Diary of Chanjun 데이터 분석가의 다이어리
AE1. Auto Encoder란 무엇일까?
학습목적
이 포스팅에선 GAN을 배우기 전 이미지를 생성하는 오토인코더에 대해서 알아보도록 하겠습니다.
오토인코더란?
- 오토인코더(Autoencoder)의 기본적인 개념은 저차원(직관적)인 데이터를 고차원(추상적)의 데이터로 표현(representation)하고 다시 저차원의 데이터로 만들어 데이터 생성, 노이즈 제거, 특성 변화 등을 할 수 있도록 만들어주는 기법으로, 오토 인코더라고 불리지만 실제로는 인코더와 디코더가 같이 묶여있는 네트워크입니다.
- 저차원과 고차원
- 인코더는 차원 축소라고 표현되나, 잠재공간(Latent Space)가 늘어나며 추상적인 고차원 데이터로 변환됩니다..
- 128x128x3의 이미지는 사람의 눈으로도 판단 가능한 직관적인 데이터지만 2x2x32 같이 잠재공간(Latent Space)가 많은 경우는 사람의 눈으로는 판단할 수 없는 추상적인 데이터라고 볼 수 있습니다.
- 결국 차원이란 축에 따라 달라지는 표현일 수 있다는 것을 생각해야합니다.
- 저차원과 고차원
출처: https://excelsior-cjh.tistory.com/187
오토인코더 실습
from google.colab import drive
drive.mount("/content/drive")
%cd drive/MyDrive/cj/portfolio/Chanjun-kim.github.io/_ipynb
!pwd
Mounted at /content/drive
/content/drive/MyDrive/cj/portfolio/Chanjun-kim.github.io/_ipynb
/content/drive/MyDrive/cj/portfolio/Chanjun-kim.github.io/_ipynb
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')
class Autoencoder():
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
)
출처 : https://github.com/davidADSP/GDL_code
- 설명
- 최초의 데이터를 28 x 28 데이터를 인풋으로 받습니다.
- Encoder와 Decoder의 깊이는 len(encoder_conv_filters), len(decoder_conv_t_filters)로 정해지므로 각각 4개의 Convolution 층을 갖게 됩니다.
- Convolution Filter의 차원은 1(Grey Scale Input) - (32 - 64 - 64 -64, Encoder Convolution Filter) - 2(represention) - (64 - 64 - 32, Decoder Convolution Filter) - 1(Grey Scale Output) 으로 구성됩니다.
- Encoder와 Decoder 모두 Activation Function은 LeakyRelu를 활용합니다.
- kernel size는 3x3으로 진행됩니다.
- stride를 2로 주어 차원 1/2로 축소합니다.
- encoder의 최종 차원을 2차원으로 만들어 잠재공간에 어떻게 분포되는지 확인하겠습니다.
- decoder는 encoder에 의해 줄어진 차원을 다시 키워줄 수 있도록 Conv2DTranspose를 사용하여 28 x 28의 크기로 점차 늘려줍니다.
- loss function은 MSE를 사용하고 최종 Activation function은 Sigmoid를 사용합니다.
- Optimizer는 Adam을 활용합니다.
- Auto Encoder 같은 경우는 Input과 Output이 같습니다. 그렇기 때문에 마지막 fit 부분에 x = x_train, y = x_train으로 들어가게 됩니다.
- 주의점
- 네트워크 상에는 BN Layer를 추가할 수 있게 되어있지만 사실 사용하면 안됩니다. batch normalization 이름에서 나오듯 특정 데이터를 가지고만 normalization을 하기 때문에 generator 모델에서는 그 값이 대표하지 않을 가능성이 농후하기 때문입니다.
- 여기서 2차원까지 줄인건 mnist데이터셋의 차원이 충분히 표현될 수 있기 때문입니다. 복잡한 이미지셋은 당연히 더 고차원으로 줄여야합니다.
AE = Autoencoder(
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
)
AE.compile(0.00005)
AE.model.summary()
Model: "model_2"
_________________________________________________________________
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 (LeakyReLU) (None, 28, 28, 32) 0
_________________________________________________________________
encoder_conv_1 (Conv2D) (None, 14, 14, 64) 18496
_________________________________________________________________
leaky_re_lu_1 (LeakyReLU) (None, 14, 14, 64) 0
_________________________________________________________________
encoder_conv_2 (Conv2D) (None, 7, 7, 64) 36928
_________________________________________________________________
leaky_re_lu_2 (LeakyReLU) (None, 7, 7, 64) 0
_________________________________________________________________
encoder_conv_3 (Conv2D) (None, 7, 7, 64) 36928
_________________________________________________________________
leaky_re_lu_3 (LeakyReLU) (None, 7, 7, 64) 0
_________________________________________________________________
flatten (Flatten) (None, 3136) 0
_________________________________________________________________
encoder_output (Dense) (None, 2) 6274
_________________________________________________________________
model_1 (Functional) (None, 28, 28, 1) 102017
=================================================================
Total params: 200,963
Trainable params: 200,963
Non-trainable params: 0
_________________________________________________________________
AE.encoder.summary()
Model: "model"
_________________________________________________________________
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 (LeakyReLU) (None, 28, 28, 32) 0
_________________________________________________________________
encoder_conv_1 (Conv2D) (None, 14, 14, 64) 18496
_________________________________________________________________
leaky_re_lu_1 (LeakyReLU) (None, 14, 14, 64) 0
_________________________________________________________________
encoder_conv_2 (Conv2D) (None, 7, 7, 64) 36928
_________________________________________________________________
leaky_re_lu_2 (LeakyReLU) (None, 7, 7, 64) 0
_________________________________________________________________
encoder_conv_3 (Conv2D) (None, 7, 7, 64) 36928
_________________________________________________________________
leaky_re_lu_3 (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
_________________________________________________________________
AE.decoder.summary()
Model: "model_1"
_________________________________________________________________
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_4 (LeakyReLU) (None, 7, 7, 64) 0
_________________________________________________________________
decoder_conv_t_1 (Conv2DTran (None, 14, 14, 64) 36928
_________________________________________________________________
leaky_re_lu_5 (LeakyReLU) (None, 14, 14, 64) 0
_________________________________________________________________
decoder_conv_t_2 (Conv2DTran (None, 28, 28, 32) 18464
_________________________________________________________________
leaky_re_lu_6 (LeakyReLU) (None, 28, 28, 32) 0
_________________________________________________________________
decoder_conv_t_3 (Conv2DTran (None, 28, 28, 1) 289
_________________________________________________________________
activation (Activation) (None, 28, 28, 1) 0
=================================================================
Total params: 102,017
Trainable params: 102,017
Non-trainable params: 0
_________________________________________________________________
mnist
(x_train, y_train), (x_test, y_test) = mnist.load_data()
Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/mnist.npz
11493376/11490434 [==============================] - 0s 0us/step
n = 10
plt.figure(figsize=(20, 4))
for i in range(n):
# display original
ax = plt.subplot(2, n, i + 1)
plt.imshow(x_train[i])
plt.title(y_train[i])
plt.gray()
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
plt.show()
- 픽셀값을 0~1까지로 정규화해주고, keras에서 학습할 수 있도록 차원을 하나 늘려준다.(grey scale)
x_train, x_test = x_train / 255.0, x_test / 255.0
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
x_train.shape
(60000, 28, 28, 1)
- X값이 곧 Y값이 되므로 x, y 모두에 x_train을 넣어준다
AE.train(x_train, x_train, batch_size = 32, epochs = 100)
Epoch 1/100
1875/1875 [==============================] - 40s 4ms/step - loss: 0.0724
Epoch 2/100
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0572
Epoch 3/100
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0546
Epoch 4/100
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0531
Epoch 5/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0519
Epoch 6/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0507
Epoch 7/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0497
Epoch 8/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0489
Epoch 9/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0481
Epoch 10/100
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0475
Epoch 11/100
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0470
Epoch 12/100
1875/1875 [==============================] - 8s 5ms/step - loss: 0.0466
Epoch 13/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0463
Epoch 14/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0459
Epoch 15/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0456
Epoch 16/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0454
Epoch 17/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0451
Epoch 18/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0449
Epoch 19/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0447
Epoch 20/100
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0445
Epoch 21/100
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0443
Epoch 22/100
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0442
Epoch 23/100
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0440
Epoch 24/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0438
Epoch 25/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0437
Epoch 26/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0436
Epoch 27/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0434
Epoch 28/100
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0433
Epoch 29/100
1875/1875 [==============================] - 8s 5ms/step - loss: 0.0432
Epoch 30/100
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0430
Epoch 31/100
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0429
Epoch 32/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0428
Epoch 33/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0427
Epoch 34/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0426
Epoch 35/100
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0425
Epoch 36/100
1875/1875 [==============================] - 8s 5ms/step - loss: 0.0424
Epoch 37/100
1875/1875 [==============================] - 8s 5ms/step - loss: 0.0423
Epoch 38/100
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0422
Epoch 39/100
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0421
Epoch 40/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0420
Epoch 41/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0420
Epoch 42/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0419
Epoch 43/100
1875/1875 [==============================] - 8s 5ms/step - loss: 0.0418
Epoch 44/100
1875/1875 [==============================] - 8s 5ms/step - loss: 0.0417
Epoch 45/100
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0417
Epoch 46/100
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0416
Epoch 47/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0415
Epoch 48/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0415
Epoch 49/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0414
Epoch 50/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0413
Epoch 51/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0413
Epoch 52/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0412
Epoch 53/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0411
Epoch 54/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0411
Epoch 55/100
1875/1875 [==============================] - 8s 5ms/step - loss: 0.0410
Epoch 56/100
1875/1875 [==============================] - 8s 5ms/step - loss: 0.0410
Epoch 57/100
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0409
Epoch 58/100
1875/1875 [==============================] - 8s 5ms/step - loss: 0.0408
Epoch 59/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0408
Epoch 60/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0407
Epoch 61/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0407
Epoch 62/100
1875/1875 [==============================] - 8s 5ms/step - loss: 0.0406
Epoch 63/100
1875/1875 [==============================] - 8s 5ms/step - loss: 0.0406
Epoch 64/100
1875/1875 [==============================] - 8s 5ms/step - loss: 0.0405
Epoch 65/100
1875/1875 [==============================] - 8s 5ms/step - loss: 0.0405
Epoch 66/100
1875/1875 [==============================] - 8s 5ms/step - loss: 0.0404
Epoch 67/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0404
Epoch 68/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0404
Epoch 69/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0403
Epoch 70/100
1875/1875 [==============================] - 8s 5ms/step - loss: 0.0403
Epoch 71/100
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0402
Epoch 72/100
1875/1875 [==============================] - 8s 5ms/step - loss: 0.0402
Epoch 73/100
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0402
Epoch 74/100
1875/1875 [==============================] - 8s 5ms/step - loss: 0.0401
Epoch 75/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0401
Epoch 76/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0401
Epoch 77/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0400
Epoch 78/100
1875/1875 [==============================] - 8s 5ms/step - loss: 0.0400
Epoch 79/100
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0400
Epoch 80/100
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0399
Epoch 81/100
1875/1875 [==============================] - 8s 5ms/step - loss: 0.0399
Epoch 82/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0399
Epoch 83/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0398
Epoch 84/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0398
Epoch 85/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0398
Epoch 86/100
1875/1875 [==============================] - 8s 5ms/step - loss: 0.0397
Epoch 87/100
1875/1875 [==============================] - 8s 5ms/step - loss: 0.0397
Epoch 88/100
1875/1875 [==============================] - 8s 5ms/step - loss: 0.0397
Epoch 89/100
1875/1875 [==============================] - 8s 5ms/step - loss: 0.0396
Epoch 90/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0396
Epoch 91/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0396
Epoch 92/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0395
Epoch 93/100
1875/1875 [==============================] - 8s 5ms/step - loss: 0.0395
Epoch 94/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0395
Epoch 95/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0395
Epoch 96/100
1875/1875 [==============================] - 9s 5ms/step - loss: 0.0394
Epoch 97/100
1875/1875 [==============================] - 8s 5ms/step - loss: 0.0394
Epoch 98/100
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0394
Epoch 99/100
1875/1875 [==============================] - 8s 5ms/step - loss: 0.0394
Epoch 100/100
1875/1875 [==============================] - 8s 4ms/step - loss: 0.0393
latent_splace_x_test = pd.DataFrame(AE.encoder.predict(x_test), columns = ["x", "y"])
latent_splace_x_test["label"] = y_test
latent_splace_x_test["label"] = latent_splace_x_test.label.astype(str)
- 각각의 숫자마다 Encoder를 거친 잠재공간을 뿌려보았습니다.
- 대부분의 숫자들이 자기만의 영역을 가지고 있습니다.
- 하지만 3, 5, 8이 가까이 붙어있고, 4, 9가 가까이 붙어있습니다.
- 이것이 오토인코더를 거칠 때 어떤 영향을 미치는 지 validation set을 통하여 알아보겠습니다.
(
ggplot() +
geom_point(data = latent_splace_x_test, mapping = aes(x = "x", y = "y", color = "label"), alpha = 0.5) +
geom_label(data = latent_splace_x_test.groupby("label").mean(["x", "y"]).reset_index(), mapping = aes(x = "x", y = "y", label = "label"), size = 10, alpha = 0.5) +
stat_hull(data = latent_splace_x_test, mapping = aes(x = "x", y = "y", color = "label")) +
theme_bw()
)
<ggplot: (8763240984113)>
- 오토인코더를 거친 숫자들입니다.
- 잘 구분된 숫자들은 확실히 그 숫자들을 띄는 것 같지만, 4, 5, 9 같이 자신의 영역을 확실히 구분하지 못한 숫자들은 뭉개지거나 혹은 다른 숫자 더 비슷하게 보이기도 합니다.
predict_x = AE.model.predict(x_test)
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(predict_x[i], (28, 28)))
plt.title(y_test[i])
plt.gray()
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
plt.show()
# 비슷한 좌표 생성
latent_convex = latent_splace_x_test.groupby("label")[["x", "y"]].apply(lambda x : ConvexHull(x))
latent_convex = pd.DataFrame(latent_convex.apply(lambda x : x.points[random.randint(0, 100)]).to_list(), columns = ["x", "y"]).reset_index().rename(columns = {"index" : "label"})
latent_convex["label"] = latent_convex["label"].astype(str)
(
ggplot() +
geom_point(data = latent_splace_x_test, mapping = aes(x = "x", y = "y", color = "label"), alpha = 0.5) +
stat_hull(data = latent_splace_x_test, mapping = aes(x = "x", y = "y", color = "label")) +
geom_point(data = latent_convex, mapping = aes(x = "x", y = "y"), color = "red") +
geom_label(data = latent_convex, mapping = aes(x = "x", y = "y", label = "label"), size = 10, alpha = 0.5) +
theme_bw()
)
<ggplot: (8763250292769)>
predict_decoder = AE.decoder.predict(np.reshape(latent_convex[["x", "y"]].to_numpy(), (10, 2, 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(predict_decoder[i], (28, 28)))
plt.title(latent_convex.label[i])
plt.gray()
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
plt.show()
# 엉뚱한 좌표 생성
random_points = [[random.randint(int(latent_splace_x_test.x.min()), int(latent_splace_x_test.x.max())) + random.random(), random.randint(int(latent_splace_x_test.y.min()), int(latent_splace_x_test.y.max())) + random.random()] for _ in range(10)]
random_points = pd.DataFrame(random_points, columns = ["x", "y"]).reset_index().rename(columns = {"index" : "label"})
random_points["label"] = [f"random{i}" for i in range(10)]
(
ggplot() +
geom_point(data = latent_splace_x_test, mapping = aes(x = "x", y = "y", color = "label"), alpha = 0.5) +
stat_hull(data = latent_splace_x_test, mapping = aes(x = "x", y = "y", color = "label")) +
geom_point(data = random_points, mapping = aes(x = "x", y = "y"), color = "red") +
geom_label(data = random_points, mapping = aes(x = "x", y = "y", label = "label"), size = 10, alpha = 0.5) +
theme_bw()
)
<ggplot: (8763239487357)>
predict_decoder = AE.decoder.predict(np.reshape(random_points[["x", "y"]].to_numpy(), (10, 2, 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(predict_decoder[i], (28, 28)))
plt.title(random_points.label[i])
plt.gray()
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
plt.show()
-
결론 :
- 오토인코더를 왜 쓰는가?
- 인코더를 거치면 일종의 차원축소처럼 보이게되는데 이 때 확실한 영역을 구분하게 된다면, 그 주변의 좌표들을 통하여 비슷한 이미지를 생성하는데 기여할 수 있습니다.
- 오터인코더의 한계?
- 모든 데이터가 인코더를 거치며 정확히 구분되지 않을 뿐 아니라 잠재 공간 모두에 분포하지 않습니다. 즉, 잠재공간이 존재하지만 엉뚱한 곳을 찍었을 때는 전혀 다른 데이터를 생성할 수 있습니다.
- 오토인코더를 왜 쓰는가?
code : https://github.com/Chanjun-kim/Chanjun-kim.github.io/blob/main/_ipynb/2021-07-25-AE1_AutoEncoder.ipynb
Written on July 25th, 2021 by Chanjun Kim참고자료 : https://www.tensorflow.org/tutorials/generative/autoencoder?hl=ko
참고자료 : https://github.com/davidADSP/GDL_code
Feel free to share!