tensorflow自定义网络结构

class MyDense(tf.keras.layers.Layer):
    def __init__(self,units): # units 神经元个数
        super().__init__() # 必须写
        self.units = units
    def build(self,input_shape):
        self.w = self.add_variable(
            name="w",
            shape=[input_shape[-1],self.units],
            initializer = tf.initializers.RandomNormal()
        )
        self.b = self.add_variable(name="b",shape=[self.units],initializer = tf.initializers.Zeros()) # b一般是全0
    def call(self,input):
        # wx+b
        return input @ self.w + self.b
        return tf.nn.relu(input @ self.w + self.b)

class MyModel(tf.keras.Model):
    def __init__(self):
        super().__init__()
        self.fc1 = MyDense(512)
        self.fc2 = MyDense(256)
        self.fc3 = MyDense(128)
        self.fc4 = MyDense(10)
    def call(self,input):
        self.fc1.out = self.fc1(input)
        self.fc2.out = self.fc2(self.fc1.out)
        self.fc3.out = self.fc3(self.fc2.out)
        self.fc4.out = self.fc4(self.fc3.out)
        return self.fc4.out
myModel = MyModel()
myModel.build(input_shape=(None,784))
myModel.summary()

# 模型保存
# 1，保存模型
# model.save("xxx.h5")
# tensorflow.keras.models.load_model("xxxx.h5")
​
# 2，保存权重参数
# model.save_weights("xxxx.ckpt")
# model.load_weights("xxxx.ckpt")
​
# 3，save_model 此时保存的模型具有平台无关性,移植性好 1.15及之后版本
# tensorflow.keras.models.save_model(model,"foldername")  生成文件夹，里面有pb文件
# tensorflow.keras.models.load_model(“foldername”) 
#   此时只导入的只有model结构与weight参数  model.compile还需要自己写