[CNN] 개념 및 실습 Tensorflow

# 합성곱 신경망 (CNN)
- 주로 이미지 데이터 학습할 때 사용
- 모든 학습은 오차역전파 알고리즘 -> 경사하강법
  - 가중치가 학습되는 방식
  - Momentum : 경사하강법에 Momentum(탄성)을 더해줌
  - NAG : 직전 모멘텀 항에서 그레디언트를 계산하여 예측
  - AdaGrad : 매 시점마다 별도의 학습률을 사용
  - RMSprop : 최근 값의 영향은 크고, 오래된 값의 영향은 대폭 낮춤
  - Adam : Momentum + RMSpropp ; 가장 좋은 성능, 가장 많이 사용됨

- DNN의 문제를 해결해 준다
- 복잡도 증가 시, 과적합 문제 발생
- 과적합 문제 해결 방법 존재
  - 규제화
  - 드롭아웃
  - 배치 정규화 (미니배치, 공변량 변화)
- 합성곱 층(Layer)
  - 입력층 : 이미지 데이터를 입력 ; Con2D(); 이전 DNN에서는 Flatten()으로 크기를 지정해줬는데, 그럴 필요가 이제 없다
  - 첫번째 합성곱 층 : 모든 픽셀이 아닌, 수용장 안의 픽셀만 연결
  - 두번째 합성곱 층 : 첫번째 층의 (작은/일부) 사각 영역에 위치한 뉴런에 연결
- 합성곱 연산
  - 입력 데이터 중 일부(2*2) 부분이 필터를 통해, 하나의 값으로 출력 ; 패딩은 없던 데이터를 붙이기에, noise 발생
  - 필터 : Kernel
  - output = (FeatureMap - KernelSize + 2*Padding)/(strides) + 1
    - Padding*2인 이유는, 상하좌우
- 폴링 Poling
  - 2차원 데이터의 공간을 줄이는 연산
  - 데이터 차원 감소 ; 계산 효율, 메모리 감소
  - 불변성을 갖기 위해
- 흐름
  - 학습 방식 : Convolution →ReLU  → Pooling → Convolution → ReLU → Pooling → Fully 
Connected
  - Sequential → Conv1D → MaxPooling1D → Flatten → Dense → model.compile() ; Losses, Optimizers → model.fit()
  - 성능 향상 방법 : Activation function뒤에 dropout을 적용 ,Batch normalization
    - Conv Layer → Batch Normalization → Activation function → Pooling Layer
- 일반화
  - 성능을 높이는 것이 목적

---

• fashion mnist 데이터로 CNN 분석

### fashion_mnist 데이터

import tensorflow as tf
from tensorflow import keras
import numpy as np
import matplotlib.pyplot as plt
from tensorflow.keras import datasets, layers, models

# 1. 데이터 불러오기
fashion_mnist = keras.datasets.fashion_mnist
(train_images, train_labels), (test_images, test_labels) = fashion_mnist.load_data()
print(train_images.shape, train_labels.shape)
print(test_images.shape, test_labels.shape)

train_images = train_images.reshape((60000, 28, 28, 1)) # 28*28 2차원 이미지 -> 흑백이미지를 28*
test_images = test_images.reshape((10000, 28, 28, 1))

# 2. 데이터 전처리
train_images = train_images / 255.0
test_images = test_images / 255.0

# 3. 모델 생성(CNN)
model = models.Sequential()
model.add(layers.Conv2D( 
    # filters=16, kernel_size= 3, strides=(1, 1), padding='same', 
    32, (3,3),
    activation='relu', input_shape=(28, 28, 1)))
model.add(layers.MaxPool2D(pool_size=(2, 2))) #, strides=2

model.add(layers.Conv2D(
    # filters=32, kernel_size= 3,strides=(1, 1), padding='same', 
    64,(3,3),
    activation='relu'))
model.add(layers.MaxPool2D(pool_size=(2, 2))) #, strides=2
   
model.add(layers.Conv2D(
    # filters=64, kernel_size= 3, strides=(1, 1), padding='same', 
    64, (3,3), 
     activation='relu'))
# model.add(layers.MaxPool2D(pool_size=(2, 2), strides=2))
   
model.add(layers.Flatten())
model.add(layers.Dense(64, activation = 'relu')) # 3차원을 1차원으로 축소
model.add(layers.Dense(10, activation = 'softmax'))
model.summary()
# 파라미터 수 약 3만7천개
# DNN에서는 9만개정도였는데, 줄었다.

# 4. 모델 컴파일 및 학습
model.compile(optimizer='adam',
        loss = 'categorical_crossentropy', 
        metrics=['accuracy'])
model.fit(train_images, keras.utils.to_categorical(train_labels), epochs=5)

# 5. 모델 평가
test_loss, test_acc = model.evaluate(test_images, keras.utils.to_categorical(test_labels))
print(f"Test Accuracy: {test_acc}")

---

• MASK 데이터로 CNN 분석

### MASK 데이터

import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Conv2D, MaxPooling2D, Flatten, Dense, Dropout
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from sklearn.model_selection import train_test_split

# 이미지 데이터 경로 설정
base_dir = '/content/drive/MyDrive/Colab Notebooks/maskdetection/face_maskdetection'

# 데이터 전처리 및 분할
train_datagen = ImageDataGenerator(
    rescale=1.0 / 255,  # 픽셀 값을 0~1 범위로 조정
    rotation_range=20,  # 이미지 회전
    width_shift_range=0.2,  # 수평 방향 이동
    height_shift_range=0.2, # 수직 방향 이동
    shear_range=0.2,     # 전단 변형
    zoom_range=0.2,      # 확대 축소
    horizontal_flip=True,# 수평 뒤집기
    fill_mode='nearest',
    validation_split=0.2  # 80% 훈련, 20% 검증
)

train_generator = train_datagen.flow_from_directory(
    base_dir,
    target_size=(28, 28),
    batch_size=32,
    class_mode='categorical',
    subset='training'  # 훈련 데이터
)

validation_generator = train_datagen.flow_from_directory(
    base_dir,
    target_size=(28, 28),
    batch_size=32,
    class_mode='categorical',
    subset='validation'  # 검증 데이터
)

# 데이터 확인
print(f"Classes found: {train_generator.class_indices}")  # Classes found: {'with_mask': 0, 'without_mask': 1}
print(f"Training samples: {train_generator.samples}") # Training samples: 1101
print(f"Validation samples: {validation_generator.samples}") # Validation samples: 275
print(train_generator.filenames[:5])  # 훈련 데이터 파일 이름 ; ['with_mask/24-with-mask.jpg', 'with_mask/240-with-mask.jpg', 'with_mask/241-with-mask.jpg', 'with_mask/242-with-mask.jpg', 'with_mask/243-with-mask.jpg']
print(validation_generator.filenames[:5])  # 검증 데이터 파일 이름 ; ['with_mask/0-with-mask.jpg', 'with_mask/1-with-mask.jpg', 'with_mask/10-with-mask.jpg', 'with_mask/100-with-mask.jpg', 'with_mask/101-with-mask.jpg']

### CNN ; model
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Conv2D, MaxPooling2D, Flatten, Dense

model = Sequential([
    Conv2D(32, (3, 3), activation='relu', input_shape=(28, 28, 3)),
    MaxPooling2D(2, 2),
    Dropout(0.2),
    Conv2D(64, (3, 3), activation='relu'),
    MaxPooling2D(2, 2),
    Dropout(0.2),
    Conv2D(32, (3, 3), activation='relu'),
    MaxPooling2D(2, 2),
    Dropout(0.3),
    Flatten(),
    Dense(128, activation='relu'),
    Dense(2, activation='softmax')  # 클래스 수에 따라 변경 (2개 클래스: with_mask, without_mask)
])

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

epochs = 20

# 모델 학습
history = model.fit(
    train_generator,
    validation_data=validation_generator,
    epochs=20,
)

# 모델 평가
test_loss, test_acc = model.evaluate(validation_generator)
print(f"Validation Loss: {test_loss}")
print(f"Validation Accuracy: {test_acc}")

import matplotlib.pyplot as plt

# 정확도 그래프
plt.figure(figsize=(10, 5))
plt.plot(history.history['accuracy'], label='Training Accuracy')
plt.plot(history.history['val_accuracy'], label='Validation Accuracy')
plt.title('Training and Validation Accuracy')
plt.xlabel('Epoch')
plt.ylabel('Accuracy')
plt.legend()

# 손실 그래프
plt.figure(figsize=(10, 5))
plt.plot(history.history['loss'], label='Training Loss')
plt.plot(history.history['val_loss'], label='Validation Loss')
plt.title('Training and Validation Loss')
plt.xlabel('Epoch')
plt.ylabel('Loss')
plt.legend()
plt.show()

---

+ 추가로, Epoch 50번 실행 시, 약간 더 좋은 성능을 보인다.

Validation Loss: 0.11369598656892776

Validation Accuracy: 0.9672726988792419

---

• Shape 데이터로 CNN 분

### Shape 데이터

import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Conv2D, MaxPooling2D, Flatten, Dense, Dropout
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from sklearn.model_selection import train_test_split

# 이미지 데이터 경로 설정
base_dir = '/content/drive/MyDrive/Colab Notebooks/shapes'

# 데이터 전처리 및 분할
train_datagen = ImageDataGenerator(
    rescale=1.0 / 255,
    validation_split=0.2  # 80% 훈련, 20% 검증
)

train_generator = train_datagen.flow_from_directory(
    base_dir,
    target_size=(28, 28),
    batch_size=32,
    class_mode='categorical',
    subset='training'  # 훈련 데이터
)

validation_generator = train_datagen.flow_from_directory(
    base_dir,
    target_size=(28, 28),
    batch_size=32,
    class_mode='categorical',
    subset='validation'  # 검증 데이터
)

# 데이터 확인
print(f"Classes found: {train_generator.class_indices}")  # {'circles': 0, 'squares': 1, 'triangles': 2}
print(f"Training samples: {train_generator.samples}") # Training samples: 240
print(f"Validation samples: {validation_generator.samples}") # Validation samples: 60


# # 이미지 데이터 전처리
# datagen = ImageDataGenerator(
#     rescale=1./255,     # 픽셀 값을 0~1 범위로 조정
#     rotation_range=20,  # 이미지 회전
#     width_shift_range=0.2,  # 수평 방향 이동
#     height_shift_range=0.2, # 수직 방향 이동
#     shear_range=0.2,     # 전단 변형
#     zoom_range=0.2,      # 확대 축소
#     horizontal_flip=True,# 수평 뒤집기
#     fill_mode='nearest'
# )

### CNN ; model
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Conv2D, MaxPooling2D, Flatten, Dense

model = Sequential([
    Conv2D(32, (3, 3), activation='relu', input_shape=(28, 28, 3)),
    MaxPooling2D(2, 2),
    Dropout(0.2),
    Conv2D(64, (3, 3), activation='relu'),
    MaxPooling2D(2, 2),
    Dropout(0.2),
    Conv2D(128, (3, 3), activation='relu'),
    MaxPooling2D(2, 2),
    Dropout(0.3),
    Flatten(),
    Dense(128, activation='relu'),
    Dense(3, activation='softmax')  # 클래스 수에 따라 변경 (3개 클래스: circles, squares, triangles)
])

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

epochs = 20

# 모델 학습
# history = model.fit(
#     X_train, y_train,
#     validation_data=(X_val, y_val),
#     epochs=epochs,
#     batch_size=32
# )


# 모델 학습
history = model.fit(
    train_generator,
    validation_data=validation_generator,
    epochs=20,
)

# 모델 평가
# test_loss, test_acc = model.evaluate(X_val, y_val, batch_size=32)
# print(f'\nValidation accuracy: {test_acc}')

# 모델 평가
test_loss, test_acc = model.evaluate(validation_generator)
print(f"Validation Loss: {test_loss}")
print(f"Validation Accuracy: {test_acc}")

import matplotlib.pyplot as plt

# 정확도 그래프
plt.figure(figsize=(10, 5))
plt.plot(history.history['accuracy'], label='Training Accuracy')
plt.plot(history.history['val_accuracy'], label='Validation Accuracy')
plt.title('Training and Validation Accuracy')
plt.xlabel('Epoch')
plt.ylabel('Accuracy')
plt.legend()

# 손실 그래프
plt.figure(figsize=(10, 5))
plt.plot(history.history['loss'], label='Training Loss')
plt.plot(history.history['val_loss'], label='Validation Loss')
plt.title('Training and Validation Loss')
plt.xlabel('Epoch')
plt.ylabel('Loss')
plt.legend()
plt.show()

성능이 별로 좋지는 않아 보인다... 약 76%