Deep Learning 09 - Transfer Learning (전이학습), Pretrained Network

20201104

 

Transfer Learning (전이학습)

학습 데이터가 부족한 분야의 모델 구축을 위해 데이터가 풍부한 분야에서 훈련된 모델을 재사용하는 머신러닝 학습 기법

• Pretrained Network
  • Inception - Google
  • ResNet - MS
  • VGG - 연습
• CNN
  • 기존 Network에서 Convolution Layer - 사용할 데이터 특성
  • -> Feature Map 추출
  • -> Classifier (분류기) (FC Layer)에 연결하여 학습

Convolution Layer	FC Layer
특성 추출 (Feature Extension)	Classifier

• 데이터 증식까지 포함한 model을 합쳐서 구현하고, Convolution Layer를 동결한다.
  • 1) Feature (map) Extraction
    • 기존 가중치는 그대로 나둔 뒤, 새로운 레이어를 추가해서 이를 학습하고 최종 결과를 내게끔 학습
  • 2) Fine Tunung (미세 조정)
    • 새로운 데이터로 다시한번 가중치를 세밀하게 조정하도록 학습. 기존 데이터는 기존대로 분류

 

# 1104_cat_dog_transfer_learning_feature_extraction.ipynb

import tensorflow as tf

gpus = tf.config.experimental.list_physical_devices('GPU')
if gpus:
    # 특정 GPU에 1GB 메모리만 할당하도록 제한
    try:
        tf.config.experimental.set_visible_devices(gpus[1], 'GPU')
        tf.config.experimental.set_virtual_device_configuration(
            gpus[1],
            [tf.config.experimental.VirtualDeviceConfiguration(memory_limit=1024)])
    except RuntimeError as e:
    # 프로그램 시작시에 가상 장치가 설정되어야만 합니다
        print(e)

 

 

from tensorflow.keras.applications import VGG16

# VGG16은 Convolution Layer 16개, Fc Layer는 3개로 구성된 Network

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

base_model.summary()

Output:

Model: "vgg16"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_1 (InputLayer)         [(None, 150, 150, 3)]     0         
_________________________________________________________________
block1_conv1 (Conv2D)        (None, 150, 150, 64)      1792      
_________________________________________________________________
block1_conv2 (Conv2D)        (None, 150, 150, 64)      36928     
_________________________________________________________________
block1_pool (MaxPooling2D)   (None, 75, 75, 64)        0         
_________________________________________________________________
block2_conv1 (Conv2D)        (None, 75, 75, 128)       73856     
_________________________________________________________________
block2_conv2 (Conv2D)        (None, 75, 75, 128)       147584    
_________________________________________________________________
block2_pool (MaxPooling2D)   (None, 37, 37, 128)       0         
_________________________________________________________________
block3_conv1 (Conv2D)        (None, 37, 37, 256)       295168    
_________________________________________________________________
block3_conv2 (Conv2D)        (None, 37, 37, 256)       590080    
_________________________________________________________________
block3_conv3 (Conv2D)        (None, 37, 37, 256)       590080    
_________________________________________________________________
block3_pool (MaxPooling2D)   (None, 18, 18, 256)       0         
_________________________________________________________________
block4_conv1 (Conv2D)        (None, 18, 18, 512)       1180160   
_________________________________________________________________
block4_conv2 (Conv2D)        (None, 18, 18, 512)       2359808   
_________________________________________________________________
block4_conv3 (Conv2D)        (None, 18, 18, 512)       2359808   
_________________________________________________________________
block4_pool (MaxPooling2D)   (None, 9, 9, 512)         0         
_________________________________________________________________
block5_conv1 (Conv2D)        (None, 9, 9, 512)         2359808   
_________________________________________________________________
block5_conv2 (Conv2D)        (None, 9, 9, 512)         2359808   
_________________________________________________________________
block5_conv3 (Conv2D)        (None, 9, 9, 512)         2359808   
_________________________________________________________________
block5_pool (MaxPooling2D)   (None, 4, 4, 512)         0         
=================================================================
Total params: 14,714,688
Trainable params: 14,714,688
Non-trainable params: 0

 

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

base_dir = './data/cat_dog_small'
train_dir = os.path.join(base_dir, 'train')
validation_dir = os.path.join(base_dir, 'validation')
test_dir = os.path.join(base_dir, 'test')

datagen = ImageDataGenerator(rescale=1/255)
batch_size=20

def extract_feature(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 x_data_batch, t_data_batch in generator:
        feature_batch = base_model.predict(x_data_batch)
        features[i*batch_size:(i+1)*batch_size] = feature_batch
        labels[i*batch_size:(i+1)*batch_size] = t_data_batch
        
        i += 1
        if i * batch_size >= sample_count:
            break;
    
    return features, labels

train_features, train_labels = extract_feature(train_dir,2000)
validation_features, validation_labels = extract_feature(validation_dir,1000)
test_features, test_labels = extract_feature(test_dir,1000)

Output:

Found 2000 images belonging to 2 classes.
Found 1000 images belonging to 2 classes.
Found 1000 images belonging to 2 classes.

 

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

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Dropout
from tensorflow.keras.optimizers import RMSprop

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

model.compile(optimizer=RMSprop(learning_rate=2e-5),
              loss='binary_crossentropy',
              metrics=['accuracy'])

history = model.fit(train_features,
                    train_labels,
                    epochs=30,
                    batch_size=20,
                    validation_data=(validation_features, validation_labels)) 

Output:

## 생략
Epoch 27/30
100/100 [==============================] - 0s 5ms/step - loss: 0.1009 - accuracy: 0.9675 - val_loss: 0.2379 - val_accuracy: 0.9050
Epoch 28/30
100/100 [==============================] - 0s 5ms/step - loss: 0.0916 - accuracy: 0.9725 - val_loss: 0.2413 - val_accuracy: 0.9030
Epoch 29/30
100/100 [==============================] - 0s 5ms/step - loss: 0.0852 - accuracy: 0.9760 - val_loss: 0.2472 - val_accuracy: 0.9010
Epoch 30/30
100/100 [==============================] - 0s 5ms/step - loss: 0.0847 - accuracy: 0.9770 - val_loss: 0.2374 - val_accuracy: 0.9030
[ ]
1234567891011121314151617181920

 

%matplotlib inline
import matplotlib.pyplot as plt

train_acc = history.history['accuracy']
val_acc = history.history['val_accuracy']
train_loss = history.history['loss']
val_loss = history.history['val_loss']

plt.plot(train_acc, 'bo', color='r', label='training accuracy')
plt.plot(val_acc, 'b', color='b', label='validation accuracy')
plt.title('Training and Validation Accuracy')
plt.legend()
plt.show()

plt.plot(train_loss, 'bo', color='r', label='training loss')
plt.plot(val_loss, 'b', color='b', label='validation loss')
plt.title('Training and Validation Loss')
plt.legend()
plt.show()

밑에 생략