tf2训练模型的编写方法集合

 

tf.keras.backend.clear_session()
 
model = models.Sequential()
model.add(layers.Dense(64, input_dim=64,
                kernel_regularizer=regularizers.l2(0.01), 
                activity_regularizer=regularizers.l1(0.01),
                kernel_constraint = constraints.MaxNorm(max_value=2, axis=0))) 
model.add(layers.Dense(10,
        kernel_regularizer=regularizers.l1_l2(0.01,0.01),activation = "sigmoid"))
model.compile(optimizer = "rmsprop",
        loss = "sparse_categorical_crossentropy",metrics = ["AUC"])
model.summary()

API法

model=tf.keras.Sequential()
model.add(tf.keras.layers.Dense(10,input_shape=(3,),activation='relu'))
model.add(tf.keras.layers.Dense(6,activation='relu'))

model.add(tf.keras.layers.Dropout(0.2))
model.add(tf.keras.layers.Dense(1,activation='sigmoid') )

model.compile(optimizer=tf.keras.optimizers.Adam(0.01),
                              loss='mse', # 'binary_crossentropy',
                              metrics=['acc'])

model.fit(x,y,epochs=500)

函数法

input=tf.keras.Input(shape=(28,28))
x=tf.keras.layers.Flatten()(input)
x=tf.keras.layers.Dense(32,activation='relu')(x)
x=tf.keras.layers.Dropout(0.5)(x)
x=tf.keras.layers.Dense(64,activation='relu')(x)
output=tf.keras.layers.Dense(10,activation='softmax')(x)

model=tf.keras.Model(input,output,)

model.compile(  optimizer='adam',
              loss='sparse_categorical_crossentropy',   #连续编码用，只有一个序号；如果用                                         #one-hot独热编码(即将一个序号编码成独1多0输出）时删掉 sparse_
              metrics=['acc'] )

分步详解法

w1 = tf.Variable(tf.random.normal([2, 11]), dtype=tf.float32)
b1 = tf.Variable(tf.constant(0.01, shape=[11]))

w2 = tf.Variable(tf.random.normal([11, 1]), dtype=tf.float32)
b2 = tf.Variable(tf.constant(0.01, shape=[1]))

lr = 0.01    # 学习率
epoch = 300  # 循环轮数

# 训练部分
for epoch in range(epoch):
    for step, (x_train, y_train) in enumerate(train_db):
        with tf.GradientTape() as tape:         # 记录梯度信息

            h1 = tf.matmul(x_train, w1) + b1  # 记录神经网络乘加运算
            h1 = tf.nn.relu(h1)
            y = tf.matmul(h1, w2) + b2
         
            # 采用均方误差损失函数mse = mean(sum(y-out)^2)
            loss_mse = tf.reduce_mean(tf.square(y_train - y))
            # 添加l2正则化
            loss_regularization = []
            # tf.nn.l2_loss(w)=sum(w ** 2) / 2
            loss_regularization.append(tf.nn.l2_loss(w1))
            loss_regularization.append(tf.nn.l2_loss(w2))
            # 求和
            # 例：x=tf.constant(([1,1,1],[1,1,1]))
            #   tf.reduce_sum(x)
            # >>>6
            # loss_regularization = tf.reduce_sum(tf.stack(loss_regularization))
            loss_regularization = tf.reduce_sum(loss_regularization)
            loss = loss_mse + 0.03 * loss_regularization #REGULARIZER = 0.03            

        # 计算loss对各个参数的梯度      
        grads = tape.gradient(loss, [w1, b1, w2, b2])

        # 实现梯度更新
        # w1 = w1 - lr * w1_grad tape.gradient是自动求导结果与[w1, b1, w2, b2] 索引为0，1，2，3 
        w1.assign_sub(lr * grads[0])
        b1.assign_sub(lr * grads[1])
        w2.assign_sub(lr * grads[2])
        b2.assign_sub(lr * grads[3])

    # 每20个epoch，打印loss信息
    if epoch % 20 == 0:
        print('epoch:', epoch, 'loss:', float(loss))