tensorflow2中创建单层的tf.keras.layers.Dense的用法

tf.keras.layers.Dense相当于在全连接层中添加一个层

Dense实现以下操作：output = activation（dot（input，kernel）+ bias） 其中，activation是用activation参数传递的逐元素激活函数，kernel是该层创建的权重矩阵，bias是由图层创建的偏差向量（仅在use_bias为True时适用）。

用法：

tf.keras.layers.Dense
(
    units,
    activation=None,
    use_bias=True,
    kernel_initializer='glorot_uniform',
    bias_initializer='zeros',
    kernel_regularizer=None,
    bias_regularizer=None,
    activity_regularizer=None,
    kernel_constraint=None,
    bias_constraint=None,
    **kwargs
)

部分参数说明：

• units：输出的维度大小，改变inputs的最后一维
• activation：激活函数，即神经网络的非线性变化，‘relu’，'selu’等。
• use_bias：使用，bias为True（默认使用），不用，bias改成False即可，是否使用偏置项

示例：

import tensorflow as tf

x = tf.random.normal([4, 784])

net = tf.keras.layers.Dense(512)
out = net(x)

out.shape
Out[9]: TensorShape([4, 512])

net.kernel.shape, net.bias.shape
Out[10]: (TensorShape([784, 512]), TensorShape([512]))

build(input_shape)：定义权重的方法

net = tf.keras.layers.Dense(10)
net.bias
# AttributeError: 'Dense' object has no attribute 'bias'

net.get_weights()  # w
Out[6]: []
net.weights  # b
Out[7]: []

net.build(input_shape = (None, 4))
net.kernel.shape, net.bias.shape
Out[9]: (TensorShape([4, 10]), TensorShape([10]))

net.build(input_shape = (None, 20))
net.kernel.shape, net.bias.shape
Out[12]: (TensorShape([20, 10]), TensorShape([10]))

net.build(input_shape = (2, 4))
net.kernel
Out[14]:
<tf.Variable 'kernel:0' shape=(4, 10) dtype=float32, numpy=
array([[ 0.46936738,  0.61390924, -0.048132  , -0.14290583,  0.06667054,
         0.20691943,  0.45904446, -0.1339497 , -0.04281723, -0.62330157],
       [-0.49840224,  0.05999857, -0.19198364, -0.21388018, -0.5439534 ,
        -0.47478867, -0.6437938 ,  0.17085022,  0.3079493 , -0.5329439 ],
       [-0.63211656, -0.5696165 ,  0.2264719 ,  0.15284914, -0.57099473,
         0.50631917,  0.47583818, -0.16816309, -0.23766968, -0.3798188 ],
       [-0.40847504, -0.18635285, -0.3801926 ,  0.4729545 ,  0.12844968,
        -0.6282965 ,  0.5639292 ,  0.47344756,  0.5877563 , -0.12006855]],
      dtype=float32)>

当input_shape和具体创建时的shape不一致时，会报错

net.build(input_shape = (None, 20))

net.kernel.shape, net.bias.shape
Out[16]: (TensorShape([20, 10]), TensorShape([10]))

# 20 != 12时，会报错
out = net(tf.random.randn((4, 12)))
# AttributeError: module 'tensorflow_core._api.v2.random' has no attribute 'randn'

out = net(tf.random.normal((4, 20)))
out.shape
Out[19]: TensorShape([4, 10])