CNN(Convolutional Neural Network)-VGG-16

VGG-16 Model

• University of Oxford - Visual Geometry Group
• 2014 ILSVRC 2nd Model
• ImageNet Large Scale Visual Recognition Challenge (ILSVRC)
• 사전 학습된 모델을 사용하여 특성 추출(Feature Extraction)
  • Feature Extraction : Parameter 재사용
  • Classification : Parameter 학습

---

1. Image_File Directory Setting

train_dir = 'train'
valid_dir = 'validation'
test_dir = 'test'

---

2. Import VGG-16 Model

2-1. conv_base

from tensorflow.keras.applications import VGG16

conv_base = VGG16(weights = 'imagenet',
                  include_top = False,
                  input_shape = (150, 150, 3))

---

2-2. Model Information

conv_base.summary()

---

3. Feature Extraction

3-1. 특징추출 함수 정의 : extract_feature( )

from tensorflow.keras.preprocessing.image import ImageDataGenerator
import numpy as np

datagen = ImageDataGenerator(rescale = 1./255)

batch_size = 20

def extract_features(directory, sample_count):
    features = np.zeros(shape = (sample_count, 4, 4, 512))
    labels = np.zeros(shape = (sample_count))
    
    generator = datagen.flow_from_directory(directory,
                                            target_size = (150, 150),
                                            batch_size = batch_size,
                                            class_mode = 'binary')
    
    i = 0
    for inputs_batch, labels_batch in generator:
        features_batch = conv_base.predict(inputs_batch)
        features[i * batch_size : (i + 1) * batch_size] = features_batch
        labels[i * batch_size : (i + 1) * batch_size] = labels_batch
        i += 1
        if i * batch_size >= sample_count:
            break
    return features, labels

---

3-2. 특징추출 함수 적용

%%time  #1분

train_features, train_labels = extract_features(train_dir, 2000)
valid_features, valid_labels = extract_features(valid_dir, 1000)
test_features, test_labels = extract_features(test_dir, 1000)

train_features.shape, valid_features.shape, test_features.shape

((2000, 4, 4, 512), (1000, 4, 4, 512), (1000, 4, 4, 512))

---

3-3. Reshape Features

train_features = np.reshape(train_features, (2000, 4 * 4 * 512))
valid_features = np.reshape(valid_features, (1000, 4 * 4 * 512))
test_features = np.reshape(test_features, (1000, 4 * 4 * 512))

train_features.shape, valid_features.shape, test_features.shape

((2000, 8192), (1000, 8192), (1000, 8192))

---

4. Keras CNN Modeling with VGG-16 Featured Data

4-1. Model Define

from tensorflow.keras import models, layers

model = models.Sequential()
model.add(layers.Dense(256, activation = 'relu', input_dim = 4 * 4 * 512))
model.add(layers.Dropout(0.5))
model.add(layers.Dense(1, activation = 'sigmoid'))

model.summary()

---

4-2. Model Compile

model.compile(loss = 'binary_crossentropy',
              optimizer = 'adam',
              metrics = ['accuracy'])

---

4-3. Model Fit

%%time

Hist_dandc = model.fit(train_features, train_labels,
                       epochs = 100,
                       batch_size = 20,
                       validation_data = (valid_features, valid_labels))

---

4-4. 학습 결과 시각화

#Loss Visualization

import matplotlib.pyplot as plt

epochs = range(1, len(Hist_dandc.history['loss']) + 1)

plt.figure(figsize = (9, 6))
plt.plot(epochs, Hist_dandc.history['loss'])
plt.plot(epochs, Hist_dandc.history['val_loss'])

plt.title('Training & Validation Loss')
plt.xlabel('Epochs')
plt.ylabel('Loss')
plt.legend(['Training Loss', 'Validation Loss'])
plt.grid()
plt.show()

#Accuracy Visualization

import matplotlib.pyplot as plt

epochs = range(1, len(Hist_dandc.history['loss']) + 1)

plt.figure(figsize = (9, 6))
plt.plot(epochs, Hist_dandc.history['accuracy'])
plt.plot(epochs, Hist_dandc.history['val_accuracy'])

plt.title('Training & Validation Accuracy')
plt.xlabel('Epochs')
plt.ylabel('Accuracy')
plt.legend(['Training Accuracy', 'Validation Accuracy'])
plt.grid()
plt.show()

---

4-5. Model Evaluate

loss, accuracy = model.evaluate(test_features, test_labels)

print('Loss = {:.5f}'.format(loss))
print('Accuracy = {:.5f}'.format(accuracy))

---

 

Fine Tunig

1. Data Preprocessing

1-1. ImageDataGenerator( ) & flow_from_directory( )

from tensorflow.keras.preprocessing.image import ImageDataGenerator

train_datagen = ImageDataGenerator(rescale = 1./255)
valid_datagen = ImageDataGenerator(rescale = 1./255)

train_generator = train_datagen.flow_from_directory(
                  train_dir,
                  target_size = (150, 150),
                  batch_size = 20,
                  class_mode = 'binary')

valid_generator = valid_datagen.flow_from_directory(
                  valid_dir,
                  target_size = (150, 150),
                  batch_size = 20,
                  class_mode = 'binary')

---

2. Import VGG-16 Model & Some Layers Freezing

2-1. conv_base

from tensorflow.keras.applications import VGG16

conv_base = VGG16(weights = 'imagenet',
                  include_top = False,
                  input_shape = (150, 150, 3))

---

2-2. Model Information

conv_base.summary()

---

2-3. 'block5_conv1' Freezing

print('conv_base 동결 전 훈련 가능 가중치의 종류:', len(conv_base.trainable_weights))

conv_base 동결 전 훈련 가능 가중치의 종류: 26

set_trainable = False

for layer in conv_base.layers:
    if layer.name == 'block5_conv1':
        set_trainable = True
        
    if set_trainable:
        layer.trainable = True
    else:
        layer.trainable = False

print('conv_base 동결 후 훈련 가능 가중치의 종류:', len(conv_base.trainable_weights))

conv_base 동결 후 훈련 가능 가중치의 종류: 6

conv_base.summary()

---

3.Keras CNN Modeling with VGG-16 Freezed Layers

3-1. Model Define

from tensorflow.keras import models, layers

model = models.Sequential(name = 'CNN_VGG16')
model.add(conv_base)

model.add(layers.Flatten())
model.add(layers.Dropout(0.4))
model.add(layers.Dense(256, activation = 'relu'))
model.add(layers.Dense(1, activation = 'sigmoid'))

model.summary()

---

3-2. Model Compile

• 모델 학습방법 설정
  • 이미 학습된 Weight 값을 Tuning
  • 매우 작은 Learnig Rate 지정
  • optimizers.Adam(lr = 0.000005)

from tensorflow.keras import optimizers

model.compile(loss = 'binary_crossentropy',
              optimizer = optimizers.Adam(lr = 0.000005),
              metrics = ['accuracy'])

---

3-3. Model Fit

%%time

Hist_dandc = model.fit(train_generator,
                       steps_per_epoch = 100,
                       epochs = 100,
                       validation_data = valid_generator,
                       validation_steps = 50)

---

3-4.학습 결과 시각화

#Loss Visualization

import matplotlib.pyplot as plt

epochs = range(1, len(Hist_dandc.history['loss']) + 1)

plt.figure(figsize = (9, 6))
plt.plot(epochs, Hist_dandc.history['loss'])
plt.plot(epochs, Hist_dandc.history['val_loss'])

plt.title('Training & Validation Loss')
plt.xlabel('Epochs')
plt.ylabel('Loss')
plt.legend(['Training Loss', 'Validation Loss'])
plt.grid()
plt.show()

#Accuracy Visualization

import matplotlib.pyplot as plt

epochs = range(1, len(Hist_dandc.history['loss']) + 1)

plt.figure(figsize = (9, 6))
plt.plot(epochs, Hist_dandc.history['accuracy'])
plt.plot(epochs, Hist_dandc.history['val_accuracy'])

plt.title('Training & Validation Accuracy')
plt.xlabel('Epochs')
plt.ylabel('Accuracy')
plt.legend(['Training Accuracy', 'Validation Accuracy'])
plt.grid()
plt.show()

---

3-5. Model Evaluate

test_datagen = ImageDataGenerator(rescale = 1./255)

test_generator = test_datagen.flow_from_directory(
                 test_dir,
                 target_size = (150, 150),
                 batch_size = 20,
                 class_mode = 'binary')

Found 1000 images belonging to 2 classes.

loss, accuracy = model.evaluate(test_generator,
                                steps = 50)

print('Loss = {:.5f}'.format(loss))
print('Accuracy = {:.5f}'.format(accuracy))