pretrained network, Transfer Learning (전이학습)

이전에 작업한 "MNIST"는 gray-scale 이었어요 
→ 속도도 빠르고 단순한 image였기 때문에 (convolution layer3개)+(pooling layer)+(FC layer) 
= 높은정확도(99%) = GPU가 아닌 CPU (1시간이내)

하지만, 실무에서 사용하는 데이터(image)는 
(1) size가 커요 + color = 1개 image의 용량이 크고 학습도 잘 안되요
(2) convolution Layer, pooling layer 가 많이필요해서 시간↑
(3) cpu로는 학습이 안되요! → gpu를 사용 (그래도 오래걸려요)
→ hyperparameter(내가 조절하는 작업들)을 하기 위해 상당히 오랜시간이 걸릴수 밖에 없어요!!
→ pretrained network을 이용하면 좋아요!!
   (기존에 이미 학습이 끝나서 완성된 Model [Nevral Network]을 이용) → Transfer Learning (전이학습)

고해상도 컬러이미지에대해 높은 성능을 나타내는 pretrained Network이 존재
연습용 : VGG16, VGG19            # 숫자는 Layers의 수
상용적 : ResNet (MS), Inception (Google) → Converution Layer, FC Layerd의 합 : 각각 150계층, 30계층 
연구용 : efficientNet , ECT
= 위 모두 Reference(참고) [각각 사용하는 방법]가 존재

"ImageNet" [대용량 이미지 ,1000개 class]
input → Pretrained Network → output
1000개의 class에 대한 확률

- Pretrained Network
(1) Feature Extraction
(2) Classifier
ImageNet이 사용하는 classifier는 못써서 Feature Extraction만 사용해서 특성만 뽑는다

 

model = VGG(include_top=False) 를 줄수 있는데

여기서 top은 classfier

         bottom 은 Feature Extraction

---

# Pretrained Network을 가져와서 확인해보아요!

# Pretrained Network의 종류는 상당히 많아요!

# (VGG16, VGG19, ResNet, Inception, EfficientNet, MobileNet, 등등등)

# tensorflow.keras안에서 제공되는 VGG16을 이용해 보아요!

 

from tensorflow.keras.applications import VGG16

 

model_base = VGG16(weights='imagenet',        # 어떤 data를 기반으로 만들었냐

                   include_top=False,                    # top은 classfier

                   input_shape=(150,150,3))           # 들어갈때의 크기

 

# model_base = VGG16(weights='imagenet',

#                    include_top=True)

# # (224, 224, 3)

 

print(model_base.summary())

 

# 이렇게 Pretrained Network을 가져다 사용할 수 있어요!

 

# 이제 두가지 선택지가 존재!

 

# 1. 우리가 가지고 있는 이미지(개와 고양이-4000개)를 Convolution layer(VGG16)에 통과

#    특성을 추출한 이미지(Feature Map)가 나와요!  => (None, 4, 4, 512) 

#    이 결과를 가지고 FC Layer를 이용해서 학습시키는거예요!

#    장점 : 빨라요!(filter를 update할 필요가 없어요!, classifier(DNN)만 학습하면 되요! )

 

# 2. VGG16에 Dense Layer를 붙여서 전체 network을 확장시켜서 사용

#    학습하기 위해 우리의 데이터가 항상 convolution layer를 통과해요 => 속도가 느려져요!

 

---

# 우리가 가지고 있는 이미지(개와 고양이-4000개)를 Convolution layer(VGG16)에 통과

# 특성을 추출한 이미지(Feature Map)가 나와요!  => (None, 4, 4, 512) 

 

import numpy as np

from tensorflow.keras.preprocessing.image import ImageDataGenerator

from tensorflow.keras.applications import VGG16

 

model_base = VGG16(weights='imagenet',

                   include_top=False,

                   input_shape=(150,150,3))

 

train_dir = '/content/drive/MyDrive/융복합 프로젝트형 AI 서비스 개발(2021.06)/09월/30일(목요일)/cat_dog_small/train'

valid_dir = '/content/drive/MyDrive/융복합 프로젝트형 AI 서비스 개발(2021.06)/09월/30일(목요일)/cat_dog_small/validation'

test_dir = '/content/drive/MyDrive/융복합 프로젝트형 AI 서비스 개발(2021.06)/09월/30일(목요일)/cat_dog_small/test'

 

datagen = ImageDataGenerator(rescale=1/255)

batch_size = 20

 

# directory : ImageDataGenerator로 Image를 가져올 폴더(디렉토리)

# sample_count : directory안에 있는 Image의 개수

def extract_feature(directory, sample_count):

    # 이 함수의 결과는..각 이미지의 특성을 뽑아낸 Feature Map이 나와야 해요!

    # 해당 이미지의 label도 결과로 나와야 해요!

    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),

        classes=['cats', 'dogs'],

        batch_size=batch_size,

        class_mode='binary'

    )

 

    i = 0

 

    # x_data : 20장의 이미지 픽셀 데이터, t_data : 20장의 레이블

    for x_data, t_data in generator:

        feature_batch = model_base.predict(x_data)

 

        features[i*batch_size:(i+1)*batch_size] = feature_batch

        labels[i*batch_size:(i+1)*batch_size] = t_data

 

        i += 1

        if i * batch_size >= sample_count:

            break

 

    return features, labels

 

train_feature, train_label = extract_feature(train_dir,2000)

valid_feature, valid_label = extract_feature(valid_dir,1000)

 

---

# 위의 결과로 feature map과 label에 대한 이미지 특성 데이터를 확보

 

# 이렇게 얻은 데이터로 DNN 해야 해요!

 

from tensorflow.keras.models import Sequential

from tensorflow.keras.layers import Dense, Dropout

from tensorflow.keras.optimizers import Adam

 

model = Sequential()

 

model.add(Dense(units=256,

                activation='relu',

                input_shape=(4*4*512,)))

 

model.add(Dropout(rate=0.5))

 

model.add(Dense(units=1,

                activation='sigmoid'))

 

model.compile(optimizer=Adam(learning_rate=1e-5),

              loss='binary_crossentropy',

              metrics=['accuracy'])

 

train_feature_result = np.reshape(train_feature,(2000,4*4*512))

valid_feature_result = np.reshape(valid_feature,(1000,4*4*512))

 

histroy = model.fit(train_feature_result,

                    train_label,

                    epochs=30,

                    batch_size=20,

                    validation_data=(valid_feature_result, valid_label),

                    verbose=1)

---

# accuracy와 loss 비교

import matplotlib.pyplot as plt

 

train_acc = histroy.history['accuracy']

valid_acc = histroy.history['val_accuracy']

 

train_loss = histroy.history['loss']

valid_loss = histroy.history['val_loss']

 

fig = plt.figure()

fig_1 = fig.add_subplot(1,2,1)

fig_2 = fig.add_subplot(1,2,2)

 

fig_1.plot(train_acc, color='b', label='training_accuracy')

fig_1.plot(valid_acc, color='r', label='validation_accuracy')

fig_1.legend()

 

fig_2.plot(train_loss, color='b', label='training_loss')

fig_2.plot(valid_loss, color='r', label='validation_loss')

fig_2.legend()

 

plt.tight_layout()

plt.show()