該程式演示將一個預訓練好的模型在新資料集上重新fine-tuning的過程。我們凍結卷積層，只調整全連線層。

1. 在MNIST資料集上使用前五個數字[0…4]訓練一個卷積網路。
2. 在後五個數字[5…9]用卷積網路做分類，凍結卷積層並且微調全連線層

一、變數初始化

now = datetime.datetime.now

batch_size = 128
nb_classes = 5
nb_epoch = 5

# 輸入影象的維度
img_rows, img_cols = 28, 28
# 使用卷積濾波器的數量
nb_filters = 32
# 用於max pooling的pooling面積的大小
 

pool_size = 2
# 卷積核的尺度
kernel_size = (3,3)
input_shape = (img_rows, img_cols, 1)

# 資料，在訓練和測試資料集上混洗和拆分
(X_train, y_train), (X_test, y_test) = mnist.load_data()
X_train_lt5 = X_train[y_train < 5]
y_train_lt5 = y_train[y_train < 5]
X_test_lt5 = X_test[y_test < 5]
y_test_lt5 = y_test[y_test < 5]

X_train_gte5
 
 = X_train[y_train >= 5]
#使標籤從0~4，故-5
y_train_gte5 = y_train[y_train >= 5] - 5
X_test_gte5 = X_test[y_test >= 5]
y_test_gte5 = y_test[y_test >= 5] - 5

二、模型的訓練函式

def train_model(model, train, test, nb_classes):
    #train[0]是圖片，train[1]是標籤
    X_train = train[0].reshape((train[0].shape[0
 
],) + input_shape)#1D+3D=4D
    X_test = test[0].reshape((test[0].shape[0],) + input_shape)
    X_train = X_train.astype('float32')
    X_test = X_test.astype('float32')
    X_train /= 255
    X_test /= 255
    print('X_train shape:', X_train.shape)
    print(X_train.shape[0], 'train samples')
    print(X_test.shape[0], 'test samples')
    Y_train = np_utils.to_categorical(train[1], nb_classes)
    Y_test = np_utils.to_categorical(test[1], nb_classes)
    model.compile(loss='categorical_crossentropy',
                  optimizer='adadelta',
                  metrics=['accuracy'])
    t = now()
    model.fit(X_train, Y_train,
              batch_size=batch_size, nb_epoch=nb_epoch,
              verbose=1,
              validation_data=(X_test, Y_test))
    print('Training time: %s' % (now() - t))
    score = model.evaluate(X_test, Y_test, verbose=0)
    print('Test score:', score[0])
    print('Test accuracy:', score[1])

三、建立模型，構建卷積層（特徵層）和全連線層（分類層）

feature_layers = [
    Convolution2D(nb_filters, kernel_size,
                  padding='valid',
                  input_shape=input_shape),
    Activation('relu'),
    Convolution2D(nb_filters, kernel_size),
    Activation('relu'),
    MaxPooling2D(pool_size=(pool_size, pool_size)),
    Dropout(0.25),
    Flatten(),
]
classification_layers = [
    Dense(128),
    Activation('relu'),
    Dropout(0.5),
    Dense(nb_classes),
    Activation('softmax')
]
model = Sequential(feature_layers + classification_layers)

四、對模型進行預訓練

train_model(model,
            (X_train_lt5, y_train_lt5),
            (X_test_lt5, y_test_lt5), nb_classes)

五、凍結預訓練模型的特徵層

for l in feature_layers:
    l.trainable = False

六、fine_tuning分類層

train_model(model,
            (X_train_gte5, y_train_gte5),
            (X_test_gte5, y_test_gte5), nb_classes)

原始碼地址：