今天延續之前的主題，我們將使用EfficientNetB0的架構，但不使用預訓練權重，參考了Keras文檔的文章，我們將input_shape設定為EfficientNetB0預設的值224X224，等等會利用Keras提供的函數，幫我們載入並縮放這些圖片，由於驗證集的部分不希望使用資料增強，所以分成了兩個資料夾，使用兩個ImageDataGenerator去處理。

input_shape = (224, 224)
batch_size = 64

from keras.preprocessing.image import ImageDataGenerator

traindatagen = ImageDataGenerator(
    width_shift_range = 0.1,
    height_shift_range = 0.1,
    zoom_range = 0.2,
    shear_range = 0.1,
    rotation_range = 25,
    horizontal_flip = True,
    rescale = 1/255.
)

validdatagen = ImageDataGenerator(
    rescale = 1/255.
)

train = traindatagen.flow_from_directory(
    img_directory + 'train',
    target_size=input_shape,
    color_mode="rgb",
    class_mode="binary",
    batch_size=batch_size,
    shuffle=True,
    interpolation="lanczos",
)

valid = validdatagen.flow_from_directory(
    img_directory + 'validation',
    target_size=input_shape,
    color_mode="rgb",
    class_mode="binary",
    batch_size=batch_size,
    shuffle=True,
    interpolation="lanczos",
)

你應該可以看到它印出了兩行結果，表示完成了對路徑的掃瞄，數量比例也大致相符。接著引入我們的主角。

from tensorflow.keras.applications import EfficientNetB0
from tensorflow.keras.layers import Dense, Dropout
from tensorflow.keras.models import Model

model = EfficientNetB0(include_top=True, weights=None, input_shape=(*input_shape,3), classes=1, activation='sigmoid', pooling='avg')

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

再來設定一些回調函數。

from tensorflow.keras.callbacks import ModelCheckpoint, CSVLogger, TerminateOnNaN, EarlyStopping
mcp = ModelCheckpoint(filepath='EfficientNetB0-{epoch:02d}.h5', monitor='val_loss', verbose=0, save_best_only=True, save_weights_only=False, mode='auto', save_freq='epoch')
log = CSVLogger(filename='EfficientNetB0.csv', separator=',', append=False)
ton = TerminateOnNaN()
esl = EarlyStopping(monitor='val_loss', patience=10, mode='auto', restore_best_weights=True)
esa = EarlyStopping(monitor='val_accuracy', patience=10, mode='auto', restore_best_weights=True)

只要經過漫長的等待就可以收穫勝利的果實了。

hist = model.fit(
    x = train,
    steps_per_epoch = train.samples // batch_size,
    epochs = 50,
    validation_data = valid,
    validation_steps = valid.samples // batch_size,
    callbacks = [mcp, log, ton, esl, esa]
)