데이터분리하여 이미지증강 코딩

Cats-v-dogs dataset 으로, 데이터를 분리하고, 이미지 증강을 활용하요, 인공지능 학습시킨다.

# /tmp 디렉토리 안에다, 학습을 위한 데이터를 분류하기 위해

# cats-v-dogs 디렉토리를 만들고, 

# 그 아래 training 과 testing 디렉토리 만든 후

# 각각 디렉토리 안에  cats 와 dogs 디렉토리를 만든다.

try:

    #YOUR CODE GOES HERE

    os.mkdir('/tmp/cats-v-dogs')

    os.mkdir('/tmp/cats-v-dogs/training')

    os.mkdir('/tmp/cats-v-dogs/testing')

    os.mkdir('/tmp/cats-v-dogs/training/cats')

    os.mkdir('/tmp/cats-v-dogs/training/dogs')

    os.mkdir('/tmp/cats-v-dogs/testing/cats')

    os.mkdir('/tmp/cats-v-dogs/testing/dogs')

##  split_data 라는 파이썬 함수를 만든다.

# SOURCE directory 

# TRAINING directory 

# TESTING directory 

# SPLIT SIZE 

# 소스 경로에 있는 파일들을 잘 섞은 후

# 스플릿 사이즈만큼 학습용과 테스트용 파일로 분리한 후

# 트레이닝 디렉토리와 테스팅 디렉토리에 파일을 저장 시켜 주는 함수 작성 (파일 크기가 0보다 큰 파일만 저장)

# 힌트.

# os.listdir(DIRECTORY) gives you a listing of the contents of that directory

# os.path.getsize(PATH) gives you the size of the file

# copyfile(source, destination) copies a file from source to destination

# random.sample(list, len(list)) shuffles a list

def split_data(SOURCE, TRAINING, TESTING, SPLIT_SIZE):

# YOUR CODE STARTS HERE

    # 1. 파일을 잘 섞어준다. 

    file_names = os.listdir(SOURCE)

    shuffled_names = random.sample(file_names, len(file_names))

    # 2. 트레이닝셋과 테스트셋으로 나눈다. (비율 계산)

    # 나누기 위한 기준이 되는 인덱스를 먼저 계산한다. 

    index = int(len(shuffled_names) * SPLIT_SIZE)

    training_names = shuffled_names[ 0 : index ]

    testing_names = shuffled_names[index : ]

    # 3. 원본 경로에 있는 파일을, 트레이닝과 테스팅 경로로 복사해준다. 

    for file_name in training_names :

        if os.path.getsize(SOURCE + file_name) > 0 : 

            copyfile(SOURCE + file_name, TRAINING + file_name )

    for file_name in testing_names :

        if os.path.getsize(SOURCE + file_name) > 0 : 

            copyfile(SOURCE + file_name, TESTING + file_name )

# YOUR CODE ENDS HERE

CAT_SOURCE_DIR = "/tmp/PetImages/Cat/"

TRAINING_CATS_DIR = "/tmp/cats-v-dogs/training/cats/"

TESTING_CATS_DIR = "/tmp/cats-v-dogs/testing/cats/"

DOG_SOURCE_DIR = "/tmp/PetImages/Dog/"

TRAINING_DOGS_DIR = "/tmp/cats-v-dogs/training/dogs/"

TESTING_DOGS_DIR = "/tmp/cats-v-dogs/testing/dogs/"

split_size = .9

split_data(CAT_SOURCE_DIR, TRAINING_CATS_DIR, TESTING_CATS_DIR, split_size)

split_data(DOG_SOURCE_DIR, TRAINING_DOGS_DIR, TESTING_DOGS_DIR, split_size)

# Expected output

# 666.jpg is zero length, so ignoring

# 11702.jpg is zero length, so ignoring

from tensorflow.keras.layers import Conv2D, MaxPool2D, Flatten, Dense

# 적어도 3개의 CONVOLUTION LAYERS를 포함한 모델을 만드시오

model = tf.keras.models.Sequential([

# YOUR CODE HERE

    Conv2D(32, (3,3), activation='relu', input_shape = (300, 300, 3)), 

    MaxPool2D((2,2), 2), 

    Conv2D(64, (3,3), activation='relu'),

    MaxPool2D((2,2), 2), 

    Conv2D(128, (3,3), activation='relu'),

    MaxPool2D((2,2), 2), 

    Flatten(), 

    Dense(512, activation='relu'),

    Dense(1, activation='sigmoid')

])

model.compile(optimizer=RMSprop(learning_rate=0.001), loss='binary_crossentropy', metrics=['accuracy'])

 

# 이미지 증강도 하여, 코드를 작성하시오.

TRAINING_DIR = '/tmp/cats-v-dogs/training'

train_datagen = ImageDataGenerator(rescale=1/255.0, rotation_range=20, width_shift_range=0.2, height_shift_range=0.2, shear_range=0.2, horizontal_flip=True, 

                                   fill_mode = 'nearest')

train_generator = train_datagen.flow_from_directory(TRAINING_DIR, target_size = (300, 300), class_mode = 'binary')

VALIDATION_DIR = '/tmp/cats-v-dogs/testing'

validation_datagen = ImageDataGenerator(rescale=1/255.0)

validation_generator = validation_datagen.flow_from_directory(VALIDATION_DIR, target_size = (300, 300), class_mode = 'binary')

# Expected Output:

# Found 22498 images belonging to 2 classes.

# Found 2500 images belonging to 2 classes.

Found 22496 images belonging to 2 classes.

Found 2502 images belonging to 2 classes.

 

# 학습합니다.

history = model.fit(train_generator,

                              epochs=10,

                              verbose=1,

                              validation_data=validation_generator)