tensflow2.0系列-自动训练问题

**第一步：**首先构建神经网络模型，有如下方法：

1.tf.keras.Sequential([ ])方法：

1
def MLP(input_shape):
2
    model = tf.keras.Sequential([
3
        tf.keras.layers.Dense(512, input_shape=input_shape),
4
        tf.keras.layers.BatchNormalization(),
5
        tf.keras.layers.Activation('relu'),
6

7
        tf.keras.layers.Dense(256),
8
        tf.keras.layers.BatchNormalization(),
9
        tf.keras.layers.Activation('relu'),
10

11
        tf.keras.layers.Dense(128),
12
        tf.keras.layers.BatchNormalization(),
13
        tf.keras.layers.Activation('relu'),
14

15
        tf.keras.layers.Dense(2),
16
        tf.keras.layers.BatchNormalization(),
17
        tf.keras.layers.Activation('sigmoid')
18
    ])
19
    return model

tf.keras.Model()方法：

1
def MLP(input_dimension, output_bgs, model_name='Pretrain_MLP_P'):
2
    inputs = tf.keras.Input(shape=(input_dimension,))
3
    x = tf.keras.layers.Dense(1024)(inputs)
4
    x = tf.keras.layers.BatchNormalization()(x)
5
    x = tf.keras.layers.Activation("relu")(x)
6

7
    x = tf.keras.layers.Dense(512)(x)
8
    x = tf.keras.layers.BatchNormalization()(x)
9
    x = tf.keras.layers.Activation("relu")(x)
10

11
    x = tf.keras.layers.Dense(256)(x)
12
    x = tf.keras.layers.BatchNormalization()(x)
13
    x = tf.keras.layers.Activation("relu")(x)
14

15
    x = tf.keras.layers.Dense(256)(x)
16
    x = tf.keras.layers.BatchNormalization()(x)
17
    x = tf.keras.layers.Activation("relu")(x)
18

19
    x = tf.keras.layers.Dense(128)(x)
20
    x = tf.keras.layers.BatchNormalization()(x)
21
    x = tf.keras.layers.Activation("relu")(x)
22

23
    x = tf.keras.layers.Dense(64)(x)
24
    x = tf.keras.layers.BatchNormalization()(x)
25
    x = tf.keras.layers.Activation("relu")(x)
26

27
    x = tf.keras.layers.Dense(output_bgs)(x)
28
    x = tf.keras.layers.BatchNormalization()(x)
29
    outputs = tf.keras.layers.Activation("sigmoid")(x)
30

31
    model = tf.keras.Model(
32
        inputs=inputs,
33
        outputs=outputs,
34
        name=model_name
35
    )
36
    return model

**第二步：**实例化神经网络模型并编译：

1
model = MLP((5,))
2
    model.compile(
3
        optimizer=tf.keras.optimizers.Adam(1e-3),
4
        loss='mean_squared_error'
5
    )

**第三步：**训练：

x与y分别是输入输出数据，validation_data是验证数据该选项可选，epochs是训练次数又叫迭代次数，batch_size是在epochs次数中可以分成n组，每组batch_size个，verbose是进度条，1是显示。

1
history = model.fit(
2
        x=train_pm,
3
        y=train_bg,
4
        validation_data=(confirm_pm, confirm_bg),
5
        epochs=epochs,
6
        batch_size=epochs,
7
        verbose=1
8
    )

**可选：**自定义回调

示例为训练早停，监控'val_loss'，当在达到50 epoch之后且因为mode='min‘ 模式是最小，所以’val_loss' < 1e-6时restore_best_weights=True存储最佳模型参数权重，并停止训练。

1
model.compile(
2
        optimizer=tf.keras.optimizers.Adam(1e-3),
3
        loss='mean_squared_error'
4
    )
5
    early_stopping = tf.keras.callbacks.EarlyStopping(
6
        monitor='val_loss',
7
        patience=50,
8
        min_delta=1e-6,
9
        mode='min',
10
        restore_best_weights=True,
11
        verbose=1
12
    )
13
    model.fit(
14
        x=train_pm,
15
        y=train_bg,
16
        validation_data=[confirm_pm, confirm_bg],
17
        batch_size=batch_size,
18
        epochs=epochs,
19
        callbacks=[early_stopping],
20
        verbose=1
21
    )

ISAJ