising AutoEncoder 모델 구성

기본적인 AE와 비슷하게 Encoder로 차원을 축소하고 Decoder로 다시 원본으로 맞추는 작업을 하되, X값에는 noise가 추가된 이미지를 Y값에는 정상 이미지를 넣어 학습을 시키게 됩니다.

class Denoising_Autoencoder_V1():
    def __init__(self
        , input_dim
        , encoder_conv_filters
        , encoder_conv_kernel_size
        , encoder_conv_strides
        , decoder_conv_t_filters
        , decoder_conv_t_kernel_size
        , decoder_conv_t_strides
        , z_dim
        , use_batch_norm = False
        , use_dropout = False
        ):

        self.name = 'autoencoder'

        self.input_dim = input_dim
        self.encoder_conv_filters = encoder_conv_filters
        self.encoder_conv_kernel_size = encoder_conv_kernel_size
        self.encoder_conv_strides = encoder_conv_strides
        self.decoder_conv_t_filters = decoder_conv_t_filters
        self.decoder_conv_t_kernel_size = decoder_conv_t_kernel_size
        self.decoder_conv_t_strides = decoder_conv_t_strides
        self.z_dim = z_dim

        self.use_batch_norm = use_batch_norm
        self.use_dropout = use_dropout

        self.n_layers_encoder = len(encoder_conv_filters)
        self.n_layers_decoder = len(decoder_conv_t_filters)

        self._build()

    def _build(self):

        ### THE ENCODER
        encoder_input = Input(shape=self.input_dim, name='encoder_input')

        x = encoder_input

        for i in range(self.n_layers_encoder):
            conv_layer = Conv2D(
                filters = self.encoder_conv_filters[i]
                , kernel_size = self.encoder_conv_kernel_size[i]
                , strides = self.encoder_conv_strides[i]
                , padding = 'same'
                , name = 'encoder_conv_' + str(i)
                )

            x = conv_layer(x)

            x = LeakyReLU()(x)

            if self.use_batch_norm:
                x = BatchNormalization()(x)

            if self.use_dropout:
                x = Dropout(rate = 0.25)(x)

        shape_before_flattening = K.int_shape(x)[1:]

        x = Flatten()(x)
        encoder_output= Dense(self.z_dim, name='encoder_output')(x)

        self.encoder = Model(encoder_input, encoder_output)


        ### THE DECODER
        decoder_input = Input(shape=(self.z_dim,), name='decoder_input')

        x = Dense(np.prod(shape_before_flattening))(decoder_input)
        x = Reshape(shape_before_flattening)(x)

        for i in range(self.n_layers_decoder):
            conv_t_layer = Conv2DTranspose(
                filters = self.decoder_conv_t_filters[i]
                , kernel_size = self.decoder_conv_t_kernel_size[i]
                , strides = self.decoder_conv_t_strides[i]
                , padding = 'same'
                , name = 'decoder_conv_t_' + str(i)
                )

            x = conv_t_layer(x)

            if i < self.n_layers_decoder - 1:
                x = LeakyReLU()(x)
                
                if self.use_batch_norm:
                    x = BatchNormalization()(x)
                
                if self.use_dropout:
                    x = Dropout(rate = 0.25)(x)
            else:
                x = Activation('sigmoid')(x)

        decoder_output = x

        self.decoder = Model(decoder_input, decoder_output)

        ### THE FULL AUTOENCODER
        model_input = encoder_input
        model_output = self.decoder(encoder_output)

        self.model = Model(model_input, model_output)
        
    def compile(self, learning_rate):
        self.learning_rate = learning_rate
        optimizer = Adam(lr=learning_rate)

        self.model.compile(optimizer=optimizer, loss = tf.keras.losses.MeanSquaredError(reduction="auto", name="mean_squared_error")) 

    def train(self, x_train, y_train, batch_size, epochs, verbose = 1):
        self.model.fit(     
        x_train
        , y_train
        , batch_size = batch_size
        , shuffle = True
        , epochs = epochs
        , verbose = verbose
        )

DAE_V1 = Denoising_Autoencoder_V1(
    input_dim = (28,28,1)
    , encoder_conv_filters = [32,64,64,64]
    , encoder_conv_kernel_size = [3,3,3,3]
    , encoder_conv_strides = [1,2,2,1]
    , decoder_conv_t_filters = [64,64,32,1]
    , decoder_conv_t_kernel_size = [3,3,3,3]
    , decoder_conv_t_strides = [1,2,2,1]
    , z_dim = 2
)

DAE_V1.compile(0.00005)

DAE_V1.encoder.summary()

Model: "model_4"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
encoder_input (InputLayer)   [(None, 28, 28, 1)]       0         
_________________________________________________________________
encoder_conv_0 (Conv2D)      (None, 28, 28, 32)        320       
_________________________________________________________________
leaky_re_lu_10 (LeakyReLU)   (None, 28, 28, 32)        0         
_________________________________________________________________
encoder_conv_1 (Conv2D)      (None, 14, 14, 64)        18496     
_________________________________________________________________
leaky_re_lu_11 (LeakyReLU)   (None, 14, 14, 64)        0         
_________________________________________________________________
encoder_conv_2 (Conv2D)      (None, 7, 7, 64)          36928     
_________________________________________________________________
leaky_re_lu_12 (LeakyReLU)   (None, 7, 7, 64)          0         
_________________________________________________________________
encoder_conv_3 (Conv2D)      (None, 7, 7, 64)          36928     
_________________________________________________________________
leaky_re_lu_13 (LeakyReLU)   (None, 7, 7, 64)          0         
_________________________________________________________________
flatten (Flatten)            (None, 3136)              0         
_________________________________________________________________
encoder_output (Dense)       (None, 2)                 6274      
=================================================================
Total params: 98,946
Trainable params: 98,946
Non-trainable params: 0
_________________________________________________________________

DAE_V1.decoder.summary()

Model: "model_5"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
decoder_input (InputLayer)   [(None, 2)]               0         
_________________________________________________________________
dense (Dense)                (None, 3136)              9408      
_________________________________________________________________
reshape (Reshape)            (None, 7, 7, 64)          0         
_________________________________________________________________
decoder_conv_t_0 (Conv2DTran (None, 7, 7, 64)          36928     
_________________________________________________________________
leaky_re_lu_14 (LeakyReLU)   (None, 7, 7, 64)          0         
_________________________________________________________________
decoder_conv_t_1 (Conv2DTran (None, 14, 14, 64)        36928     
_________________________________________________________________
leaky_re_lu_15 (LeakyReLU)   (None, 14, 14, 64)        0         
_________________________________________________________________
decoder_conv_t_2 (Conv2DTran (None, 28, 28, 32)        18464     
_________________________________________________________________
leaky_re_lu_16 (LeakyReLU)   (None, 28, 28, 32)        0         
_________________________________________________________________
decoder_conv_t_3 (Conv2DTran (None, 28, 28, 1)         289       
_________________________________________________________________
activation_2 (Activation)    (None, 28, 28, 1)         0         
=================================================================
Total params: 102,017
Trainable params: 102,017
Non-trainable params: 0
_________________________________________________________________

DAE_V1.train(x_train_noisy, x_train, batch_size = 32, epochs = 200)

Epoch 1/200
1875/1875 [==============================] - 9s 4ms/step - loss: 0.0757
Epoch 2/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0592
Epoch 3/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0567
Epoch 4/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0558
Epoch 5/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0552
Epoch 6/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0546
Epoch 7/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0540
Epoch 8/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0533
Epoch 9/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0528
Epoch 10/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0523
Epoch 11/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0518
Epoch 12/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0514
Epoch 13/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0510
Epoch 14/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0506
Epoch 15/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0502
Epoch 16/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0499
Epoch 17/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0495
Epoch 18/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0492
Epoch 19/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0489
Epoch 20/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0487
Epoch 21/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0484
Epoch 22/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0482
Epoch 23/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0480
Epoch 24/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0478
Epoch 25/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0476
Epoch 26/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0474
Epoch 27/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0472
Epoch 28/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0471
Epoch 29/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0469
Epoch 30/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0468
Epoch 31/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0467
Epoch 32/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0465
Epoch 33/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0464
Epoch 34/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0463
Epoch 35/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0462
Epoch 36/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0461
Epoch 37/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0460
Epoch 38/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0458
Epoch 39/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0457
Epoch 40/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0457
Epoch 41/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0456
Epoch 42/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0455
Epoch 43/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0454
Epoch 44/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0453
Epoch 45/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0452
Epoch 46/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0452
Epoch 47/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0451
Epoch 48/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0450
Epoch 49/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0450
Epoch 50/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0449
Epoch 51/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0448
Epoch 52/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0447
Epoch 53/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0447
Epoch 54/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0446
Epoch 55/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0446
Epoch 56/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0445
Epoch 57/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0445
Epoch 58/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0444
Epoch 59/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0444
Epoch 60/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0443
Epoch 61/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0443
Epoch 62/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0442
Epoch 63/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0442
Epoch 64/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0442
Epoch 65/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0441
Epoch 66/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0441
Epoch 67/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0440
Epoch 68/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0440
Epoch 69/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0439
Epoch 70/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0439
Epoch 71/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0438
Epoch 72/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0438
Epoch 73/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0438
Epoch 74/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0437
Epoch 75/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0437
Epoch 76/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0436
Epoch 77/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0436
Epoch 78/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0436
Epoch 79/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0435
Epoch 80/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0435
Epoch 81/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0435
Epoch 82/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0434
Epoch 83/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0434
Epoch 84/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0434
Epoch 85/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0433
Epoch 86/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0433
Epoch 87/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0433
Epoch 88/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0432
Epoch 89/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0432
Epoch 90/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0432
Epoch 91/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0431
Epoch 92/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0431
Epoch 93/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0431
Epoch 94/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0430
Epoch 95/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0430
Epoch 96/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0430
Epoch 97/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0430
Epoch 98/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0429
Epoch 99/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0429
Epoch 100/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0429
Epoch 101/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0428
Epoch 102/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0428
Epoch 103/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0428
Epoch 104/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0428
Epoch 105/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0427
Epoch 106/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0427
Epoch 107/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0427
Epoch 108/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0427
Epoch 109/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0426
Epoch 110/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0426
Epoch 111/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0426
Epoch 112/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0425
Epoch 113/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0425
Epoch 114/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0425
Epoch 115/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0425
Epoch 116/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0425
Epoch 117/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0424
Epoch 118/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0424
Epoch 119/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0424
Epoch 120/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0424
Epoch 121/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0423
Epoch 122/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0423
Epoch 123/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0423
Epoch 124/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0422
Epoch 125/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0422
Epoch 126/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0422
Epoch 127/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0422
Epoch 128/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0421
Epoch 129/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0421
Epoch 130/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0421
Epoch 131/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0421
Epoch 132/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0421
Epoch 133/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0420
Epoch 134/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0420
Epoch 135/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0420
Epoch 136/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0420
Epoch 137/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0419
Epoch 138/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0419
Epoch 139/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0419
Epoch 140/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0419
Epoch 141/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0419
Epoch 142/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0418
Epoch 143/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0418
Epoch 144/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0418
Epoch 145/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0418
Epoch 146/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0418
Epoch 147/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0417
Epoch 148/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0417
Epoch 149/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0417
Epoch 150/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0417
Epoch 151/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0417
Epoch 152/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0416
Epoch 153/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0416
Epoch 154/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0416
Epoch 155/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0416
Epoch 156/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0416
Epoch 157/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0416
Epoch 158/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0415
Epoch 159/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0415
Epoch 160/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0415
Epoch 161/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0415
Epoch 162/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0414
Epoch 163/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0414
Epoch 164/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0414
Epoch 165/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0414
Epoch 166/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0414
Epoch 167/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0414
Epoch 168/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0413
Epoch 169/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0413
Epoch 170/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0413
Epoch 171/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0413
Epoch 172/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0413
Epoch 173/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0413
Epoch 174/200
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0412
Epoch 175/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0412
Epoch 176/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0412
Epoch 177/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0412
Epoch 178/200
1875/1875 [==============================] - 8s 5ms/step - loss: 0.0412
Epoch 179/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0412
Epoch 180/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0411
Epoch 181/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0411
Epoch 182/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0411
Epoch 183/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0411
Epoch 184/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0411
Epoch 185/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0411
Epoch 186/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0410
Epoch 187/200
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0410
Epoch 188/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0410
Epoch 189/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0410
Epoch 190/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0410
Epoch 191/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0410
Epoch 192/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0409
Epoch 193/200
1875/1875 [==============================] - 8s 5ms/step - loss: 0.0409
Epoch 194/200
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0409
Epoch 195/200
1875/1875 [==============================] - 8s 5ms/step - loss: 0.0409
Epoch 196/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0409
Epoch 197/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0409
Epoch 198/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0408
Epoch 199/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0408
Epoch 200/200
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0408

n = 10
plt.figure(figsize=(20, 4))
for i in range(n):
    # display original
    ax = plt.subplot(1, n, i + 1)
    plt.imshow(np.reshape(x_test_noisy[i], (28, 28)))
    plt.title(y_test[i])
    plt.gray()
    ax.get_xaxis().set_visible(False)
    ax.get_yaxis().set_visible(False)
plt.show()

png

Oops

n = 10
plt.figure(figsize=(20, 4))
for i in range(n):
    # display original
    ax = plt.subplot(1, n, i + 1)
    plt.imshow(np.reshape(x_test[i], (28, 28)))
    plt.title(y_test[i])
    plt.gray()
    ax.get_xaxis().set_visible(False)
    ax.get_yaxis().set_visible(False)
plt.show()

png

Oops

결과를 보게되면 AutoEncoder를 돌렸을 때 Convexhull이 잘 생성된 이미지는 잘 구분하지만, 결국 잘 구분되지 않은 애들은 노이즈 제거도 잘 하지 못한다는 것을 알 수 있습니다.
왜냐하면 Latent Space에 각자의 영역을 충분히 만들어내지 못했기 때문에 다시 Decoder를 통해서도 생성이 잘 되지 않습니다.

dae_predict = DAE_V1.model.predict(x_test_noisy)

n = 10
plt.figure(figsize=(20, 4))
for i in range(n):
    # display original
    ax = plt.subplot(2, n, i + 1)
    plt.imshow(np.reshape(dae_predict[i], (28, 28)))
    plt.title(y_test[i])
    plt.gray()
    ax.get_xaxis().set_visible(False)
    ax.get_yaxis().set_visible(False)
plt.show()

png

Oops

Denoising AutoEncoder 모델 구성 Version2

위의 AutoEncoder가 잠재공간 내에 각각의 이미지를 잘 표현(representation)을 하지 못했기 때문에, 잠재공간을 늘려서 학습을 하면 좀 더 잘되지 않을까 라는 시도를 해볼 수 있습니다. 차원을 flatten이나 2차원으로 만들지 않고 잠재공간(latent space)가 충분한(7, 7, 128) 상태로 학습을 시켜보겠습니다.

class Denoising_Autoencoder_V2():
    def __init__(self
        , input_dim
        , encoder_conv_filters
        , encoder_conv_kernel_size
        , encoder_conv_strides
        , decoder_conv_t_filters
        , decoder_conv_t_kernel_size
        , decoder_conv_t_strides
        , z_dim
        , use_batch_norm = False
        , use_dropout = False
        ):

        self.name = 'autoencoder'

        self.input_dim = input_dim
        self.encoder_conv_filters = encoder_conv_filters
        self.encoder_conv_kernel_size = encoder_conv_kernel_size
        self.encoder_conv_strides = encoder_conv_strides
        self.decoder_conv_t_filters = decoder_conv_t_filters
        self.decoder_conv_t_kernel_size = decoder_conv_t_kernel_size
        self.decoder_conv_t_strides = decoder_conv_t_strides
        self.z_dim = z_dim

        self.use_batch_norm = use_batch_norm
        self.use_dropout = use_dropout

        self.n_layers_encoder = len(encoder_conv_filters)
        self.n_layers_decoder = len(decoder_conv_t_filters)

        self._build()

    def _build(self):

        ### THE ENCODER
        encoder_input = Input(shape=self.input_dim, name='encoder_input')

        x = encoder_input

        for i in range(self.n_layers_encoder):
            conv_layer = Conv2D(
                filters = self.encoder_conv_filters[i]
                , kernel_size = self.encoder_conv_kernel_size[i]
                , strides = self.encoder_conv_strides[i]
                , padding = 'same'
                , name = 'encoder_conv_' + str(i)
                )

            x = conv_layer(x)

            x = LeakyReLU()(x)

            if self.use_batch_norm:
                x = BatchNormalization()(x)

            if self.use_dropout:
                x = Dropout(rate = 0.25)(x)

        # encoder_output= Conv2D(
        #         filters = self.z_dim
        #         , kernel_size = 1
        #         , strides = 1
        #         , padding = 'same'
        #         , name = 'encoder_conv_' + str(i + 1)
        #         , activation = "sigmoid"
        #         )
        
        # x = encoder_output(x)
        
        self.encoder = Model(encoder_input, x)

        ### THE DECODER

        for i in range(self.n_layers_decoder):
            conv_t_layer = Conv2DTranspose(
                filters = self.decoder_conv_t_filters[i]
                , kernel_size = self.decoder_conv_t_kernel_size[i]
                , strides = self.decoder_conv_t_strides[i]
                , padding = 'same'
                , name = 'decoder_conv_t_' + str(i)
                )

            x = conv_t_layer(x)

            if i < self.n_layers_decoder - 1:
                x = LeakyReLU()(x)
                
                if self.use_batch_norm:
                    x = BatchNormalization()(x)
                
                if self.use_dropout:
                    x = Dropout(rate = 0.25)(x)
            else:
                x = Activation('sigmoid')(x)

        decoder_output = x

        ### THE FULL AUTOENCODER
        model_input = encoder_input

        self.model = Model(model_input, decoder_output)
        
    def compile(self, learning_rate, loss_func = "BCE"):
        self.learning_rate = learning_rate
        optimizer = Adam(lr=learning_rate)

        if loss_func == "BCE" :
            self.model.compile(optimizer=optimizer, loss = tf.keras.losses.BinaryCrossentropy(reduction="auto", name="BCE")) 

        elif loss_func == "mse" :
            self.model.compile(optimizer=optimizer, loss = tf.keras.losses.MeanSquaredError(reduction="auto", name="mean_squared_error")) 
        
        else :
            self.model.compile(optimizer=optimizer, loss = tf.keras.losses.BinaryCrossentropy(reduction="auto", name="BCE")) 
          

    def train(self, x_train, y_train, batch_size, epochs, verbose = 1):
        self.model.fit(     
        x_train
        , y_train
        , batch_size = batch_size
        , shuffle = True
        , epochs = epochs
        , verbose = verbose
        )

DAE_V2 = Denoising_Autoencoder_V2(
    input_dim = (28,28,1)
    , encoder_conv_filters = [32,64,128]
    , encoder_conv_kernel_size = [3,3,3,3]
    , encoder_conv_strides = [1,2,2]
    , decoder_conv_t_filters = [64,32,1]
    , decoder_conv_t_kernel_size = [3,3,3,3]
    , decoder_conv_t_strides = [2,2,1]
    , z_dim = 256
)

DAE_V2.model.summary()

Model: "model_3"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
encoder_input (InputLayer)   [(None, 28, 28, 1)]       0         
_________________________________________________________________
encoder_conv_0 (Conv2D)      (None, 28, 28, 32)        320       
_________________________________________________________________
leaky_re_lu_5 (LeakyReLU)    (None, 28, 28, 32)        0         
_________________________________________________________________
encoder_conv_1 (Conv2D)      (None, 14, 14, 64)        18496     
_________________________________________________________________
leaky_re_lu_6 (LeakyReLU)    (None, 14, 14, 64)        0         
_________________________________________________________________
encoder_conv_2 (Conv2D)      (None, 7, 7, 128)         73856     
_________________________________________________________________
leaky_re_lu_7 (LeakyReLU)    (None, 7, 7, 128)         0         
_________________________________________________________________
decoder_conv_t_0 (Conv2DTran (None, 14, 14, 64)        73792     
_________________________________________________________________
leaky_re_lu_8 (LeakyReLU)    (None, 14, 14, 64)        0         
_________________________________________________________________
decoder_conv_t_1 (Conv2DTran (None, 28, 28, 32)        18464     
_________________________________________________________________
leaky_re_lu_9 (LeakyReLU)    (None, 28, 28, 32)        0         
_________________________________________________________________
decoder_conv_t_2 (Conv2DTran (None, 28, 28, 1)         289       
_________________________________________________________________
activation_1 (Activation)    (None, 28, 28, 1)         0         
=================================================================
Total params: 185,217
Trainable params: 185,217
Non-trainable params: 0
_________________________________________________________________

DAE_V2.compile(0.0005)

DAE_V2.train(x_train_noisy, x_train, batch_size = 128, epochs = 100)

Epoch 1/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0975
Epoch 2/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0968
Epoch 3/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0963
Epoch 4/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0958
Epoch 5/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0955
Epoch 6/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0952
Epoch 7/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0949
Epoch 8/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0947
Epoch 9/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0945
Epoch 10/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0943
Epoch 11/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0942
Epoch 12/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0940
Epoch 13/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0938
Epoch 14/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0937
Epoch 15/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0936
Epoch 16/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0934
Epoch 17/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0933
Epoch 18/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0932
Epoch 19/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0931
Epoch 20/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0930
Epoch 21/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0929
Epoch 22/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0928
Epoch 23/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0927
Epoch 24/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0927
Epoch 25/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0925
Epoch 26/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0924
Epoch 27/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0924
Epoch 28/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0924
Epoch 29/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0923
Epoch 30/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0922
Epoch 31/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0921
Epoch 32/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0921
Epoch 33/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0920
Epoch 34/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0919
Epoch 35/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0918
Epoch 36/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0918
Epoch 37/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0918
Epoch 38/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0917
Epoch 39/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0917
Epoch 40/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0916
Epoch 41/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0915
Epoch 42/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0915
Epoch 43/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0915
Epoch 44/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0914
Epoch 45/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0914
Epoch 46/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0913
Epoch 47/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0913
Epoch 48/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0912
Epoch 49/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0912
Epoch 50/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0912
Epoch 51/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0911
Epoch 52/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0911
Epoch 53/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0911
Epoch 54/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0910
Epoch 55/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0909
Epoch 56/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0910
Epoch 57/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0909
Epoch 58/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0908
Epoch 59/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0908
Epoch 60/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0908
Epoch 61/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0908
Epoch 62/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0908
Epoch 63/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0907
Epoch 64/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0907
Epoch 65/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0906
Epoch 66/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0906
Epoch 67/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0906
Epoch 68/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0906
Epoch 69/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0905
Epoch 70/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0905
Epoch 71/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0905
Epoch 72/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0904
Epoch 73/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0904
Epoch 74/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0904
Epoch 75/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0904
Epoch 76/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0903
Epoch 77/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0903
Epoch 78/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0903
Epoch 79/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0903
Epoch 80/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0902
Epoch 81/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0902
Epoch 82/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0902
Epoch 83/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0902
Epoch 84/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0901
Epoch 85/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0901
Epoch 86/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0901
Epoch 87/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0901
Epoch 88/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0900
Epoch 89/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0900
Epoch 90/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0900
Epoch 91/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0899
Epoch 92/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0900
Epoch 93/100
469/469 [==============================] - 6s 12ms/step - loss: 0.0899
Epoch 94/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0899
Epoch 95/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0899
Epoch 96/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0899
Epoch 97/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0898
Epoch 98/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0898
Epoch 99/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0898
Epoch 100/100
469/469 [==============================] - 6s 13ms/step - loss: 0.0898

n = 10
plt.figure(figsize=(20, 4))
for i in range(n):
    # display original
    ax = plt.subplot(2, n, i + 1)
    plt.imshow(np.reshape(x_test_noisy[i], (28, 28)))
    plt.title(y_test[i])
    plt.gray()
    ax.get_xaxis().set_visible(False)
    ax.get_yaxis().set_visible(False)
plt.show()

png

Oops

잠재공간을 축소했을 때보다 훨씬 좋은 결과를 얻을 수 있습니다. 결과론적으로는 잠재공간이 충분할 경우에 AutoEncoder의 라벨별 경계(Boundary)가 잘 생성될 것이라고 생각할 수 있습니다.
- 하지만 AutoEncoder는 이것을 보장하지 않습니다.(관련 : VAE)
차원을 늘리면서 기존의 loss function을 Binary Cross Entropy로 바꾸어주었습니다.(mse로 진행하면 특정 minima에 빠져버리게됩니다. 이것은 좀 더 자세히 살펴봐야할 것 같습니다.)

n = 10
plt.figure(figsize=(20, 4))
for i in range(n):
    # display original
    ax = plt.subplot(2, n, i + 1)
    plt.imshow(np.reshape(x_test[i], (28, 28)))
    plt.title(y_test[i])
    plt.gray()
    ax.get_xaxis().set_visible(False)
    ax.get_yaxis().set_visible(False)
plt.show()

png

Oops

dae_predict = DAE_V2.model.predict(x_test_noisy)

n = 10
plt.figure(figsize=(20, 4))
for i in range(n):
    # display original
    ax = plt.subplot(2, n, i + 1)
    plt.imshow(np.reshape(dae_predict[i], (28, 28)))
    plt.title(y_test[i])
    plt.gray()
    ax.get_xaxis().set_visible(False)
    ax.get_yaxis().set_visible(False)
plt.show()

png

Oops

Denoising AutoEncoder 모델 구성 Version2 - loss function : MSE

잠재공간의 차원을 늘리자 기존의 MSE Loss function 을 사용하자 loss가 0.1120으로 수렴되고 학습이 더이상 되지 않는 현상을 볼 수 있었습니다.

DAE_V2_MSE = Denoising_Autoencoder_V2(
    input_dim = (28,28,1)
    , encoder_conv_filters = [32,64,128]
    , encoder_conv_kernel_size = [3,3,3,3]
    , encoder_conv_strides = [1,2,2]
    , decoder_conv_t_filters = [64,32,1]
    , decoder_conv_t_kernel_size = [3,3,3,3]
    , decoder_conv_t_strides = [2,2,1]
    , z_dim = 256
)

DAE_V2_MSE.compile(0.0005, loss_func = "mse")

DAE_V2_MSE.train(x_train_noisy, x_train, batch_size = 128, epochs = 10)

Epoch 1/10
469/469 [==============================] - 6s 13ms/step - loss: 0.1153
Epoch 2/10
469/469 [==============================] - 6s 12ms/step - loss: 0.1120
Epoch 3/10
469/469 [==============================] - 6s 12ms/step - loss: 0.1120
Epoch 4/10
469/469 [==============================] - 6s 13ms/step - loss: 0.1120
Epoch 5/10
469/469 [==============================] - 6s 13ms/step - loss: 0.1120
Epoch 6/10
469/469 [==============================] - 6s 13ms/step - loss: 0.1120
Epoch 7/10
469/469 [==============================] - 6s 13ms/step - loss: 0.1120
Epoch 8/10
469/469 [==============================] - 6s 13ms/step - loss: 0.1120
Epoch 9/10
469/469 [==============================] - 6s 12ms/step - loss: 0.1120
Epoch 10/10
469/469 [==============================] - 6s 12ms/step - loss: 0.1120

결론 :
- Denoising AutoEncoder의 성능을 높이려면?
  - loss function과 잠재공간을 여유롭게 주는 방법들이 있을 것으로 보입니다.
  - 하지만 학습속도를 생각했을 때 input 이미지의 차원을 고려하여 설계를 하는 것이 맞을 것으로 예상되며, 너무 넓은 잠재공간이 있을 시에는 의미없는 공간들이 많이 생성될 것으로 생각됩니다.

code : https://github.com/Chanjun-kim/Chanjun-kim.github.io/blob/main/_ipynb/2021-07-27-AE3_DenoisingAE.ipynb

참고자료 : https://www.tensorflow.org/tutorials/generative/autoencoder?hl=ko
참고자료 : https://github.com/davidADSP/GDL_code
참고자료 : https://wikidocs.net/3413

Written on July 27th, 2021 by Chanjun Kim

Feel free to share!

AE3. Denoising AE - 노이즈제거 오토인코더란?

학습목적

Denosing AutoEncoder

왜 노이즈 제거에 AE를 사용할까?

오토인코더 실습

Denoising AutoEncoder 모델 구성

Denoising AutoEncoder 모델 구성 Version2

Denoising AutoEncoder 모델 구성 Version2 - loss function : MSE

결론 :

You may also enjoy:

AE1. Auto Encoder란 무엇일까?

Loss function for Regression Model

CNN에 관하여