● 케라스 전이학습(tramsfer learning)
- 새로운 모델을 만들때 기존에 학습된 모델을 사용
import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Conv2D, MaxPool2D, Dense, Flatten, BatchNormalization, Activation
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.applications import *
# 예시로 학습된 vgg 데이터 불러오기
vgg16 = VGG16(weights = 'imagenet',
input_shape = (32, 32, 3), include_top = False)
model = Sequential()
model.add(vgg16)
model.add(Flatten())
model.add(Dense(256))
model.add(BatchNormalization())
model.add(Activation('relu'))
model.add(Dense(10, activation = 'softmax'))
model.summary()
# 출력 결과
Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
vgg16 (Functional) (None, 1, 1, 512) 14714688
flatten (Flatten) (None, 512) 0
dense (Dense) (None, 256) 131328
batch_normalization (BatchN (None, 256) 1024
ormalization)
activation (Activation) (None, 256) 0
dense_1 (Dense) (None, 10) 2570
=================================================================
Total params: 14,849,610
Trainable params: 14,849,098
Non-trainable params: 512
_________________________________________________________________- vgg16 이외에 MobileNet, ResNet50, Xceoption 모델 등이 존재하여 전이 학습에 이용가능
1. 예제: Dogs vs Cats
- Kaggle 데이터 이용(https://www.kaggle.com/c/dogs-vs-cats/data)
- ImageDataGenerator의 flow_from_directory 이용
- modules import
import tensorflow as tf
from tensorflow.keras.preprocessing.image import array_to_img, img_to_array, load_img, ImageDataGenerator
from tensorflow.keras.layers import Conv2D, Flatten, MaxPool2D, Input, Dropout, Dense
from tensorflow.keras import Model
from tensorflow.keras.optimizers import Adam
import os
import zipfile
import matplotlib.image as mpimg
import matplotlib.pyplot as plt
- 데이터 로드
# 외부에서 데이터 가져오기
import wget
wget.download("https://storage.googleapis.com/mledu-datasets/cats_and_dogs_filtered.zip")
# 압축 해제
local_zip = 'cats_and_dogs_filtered.zip'
zip_ref = zipfile.ZipFile(local_zip, 'r')
# 현재 폴더에 압축해제
zip_ref.extractall()
zip_ref.close()
# 압축해제된 폴더를 기본 경로로 지정, 폴더 내의 train과 validation 폴더에 각각 접근
base_dir = 'cats_and_dogs_filtered'
train_dir = os.path.join(base_dir, 'train')
validation_dir = os.path.join(base_dir, 'validation')
# 압축해제된 폴더 내의 train cat, validation cat, train dog, validation dog 폴더에 각각 접근
train_cats_dir = os.path.join(train_dir, 'cats')
train_dogs_dir = os.path.join(train_dir, 'dogs')
validation_cats_dir = os.path.join(validation_dir, 'cats')
validation_dogs_dir = os.path.join(validation_dir, 'dogs')
train_cat_frames = os.listdir(train_cats_dir)
train_dog_frames = os.listdir(train_dogs_dir)
- 이미지 보강된 데이터 확인
# ImageDataGenerator 정의
datagen = ImageDataGenerator(
rotation_range = 40,
width_shift_range = 0.2,
height_shift_range = 0.2,
shear_range = 0.2,
zoom_range = 0.2,
horizontal_flip = True,
fill_mode = 'nearest'
)
# 이미지 로드
img_path = os.path.join(train_cats_dir, train_cat_frames[2])
img = load_img(img_path, target_size = (150, 150))
x = img_to_array(img)
x = x.reshape((1, ) + x.shape)
i = 0
for batch in datagen.flow(x, batch_size = 1):
plt.figure(i)
imgplot = plt.imshow(array_to_img(batch[0]))
i += 1
if i % 5 == 0:
break
- 학습, 검증 데이터셋의 Data Generator
train_datagen = ImageDataGenerator(
rescale = 1. / 255,
rotation_range = 40,
width_shift_range = 0.2,
height_shift_range = 0.2,
shear_range = 0.2,
zoom_range = 0.2,
horizontal_flip = True
)
train_generator = train_datagen.flow_from_directory(
train_dir,
target_size = (150, 150),
batch_size = 20,
class_mode = 'binary'
)
val_datagen = ImageDataGenerator(rescale = 1. / 255)
validation_generator = val_datagen.flow_from_directory(
validation_dir,
target_size = (150, 150),
batch_size = 20,
class_mode = 'binary'
)
# 출력 결과
Found 2000 images belonging to 2 classes.
Found 1000 images belonging to 2 classes.
- 모델 구성 및 컴파일
model = Sequential()
model.add(Conv2D(32, (3, 3), activation = 'relu', input_shape = (150, 150, 3)))
model.add(MaxPool2D(2, 2))
model.add(Conv2D(64, (3, 3), activation = 'relu'))
model.add(MaxPool2D(2, 2))
model.add(Conv2D(128, (3, 3), activation = 'relu'))
model.add(MaxPool2D(2, 2))
model.add(Conv2D(128, (3, 3), activation = 'relu'))
model.add(MaxPool2D(2, 2))
model.add(Flatten())
model.add(Dropout(0.5))
model.add(Dense(512, activation = 'relu'))
model.add(Dense(1, activation = 'sigmoid'))
model.compile(loss = 'binary_crossentropy',
optimizer = Adam(learning_rate = 1e-4),
metrics = ['acc'])
model.summary()
# 출력 결과
Model: "sequential_1"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d (Conv2D) (None, 148, 148, 32) 896
max_pooling2d (MaxPooling2D (None, 74, 74, 32) 0
)
conv2d_1 (Conv2D) (None, 72, 72, 64) 18496
max_pooling2d_1 (MaxPooling (None, 36, 36, 64) 0
2D)
conv2d_2 (Conv2D) (None, 34, 34, 128) 73856
max_pooling2d_2 (MaxPooling (None, 17, 17, 128) 0
2D)
conv2d_3 (Conv2D) (None, 15, 15, 128) 147584
max_pooling2d_3 (MaxPooling (None, 7, 7, 128) 0
2D)
flatten_1 (Flatten) (None, 6272) 0
dropout (Dropout) (None, 6272) 0
dense_2 (Dense) (None, 512) 3211776
dense_3 (Dense) (None, 1) 513
=================================================================
Total params: 3,453,121
Trainable params: 3,453,121
Non-trainable params: 0
_________________________________________________________________
- 모델 학습 및 학습 과정 시각화
history = model.fit(train_generator,
steps_per_epoch = 100,
epochs = 30,
batch_size = 256,
validation_data = validation_generator,
validation_steps = 50,
verbose = 2)
# 시각화
acc = history.history['acc']
val_acc = history.history['val_acc']
loss = history.history['loss']
val_loss = history.history['val_loss']
epochs = range(len(acc))
plt.plot(epochs, loss, 'b--', label = 'Train Loss')
plt.plot(epochs, val_loss, 'b--', label = 'Validation Loss')
plt.grid()
plt.legend()
plt.plot(epochs, acc, 'b--', label = 'Train Accuracy')
plt.plot(epochs, val_acc, 'b--', label = 'Validation Accuracy')
plt.grid()
plt.legend()
plt.show()
- 모델 저장
model.save('cats_and_dogs_model.h5')
- 사전 훈련된 모델 사용
from tensorflow.keras.optimizers import RMSprop
conv_base = VGG16(weights = 'imagenet',
input_shape = (150, 150, 3), include_top = False)
def build_model_with_pretrained(convbase):
model = Sequential()
model.add(conv_base)
model.add(Flatten())
model.add(Dense(256, activation = 'relu'))
model.add(Dense(1, activation = 'sigmoid'))
model.compile(loss = binary_crossentropy,
optimizer = RMSprop(learning_rate = 2e-5),
metrics = ['accuracy'])
return model- 파라미터 수 확인
model.build_model_with_pretrained(conv_base)
model.summary()
# 출력 결과
Model: "sequential_2"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
vgg16 (Functional) (None, 4, 4, 512) 14714688
flatten_2 (Flatten) (None, 8192) 0
dense_4 (Dense) (None, 256) 2097408
dense_5 (Dense) (None, 1) 257
=================================================================
Total params: 16,812,353
Trainable params: 16,812,353
Non-trainable params: 0
_________________________________________________________________
- 레이어 동결
- 훈련하기 전, 합성곱 기반 레이어들의 가중치 학습을 막기 위해 이를 동결
# 동결 전
print(len(model.trainable_weights))
# 출력 결과
30
# 동결 후
conv_base.trainable = False
print(len(model.trainable_weights))
# 출력 결과
4
- 모델 컴파일
- trainable 속성을 변경했기 때문에 다시 모델을 컴파일 해야함
model.compile(loss = 'binary_crossentropy',
optimizer = RMSprop(learning_rate = 2e-5),
metrics = ['accuracy'])
- 이미지 제너레이터
train_datagen = ImageDataGenerator(
rescale = 1. / 255,
rotation_range = 40,
width_shift_range = 0.2,
height_shift_range = 0.2,
shear_range = 0.2,
zoom_range = 0.2,
horizontal_flip = True
)
train_generator = train_datagen.flow_from_directory(
train_dir,
target_size = (150, 150),
batch_size = 20,
class_mode = 'binary'
)
val_datagen = ImageDataGenerator(rescale = 1. / 255)
validation_generator = val_datagen.flow_from_directory(
validation_dir,
target_size = (150, 150),
batch_size = 20,
class_mode = 'binary'
)
# 출력 결과
Found 2000 images belonging to 2 classes.
Found 1000 images belonging to 2 classes.
- 모델 재학습
history2 = model.fit(train_generator,
steps_per_epoch = 100,
epochs = 30,
batch_size = 256,
validation_data = validation_generator,
validation_steps = 50,
verbose = 2)
acc = history2.history['accuracy']
val_acc = history2.history['val_accuracy']
loss = history2.history['loss']
val_loss = history2.history['val_loss']
epochs = range(len(acc))
plt.plot(epochs, loss, 'b--', label = 'Train Loss')
plt.plot(epochs, val_loss, 'r:', label = 'Validation Loss')
plt.grid()
plt.legend()
plt.plot(epochs, acc, 'b--', label = 'Train Accuracy')
plt.plot(epochs, val_acc, 'r:', label = 'Validation Accuracy')
plt.grid()
plt.legend()
plt.show()
- 모델 저장
model.save('cats_and_dogs_with_pretrained_model.h5')
2. Feature Map 시각화
- 모델 구성
import numpy as np
from tensorflow.keras.models import load_model
from tensorflow.keras.preprocessing import image
# 저장된 모델 로드
model = load_model('cats_and_dogs_model.h5')
model.summary()
# 출력 결과
Model: "sequential_1"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d (Conv2D) (None, 148, 148, 32) 896
max_pooling2d (MaxPooling2D (None, 74, 74, 32) 0
)
conv2d_1 (Conv2D) (None, 72, 72, 64) 18496
max_pooling2d_1 (MaxPooling (None, 36, 36, 64) 0
2D)
conv2d_2 (Conv2D) (None, 34, 34, 128) 73856
max_pooling2d_2 (MaxPooling (None, 17, 17, 128) 0
2D)
conv2d_3 (Conv2D) (None, 15, 15, 128) 147584
max_pooling2d_3 (MaxPooling (None, 7, 7, 128) 0
2D)
flatten_1 (Flatten) (None, 6272) 0
dropout (Dropout) (None, 6272) 0
dense_2 (Dense) (None, 512) 3211776
dense_3 (Dense) (None, 1) 513
=================================================================
Total params: 3,453,121
Trainable params: 3,453,121
Non-trainable params: 0
_________________________________________________________________img_path = 'cats_and_dogs_filtered/validation/dogs/dog.2000.jpg'
img = image.load_img(img_path, target_size = (150, 150))
img_tensor = image.img_to_array(img)
img_tensor = img_tensor[np.newaxis, ...]
img_tensor /= 255.
print(img_tensor.shape)
# 출력 결과
(1, 150, 150, 3)plt.imshow(img_tensor[0])
plt.show()
# 레이어 중 일부만(8개) 출력
conv_output = [layer.output for layer in model.layer[:8]]
conv_output
# 출력 결과
[<KerasTensor: shape=(None, 148, 148, 32) dtype=float32 (created by layer 'conv2d')>,
<KerasTensor: shape=(None, 74, 74, 32) dtype=float32 (created by layer 'max_pooling2d')>,
<KerasTensor: shape=(None, 72, 72, 64) dtype=float32 (created by layer 'conv2d_1')>,
<KerasTensor: shape=(None, 36, 36, 64) dtype=float32 (created by layer 'max_pooling2d_1')>,
<KerasTensor: shape=(None, 34, 34, 128) dtype=float32 (created by layer 'conv2d_2')>,
<KerasTensor: shape=(None, 17, 17, 128) dtype=float32 (created by layer 'max_pooling2d_2')>,
<KerasTensor: shape=(None, 15, 15, 128) dtype=float32 (created by layer 'conv2d_3')>,
<KerasTensor: shape=(None, 7, 7, 128) dtype=float32 (created by layer 'max_pooling2d_3')>]activation_model = Model(inputs = [model.input], outputs = conv_output)
activations = activation_model.predict(img_tensor)
len(activations)
# 출력 결과
8
- 시각화
print(activations[0].shape)
plt.matshow(activations[0][0, :, :, 7], cmap = 'viridis')
plt.show()
# 출력 결과
(1, 148, 148, 32)
print(activations[0].shape)
plt.matshow(activations[0][0, :, :, 10], cmap = 'viridis')
plt.show()
# 출력 결과
(1, 148, 148, 32)
- 중간의 모든 활성화에 대해 시각화
# 각 layer에서 이미지의 변환과정을 시각화
layer_names = []
for layer in model.layers[:8]:
layer_names.append(layer.name)
images_per_row = 16
for layer_name, layer_activation in zip(layer_names, activations):
num_features = layer_activation.shape[-1]
size = layer_activation.shape[1]
num_cols = num_features // images_per_row
display_grid = np.zeros((size * num_cols, size * images_per_row))
for col in range(num_cols):
for row in range(images_per_row):
channel_image = layer_activation[0, :, :, col * images_per_row + row]
channel_image -= channel_image.mean()
channel_image /= channel_image.std()
channel_image *= 64
channel_image += 128
channel_image =np.clip(channel_image, 0, 255).astype('unit8')
display_grid[col * size : (col + 1) * size, row * size : (row + 1) * size] = channel_image
scale = 1. / size
plt.figure(figsize = (scale * display_grid.shape[1],
scale * display_grid.shape[0]))
plt.title(layer_name)
plt.grid(False)
plt.imshow(display_grid, aspect = 'auto', cmap = 'viridis')
plt.show()
'Python > Deep Learning' 카테고리의 다른 글
| [딥러닝-케라스] 케라스 텍스트 처리 및 임베딩(2) (0) | 2023.05.18 |
|---|---|
| [딥러닝-케라스] 케라스 텍스트 처리 및 임베딩(1) (0) | 2023.05.17 |
| [딥러닝-케라스] 케라스 CIFAR10 CNN 모델 (0) | 2023.05.16 |
| [딥러닝-케라스] 케라스 Fashion MNIST CNN 모델 (0) | 2023.05.14 |
| [딥러닝-케라스] 케라스 컨볼루션 신경망 (0) | 2023.05.13 |