tensorflow自定义网络层和结构

1. 自定义网络层

实现自己的层的最佳方法是扩展tf.keras.layers.Layer类并实现：

• _init_()函数，在其中执行所有与输入无关的初始化
• build()函数，获得输入张量的形状，并可以进行其余的初始化
• call()函数，构建网络结构，进行前向传播

该自定以网络相当于：tf.keras.layers.Dense()

格式：

class MyDense(tf.keras.layers.Layer):  # 继承该类
	def __init__(self, input_dim, output_dim):
        super(MyDense, self).__init__  # 继承父类中__init__()函数中的成员，否则会被覆盖掉
        # self.weight 和 self.add_variable作用一样，都是把变量添加到层中
        self.kernel = self.add_weight('w', [input_dim, output_dim])
        self.bias = self.add_weight('b', [output_dim])
        
	def call(self, inputs, training=None):  # model(x)会自动调用
        out = inputs @ self.kernel + self.bias
        
        retuen out

2. 自定义网络结构

机器学习模型中有很多是通过叠加不同的结构层组合而成的，实现自定义网络结构就要扩展tf.keras.Model 类

• _init_()函数，在其中执行所有与输入无关的初始化
• build()函数，获得输入张量的形状，并可以进行其余的初始化
• call()函数，构建网络结构，进行前向传播

该自定义网络相当于：tf.keras.Sequential()

格式：

class MyNetwork(tf.keras.Model):
    def __init__(self):
        super(MyNetwork, self).__init__()
        
        self.fc1 = MyDense(32*32*3, 256)  # 调用MyDense类的__init__方法
        self.fc2 = MyDense(256, 128)
        self.fc3 = MyDense(128, 64)
        self.fc4 = MyDense(64, 32)
        self.fc5 = MyDense(32, 10)
	def call(self):
        x = tf.reshape(inputs, [-1, 32*32*3])
        x = self.fc1(x)  # 调用MyDense类的call方法
        x = tf.nn.relu(x)
        x = self.fc2(x)  # 调用MyDense类的call方法
        x = tf.nn.relu(x)
        x = self.fc3(x)  # 调用MyDense类的call方法
        x = tf.nn.relu(x)
        x = self.fc4(x)  # 调用MyDense类的call方法
        x = tf.nn.relu(x)
        x = self.fc5(x)  # 调用MyDense类的call方法
        
        return x

3. 示例

import tensorflow as tf

(x, y), (x_val, y_val) = tf.keras.datasets.mnist.load_data()
print(x.shape, y.shape)

train_db = tf.data.Dataset.from_tensor_slices((x, y))
test_db = tf.data.Dataset.from_tensor_slices((x_val, y_val))
def processing(x, y):
    x = tf.cast(x, dtype=tf.float32)/255.
    x = tf.reshape(x, [-1, 28*28])
    x = tf.squeeze(x)
    y = tf.cast(y, dtype=tf.int32)
    y = tf.one_hot(y, depth=10)
    return x, y


batcs = 128
train_db = train_db.map(processing).shuffle(10000).batch(batcs)
test_db = test_db.map(processing).batch(batcs)

sample = next(iter(train_db))
print(sample[0].shape, sample[1].shape)

# network = tf.keras.Sequential(
#     [tf.keras.layers.Dense(256, activation='relu'),
#     tf.keras.layers.Dense(128, activation='relu'),
#     tf.keras.layers.Dense(64, activation='relu'),
#     tf.keras.layers.Dense(32, activation='relu'),
#     tf.keras.layers.Dense(10)]
# )

class MyDense(tf.keras.layers.Layer):
    def __init__(self, input_dim, output_dim):
        super(MyDense, self).__init__()

        self.kernel = self.add_weight('w', [input_dim, output_dim])
        self.bias = self.add_weight('b', [output_dim])

    def call(self, inputs):
        out = inputs @ self.kernel + self.bias
        return out

class MyNetwork(tf.keras.Model):
    def __init__(self):
        super(MyNetwork, self).__init__()
        self.fc1 = MyDense(28*28, 256)
        self.fc2 = MyDense(256, 128)
        self.fc3 = MyDense(128, 64)
        self.fc4 = MyDense(64, 32)
        self.fc5 = MyDense(32, 10)
    def call(self, inputs):
        x = self.fc1(inputs)
        x = tf.nn.relu(x)
        x = self.fc2(x)
        x = tf.nn.relu(x)
        x = self.fc3(x)
        x = tf.nn.relu(x)
        x = self.fc4(x)
        x = tf.nn.relu(x)
        x = self.fc5(x)

        return x

network = MyNetwork() # 创建该类
network.compile(optimizer=tf.keras.optimizers.Adam(lr=0.001), loss=tf.losses.CategoricalCrossentropy(from_logits=True), metrics=['accuracy']) 
network.fit(train_db, epochs=2)

输出为：

(128, 784) (128, 10)
Epoch 1/2
2020-09-07 10:17:30.047658: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library cublas64_10.dll
469/469 [==============================] - 3s 6ms/step - loss: 0.3062 - accuracy: 0.9102
Epoch 2/2
469/469 [==============================] - 3s 6ms/step - loss: 0.1062 - accuracy: 0.9681