스터디노트 (DeepLearning 13)

Transfer Learning - 오토인코더

• colab에서 작성

• 오토인코더는 입력과 출력이 동일하다
  - 자기 자신을 재생성하는 네트워크
• Latent Vector : 잠재변수
• Encoder : 입력쪽
• Decoder : 출력쪽
• 인코더는 일종의 특징추출기와 같은 역할
• 디코더는 압축된 데이터를 다시 복원하는 역할

# tensorflow

import tensorflow as tf
import numpy as np

# mnist

(train_X, train_Y), (test_X, test_Y) = tf.keras.datasets.mnist.load_data()
print(train_X.shape, train_Y.shape)

train_X = train_X / 255
test_X = test_X / 255

import matplotlib.pyplot as plt

plt.imshow(train_X[0].reshape(28, 28), cmap='gray')
plt.colorbar()
plt.show()

print(train_Y[0])

train_X = train_X.reshape(-1, 28*28)
test_X = test_X.reshape(-1, 28*28)
print(train_X.shape, train_Y.shape)

model = tf.keras.Sequential([
    tf.keras.layers.Dense(784, activation='relu', input_shape = (784,)),
    tf.keras.layers.Dense(64, activation='relu'),
    tf.keras.layers.Dense(784, activation='sigmoid')
    ])

model.compile(optimizer=tf.optimizers.Adam(), loss='mse')
model.summary()

# 학습

model.fit(train_X, train_X, epochs=10, batch_size=256)

# 자기 자신을 예측

import random

plt.figure(figsize=(4, 8))
for c in range(4):
  plt.subplot(4, 2, c*2+1)
  rand_index = random.randint(0, test_X.shape[0])
  plt.imshow(test_X[rand_index].reshape(28, 28), cmap='gray')
  plt.axis('off')

  plt.subplot(4, 2, c*2+2)
  img = model.predict(np.expand_dims(test_X[rand_index], axis = 0))
  plt.imshow(test_X[rand_index].reshape(28, 28), cmap='gray')
  plt.axis('off')

plt.show()

model.evaluate(test_X, test_X)

# CNN 활용

train_X = train_X.reshape(-1, 28, 28, 1)
test_X = test_X.reshape(-1, 28, 28, 1)

model = tf.keras.Sequential([
    tf.keras.layers.Conv2D(filters=32, kernel_size = 2, strides = (2, 2), activation = 'relu', input_shape=(28, 28, 1)), # 14 * 14
    tf.keras.layers.Conv2D(filters=64, kernel_size=2, strides = (2, 2), activation='relu'), # 7 * 7
    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(64, activation='relu'), # 특성
    tf.keras.layers.Dense(7*7*64, activation='relu'),
    tf.keras.layers.Reshape(target_shape=(7, 7, 64)),
    tf.keras.layers.Conv2DTranspose(filters=32, kernel_size=2, strides=(2, 2), padding='same', activation='relu'),
    tf.keras.layers.Conv2DTranspose(filters=1, kernel_size=2, strides=(2, 2), padding = 'same', activation = 'sigmoid')
    ])

model.compile(optimizer = tf.optimizers.Adam(), loss = 'mse')
model.summary()

# 학습

model.fit(train_X, train_X, epochs=20, batch_size=256) # 자기 자신을 집어넣음

# 결과

import random

plt.figure(figsize=(4, 8))
for c in range(4):
  plt.subplot(4, 2, c*2+1)
  rand_index = random.randint(0, test_X.shape[0])
  plt.imshow(test_X[rand_index].reshape(28, 28), cmap='gray')
  plt.axis('off')

  plt.subplot(4, 2, c*2+2)
  img = model.predict(np.expand_dims(test_X[rand_index], axis = 0))
  plt.imshow(img.reshape(28, 28), cmap='gray')
  plt.axis('off')

plt.show()

model.evaluate(test_X, test_X)

---

• ELU

# ELU

import math

x = np.arange(-5, 5, 0.01)
relu = [0 if z < 0 else z for z in x]
elu = [1.0 * (np.exp(z) - 1) if z < 0 else z for z in x]

plt.axvline(0, color='gray')
plt.plot(x, relu, 'r--', label = 'relu')
plt.plot(x, elu, 'g-', label = 'elu')
plt.legend()
plt.show()

---

# 다시 CNN

train_X = train_X.reshape(-1, 28, 28, 1)
test_X = test_X.reshape(-1, 28, 28, 1)

model = tf.keras.Sequential([
    tf.keras.layers.Conv2D(filters=32, kernel_size=2, strides = (2, 2), activation = 'elu', input_shape=(28, 28, 1)),
    tf.keras.layers.Conv2D(filters=64, kernel_size=2, strides = (2, 2), activation='elu'), # 7 * 7
    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(64, activation='elu'), # 특성
    tf.keras.layers.Dense(7*7*64, activation='elu'),
    tf.keras.layers.Reshape(target_shape=(7, 7, 64)),
    tf.keras.layers.Conv2DTranspose(filters=32, kernel_size=2, strides=(2, 2), padding='same', activation='elu'),
    tf.keras.layers.Conv2DTranspose(filters=1, kernel_size=2, strides=(2, 2), padding = 'same', activation = 'sigmoid')
])

model.compile(optimizer = tf.optimizers.Adam(), loss = 'mse')
model.fit(train_X, train_X, epochs = 20, batch_size=256)

# 결과

import random

plt.figure(figsize=(4, 8))
for c in range(4):
  plt.subplot(4, 2, c*2+1)
  rand_index = random.randint(0, test_X.shape[0])
  plt.imshow(test_X[rand_index].reshape(28, 28), cmap='gray')
  plt.axis('off')

  plt.subplot(4, 2, c*2+2)
  img = model.predict(np.expand_dims(test_X[rand_index], axis = 0))
  plt.imshow(img.reshape(28, 28), cmap='gray')
  plt.axis('off')

plt.show()

model.evaluate(test_X, test_X)

# 잠재변수 벡터 확보

latent_vector_model = tf.keras.Model(inputs = model.input, outputs = model.layers[3].output) # 3번째 layer에서 output을 뽑는 것
latent_vector = latent_vector_model.predict(train_X)
print(latent_vector.shape)
print(latent_vector[0])

# 군집

%%time
from sklearn.cluster import KMeans

kmeans = KMeans(n_clusters = 10, n_init = 10, random_state = 42)
kmeans.fit(latent_vector)

# 군집 결과

print(kmeans.labels_)
print(kmeans.cluster_centers_.shape)
print(kmeans.cluster_centers_[0])

# 시각화

import matplotlib.pyplot as plt

plt.figure(figsize=(12, 12))

for i in range(10):
  images = train_X[kmeans.labels_ == i]
  for c in range(10):
    plt.subplot(10, 10, i*10+c+1)
    plt.imshow(images[c].reshape(28, 28), cmap='gray')
    plt.axis('off')

plt.show()

• t- SNE
  - 고차원의 벡터를 저차원으로 옮겨서 시각화에 도움을 주는 방법
  - t Stochastic Nearest Neghbor
  - k - Means가 각 클러스터를 계산하기 위한 단위로 중심과 각 데이터의 거리를 측정한다면
  - t - SNE는 각 데이터의 유사도를 정의하고 원래 공간에서의 유사도와 저차원 공간에서의 유사도가 비슷해지도록 학습시킴
  - 유사도는 수학적으로 확률로 표현

# t-SNE 수행

%%time
from sklearn.manifold import TSNE

tsne = TSNE(n_components = 2, learning_rate = 100, perplexity = 15, random_state = 0)
tsne_vector = tsne.fit_transform(latent_vector[:5000])

cmap = plt.get_cmap('rainbow', 10)
fig = plt.scatter(tsne_vector[:, 0], tsne_vector[:, 1], marker = '.', c = train_Y[:5000], cmap = cmap)
cb = plt.colorbar(fig, ticks = range(10))
n_clusters = 10
tick_locs = (np.arange(n_clusters) + 0.5) * (n_clusters -1) / n_clusters
cb.set_ticks(tick_locs)
cb.set_ticklabels(range(10))

plt.show()

# 최근접이웃의 숫자 perplexities 튜닝

%%time

perplexities = [5, 10, 15, 25, 50, 100]
plt.figure(figsize = (8, 12))

for c in range(6):
  tsne = TSNE(n_components = 2, learning_rate = 100, perplexity = perplexities[c], random_state = 0)
  tsne_vector = tsne.fit_transform(latent_vector[:5000])

  plt.subplot(3, 2, c+1)
  plt.scatter(tsne_vector[:, 0], tsne_vector[:, 1], marker = '.', c = train_Y[:5000], cmap='rainbow')
  plt.title('perplexity : {0}' .format(perplexities[c]))

plt.show()

# 약간의 뎁스를 더 가져간 시각화

from matplotlib.offsetbox import TextArea, DrawingArea, OffsetImage, AnnotationBbox

plt.figure(figsize=(16, 16))

tsne = TSNE(n_components = 2, learning_rate = 100, perplexity = 15, random_state = 0)
tsne_vector = tsne.fit_transform(latent_vector[:5000])

ax = plt.subplot(1, 1, 1)
ax.scatter(tsne_vector[:, 0], tsne_vector[:, 1], marker = '.', c = train_Y[:5000], cmap='rainbow')
for i in range(200):
  imagebox = OffsetImage(train_X[i].reshape(28, 28))
  ab = AnnotationBbox(imagebox, (tsne_vector[i, 0], tsne_vector[i, 1]), frameon=False, pad = 0.0)
  ax.add_artist(ab)

ax.set_xticks([])
ax.set_yticks([])
plt.show()

너무너무 어렵다 ㅠㅠ...
💻 출처 : 제로베이스 데이터 취업 스쿨