[52일차] ASAC 빅데이터 분석가 4기(DL-5)

⚙️ VGG Model

📌 데이터 준비

import tensorflow as tf

fahion_mnist = tf.keras.datasets.fashion_mnist
(train_X, train_y), (test_X, test_y) = fahion_mnist.load_data()

train_X = train_X / 255.0
test_X = test_X / 255.0

train_X = train_X.reshape(-1,28,28,1)
test_X  = test_X.reshape(-1,28,28,1)

---

📌 데이터 펌핑하기

import numpy as np

augment_size = 30000
randidx = np.random.randint(train_X.shape[0], size=augment_size)

# 학습 데이터를 랜덤으로 가져와서 펌핑하기

x_augmented = train_X[randidx].copy()
y_augmented = train_y[randidx].copy()

x_augmented = image_gen.flow(x_augmented,
                             np.zeros(augment_size),
                             batch_size = augment_size,
                             shuffle = False).next()[0]

# 펌핑된 데이터를 전체 데이터에 추가
train_X = np.concatenate((train_X, x_augmented))
train_y = np.concatenate((train_y, y_augmented))

---

📌 VGG 모델 만들기

model_vgg_fshion= tf.keras.Sequential([
    
    Conv2D(input_shape = (28,28,1), kernel_size = (3,3), filters = 64, padding = "same", activation = "relu"),
    Conv2D(kernel_size = (3,3), filters = 64, padding = "same", activation = "relu"),
    MaxPool2D(pool_size = (2,2), strides = (2,2)),

    Conv2D(kernel_size = (3,3), filters = 128, padding = "same", activation = "relu"),
    Conv2D(kernel_size = (3,3), filters = 128, padding = "same", activation = "relu"),
    MaxPool2D(pool_size = (2,2), strides = (2,2)),


    Conv2D(kernel_size = (3,3), filters = 256, padding = "same", activation = "relu"),
    Conv2D(kernel_size = (3,3), filters = 256, padding = "same", activation = "relu"),
    Conv2D(kernel_size = (3,3), filters = 256, padding = "same", activation = "relu"),
    MaxPool2D(pool_size = (2,2), strides = (2,2)),

    # 분류를 위한 FNN
    # 특징에 대한 Flatten 작업
    Flatten(),

    # 분류용 HL
    Dense(units = 512, activation = "relu"),
    Dropout(rate = 0.2),
    Dense(units = 256, activation = "relu"),
    Dropout(rate = 0.2),
    # + Layer를 추가를 하거나..

    # 출력용
    Dense(units = 10, activation = "softmax"),

])

# 모델 저장 및 옵션

import os
cp_path = "training/cp-{epoch:04d}.ckpt"
cp_dir  = os.path.dirname(cp_path)
cp_callback = tf.keras.callbacks.ModelCheckpoint(
    cp_path,
    verbose = 1,
    save_weights_only = True
    )
es_callback = tf.keras.callbacks.EarlyStopping(
    monitor = "val_loss",
    patience = 10
)

---

📌 생성해둔 이미지로 학습하기

model_vgg_fshion.compile(
    optimizer = tf.keras.optimizers.Adam(),
    loss = "sparse_categorical_crossentropy",
    metrics = ["accuracy"]
)

model_vgg_fshion.fit(train_X, train_y,
                     epochs = 100,
                     validation_split = 0.25,
                     batch_size = 1024,
                     callbacks = [cp_callback,es_callback]
)

---

📌 이미지를 학습시에 생성해서 사용하기

generator= tf.keras.preprocessing.image.ImageDataGenerator(
    rotation_range = 10,
    zoom_range = 0.10,
    horizontal_flip = True,
    vertical_flip = False
)

train_batches = generator.flow(train_X, train_y, batch_size = 256)
val_batches = generator.flow(test_X, test_y, batch_size = 256)

history = model_vgg_fshion.fit_generator(
    train_batches,
    epochs = 20,
    validation_data = val_batches,
    use_multiprocessing = True
)