keras自定义层

tensorflow 2.x

import tensorflow as tf
from tensorflow import keras
print(tf.__version__)
print(keras.__version__)
## tf2.xxx 

class Linear(keras.layers.Layer):
    def __init__(self, units=32, input_dim=32):
        super(Linear, self).__init__()
        w_init = tf.random_normal_initializer()
        self.w = tf.Variable(
            initial_value=w_init(shape=(input_dim, units), dtype="float32"),
            trainable=True,
        )
        b_init = tf.zeros_initializer() ## 初始化
        self.b = tf.Variable(
            initial_value=b_init(shape=(units,), dtype="float32"), trainable=True
        )

    def call(self, inputs):
        return tf.matmul(inputs, self.w) + self.b # 这个地方相当于pytorch层的forward
#2.6.2
#2.6.0

############### 自定义层实现###############
#print(linear_layer.weights[0].numpy()) # 访问权重
#print(linear_layer.weights[1].numpy()) # 访问权重

import numpy as np
a=np.random.randn(10,2) 
x=tf.convert_to_tensor(a,dtype=tf.float64)# numpy转tensor
linear_layer = Linear(4, 2)# 第二个参数是输入层的维度，第一个维度是输出层的维度
y = linear_layer(x)
print(y)
print("================================================================================================")
print(np.matmul(a,linear_layer.weights[0].numpy())+linear_layer.weights[1].numpy())

tensorflow 1.xx

import tensorflow as tf
import keras.backend as K
import keras 
print(tf.__version__)
print(keras.__version__)

class Linear(keras.layers.Layer):
    def __init__(self, units=32, input_dim=32):
        super(Linear, self).__init__()
        w_init = tf.random_normal_initializer()
        self.w = tf.Variable(
            initial_value=w_init(shape=(input_dim, units), dtype="float32"),
            trainable=True,
        )
        b_init = tf.zeros_initializer()
        self.b = tf.Variable(
            initial_value=b_init(shape=(units,), dtype="float32"), trainable=True
        )

    def call(self, inputs):
        return tf.matmul(inputs, self.w) + self.b
    
#1.14.0
#2.2.0

import numpy as np 
a=np.random.randn(2,2)
x=tf.convert_to_tensor(a,dtype=tf.float32)
linear_layer = Linear(4, 2)
y = linear_layer(x)
print(K.eval(y))
print("================================================================================================")
print(np.matmul(a,K.eval(linear_layer.w))+K.eval(linear_layer.b))