（三）深度学习计算 -- 2
程序员文章站
2022-07-12 23:02:49
...
2.模型参数的访问、初始化和共享
import tensorflow as tf
2.1 access parameters
先定义一个含单隐藏层的多层感知机，用默认方式进行参数初始化，并做一次前向计算：
# 构造
net = tf.keras.models.Sequential()
net.add(tf.keras.layers.Flatten())
net.add(tf.keras.layers.Dense(units=256,activation=tf.nn.relu))
net.add(tf.keras.layers.Dense(units=10))
X = tf.random.normal(shape=(2,20))
Y = net(X)
Y
输出：
<tf.Tensor: id=131, shape=(2, 10), dtype=float32, numpy=
array([[ 0.40529358, 0.15783574, -0.11411244, -0.02750503, 0.7654685 ,
0.30040392, 0.5138651 , -0.175756 , 0.26451442, 0.06294996],
[ 0.13066906, -0.06547291, -0.11907507, -0.07975144, 0.18626948,
-0.35689768, 0.1369905 , -0.20330222, 0.43301803, 0.11998984]],
dtype=float32)>
对于使用Sequential类构造的神经网络，可以通过weights属性访问网络任一层的权重。
访问多层感知机net中隐藏层的所有参数：
net.weights[0], type(net.weights[0])
输出：
(<tf.Variable 'sequential_4/dense_6/kernel:0' shape=(20, 256) dtype=float32, numpy=
array([[ 0.02447869, 0.08890556, 0.07769121, ..., 0.12422395,
-0.08612756, -0.06894704],
[-0.13559662, 0.02131207, 0.11932039, ..., -0.09147681,
0.06214213, 0.00162295],
[ 0.11250323, 0.14524457, 0.07378155, ..., 0.10643828,
0.0880693 , 0.11410677],
...,
[ 0.11822987, -0.01228143, -0.01754822, ..., 0.00051971,
-0.0498704 , 0.03023405],
[ 0.02840948, -0.12175278, -0.10541758, ..., -0.09337974,
0.08511312, 0.0125183 ],
[ 0.08039543, 0.11873844, 0.04696544, ..., -0.1230341 ,
0.10841358, 0.10288766]], dtype=float32)>,
tensorflow.python.ops.resource_variable_ops.ResourceVariable)
其中，索引0表示隐藏层的参数。
2.2 initialize params
将权重参数初始化成均值为0、标准差为0.01的正态分布随机数，并将偏差参数清零：
class Linear(tf.keras.Model):
def __init__(self):
super().__init__()
self.d1 = tf.keras.layers.Dense(
units=10,
activation=None,
kernel_initializer=tf.random_normal_initializer(mean=0,stddev=0.01),
bias_initializer=tf.zeros_initializer()
)
self.d2 = tf.keras.layers.Dense(
units=1,
activation=None,
kernel_initializer=tf.ones_initializer(),
bias_initializer=tf.ones_initializer()
)
def call(self, input):
output = self.d1(input)
output = self.d2(output)
return output
X = tf.random.normal(shape=(2,20))
net = Linear()
Y = net(X)
Y
输出：
<tf.Tensor: id=53, shape=(2, 1), dtype=float32, numpy=
array([[0.8000398],
[0.8991333]], dtype=float32)>
net.get_weights()
输出：
[array([[ 1.08196139e-02, 5.79477428e-03, -1.06005976e-02,
8.26828275e-03, 5.04667917e-03, 1.72793642e-02,
-3.92814586e-03, -2.72732018e-03, 5.62155573e-03,
-3.02137225e-04],
[ 3.87676549e-03, 1.17851729e-02, 1.08660981e-02,
-6.99408259e-03, 1.92590989e-03, -1.83490932e-03,
6.89283572e-03, -7.06359418e-03, 3.14420962e-04,
6.61447644e-04],
[ 8.02275259e-03, 1.57884490e-02, 2.26609013e-03,
-2.23289663e-03, 9.47015267e-03, -7.17470248e-04,
1.16258478e-02, 8.79523432e-05, 3.11152334e-03,
-1.19256079e-02],
[ 6.94830623e-03, 1.22911343e-03, 2.33568228e-03,
1.12228524e-02, 2.49789469e-02, -1.10885501e-02,
-1.73985644e-03, 5.82345063e-03, 1.37769533e-02,
2.61171046e-03],
[-8.19462445e-03, -4.04411508e-03, 2.38322895e-02,
-4.24177560e-04, 1.55501980e-02, 2.83632631e-04,
1.01489350e-02, 1.03493668e-02, 3.76219791e-03,
7.25238118e-03],
[ 7.48372963e-03, 1.34342210e-02, -1.16917929e-02,
-7.98194390e-03, 3.77418124e-03, 5.94647368e-04,
-5.43513242e-03, -2.17209738e-02, 2.01016515e-02,
8.55223276e-04],
[ 6.30497979e-03, -1.42902685e-02, -5.65174408e-03,
1.05804554e-03, -1.00298449e-02, -1.40983220e-02,
9.58261453e-03, 1.90745108e-02, -1.15314289e-03,
-1.60493851e-02],
[ 1.65592954e-02, -6.49799331e-05, 7.29267858e-03,
-1.64555088e-02, 7.86533114e-04, 9.04201809e-03,
9.76761524e-03, -8.34710896e-03, -2.54614046e-03,
-1.39755512e-05],
[ 4.41787904e-03, -1.37950480e-02, 3.53181711e-03,
1.53799134e-03, -1.65310241e-02, 8.04246869e-03,
-1.56394043e-03, 4.09922330e-03, -6.75616204e-04,
3.64435394e-03],
[ 2.04873364e-03, -5.06615173e-03, -1.34313554e-02,
1.44438725e-02, -2.43533379e-03, -2.76915147e-03,
8.29426106e-03, 2.37593781e-02, -1.67158642e-03,
-9.71452612e-03],
[-6.51521317e-04, -2.00210139e-02, 4.81219077e-03,
1.39641659e-02, -9.27464478e-03, 3.03924573e-03,
8.56531877e-03, 7.31856516e-03, -1.51513061e-02,
1.61066949e-02],
[ 2.83615035e-03, 8.50357523e-04, 1.51638659e-02,
9.69374366e-03, -8.76047241e-04, 8.99826828e-03,
-5.30649349e-03, -4.62086685e-03, 1.37379104e-02,
-1.76256541e-02],
[-1.42976148e-02, 9.24051553e-03, 3.62900808e-03,
3.00289178e-03, 5.01110684e-03, 1.03468634e-03,
1.01388549e-03, 9.39975400e-03, -6.78061135e-03,
-1.49735911e-02],
[ 8.51881705e-05, -4.43281513e-03, 3.90838506e-03,
-2.92743999e-03, 2.06509116e-03, -6.18821578e-05,
8.80562980e-03, 2.54227780e-03, 2.56423809e-05,
-4.66403132e-03],
[ 8.15627631e-03, 7.13415863e-03, -2.01360248e-02,
4.30102786e-03, -4.99079004e-04, 5.28379949e-03,
1.58037301e-02, -3.75230238e-03, -8.95734876e-04,
1.39328260e-02],
[-2.48550382e-02, -6.31160010e-03, -6.22724183e-03,
-9.44277551e-03, 1.10578630e-02, -6.14645949e-04,
-2.52730539e-03, 2.29452504e-03, 9.93179274e-04,
-2.90897046e-03],
[ 6.50542649e-03, -1.21802492e-02, -1.15979975e-02,
3.04045796e-04, -5.20680938e-03, -5.98686142e-03,
1.02923429e-02, -2.00138614e-02, 2.01741327e-02,
1.01022851e-02],
[-4.86079371e-03, 3.91012290e-03, 1.68614404e-03,
-7.48395314e-03, -5.46253286e-04, -8.52062460e-03,
7.88574014e-03, 1.17413038e-02, -4.50282078e-03,
4.58895974e-03],
[ 7.92449064e-05, 3.17677343e-03, -1.36370957e-02,
-1.07877040e-02, 1.27669917e-02, 9.87697672e-03,
-4.51020384e-03, -9.63695813e-03, 5.63280517e-03,
-2.42555770e-03],
[ 1.30282128e-02, -1.77450315e-03, -9.25759878e-03,
-9.80420282e-06, -3.06460354e-03, -1.73859636e-03,
7.01046176e-03, -1.29886828e-02, -6.16477896e-03,
4.15576249e-03]], dtype=float32),
array([0., 0., 0., 0., 0., 0., 0., 0., 0., 0.], dtype=float32),
array([[1.],
[1.],
[1.],
[1.],
[1.],
[1.],
[1.],
[1.],
[1.],
[1.]], dtype=float32),
array([1.], dtype=float32)]
2.3 define initializer
使用tf.keras.initializers类中的方法实现自定义初始化：
def my_init():
"""Initializer that generates tensors initialized to 1"""
return tf.keras.initializers.Ones()
model = tf.keras.Sequential()
model.add(tf.keras.layers.Dense(units=64,kernel_initializer=my_init()))
X = tf.random.normal(shape=(2,20))
Y = model(X)
Y
输出：
<tf.Tensor: id=28, shape=(2, 64), dtype=float32, numpy=
array([[4.813542, 4.813542, 4.813542, 4.813542, 4.813542, 4.813542,
4.813542, 4.813542, 4.813542, 4.813542, 4.813542, 4.813542,
4.813542, 4.813542, 4.813542, 4.813542, 4.813542, 4.813542,
4.813542, 4.813542, 4.813542, 4.813542, 4.813542, 4.813542,
4.813542, 4.813542, 4.813542, 4.813542, 4.813542, 4.813542,
4.813542, 4.813542, 4.813542, 4.813542, 4.813542, 4.813542,
4.813542, 4.813542, 4.813542, 4.813542, 4.813542, 4.813542,
4.813542, 4.813542, 4.813542, 4.813542, 4.813542, 4.813542,
4.813542, 4.813542, 4.813542, 4.813542, 4.813542, 4.813542,
4.813542, 4.813542, 4.813542, 4.813542, 4.813542, 4.813542,
4.813542, 4.813542, 4.813542, 4.813542],
[2.223597, 2.223597, 2.223597, 2.223597, 2.223597, 2.223597,
2.223597, 2.223597, 2.223597, 2.223597, 2.223597, 2.223597,
2.223597, 2.223597, 2.223597, 2.223597, 2.223597, 2.223597,
2.223597, 2.223597, 2.223597, 2.223597, 2.223597, 2.223597,
2.223597, 2.223597, 2.223597, 2.223597, 2.223597, 2.223597,
2.223597, 2.223597, 2.223597, 2.223597, 2.223597, 2.223597,
2.223597, 2.223597, 2.223597, 2.223597, 2.223597, 2.223597,
2.223597, 2.223597, 2.223597, 2.223597, 2.223597, 2.223597,
2.223597, 2.223597, 2.223597, 2.223597, 2.223597, 2.223597,
2.223597, 2.223597, 2.223597, 2.223597, 2.223597, 2.223597,
2.223597, 2.223597, 2.223597, 2.223597]], dtype=float32)>
model.weights[0]
输出：
<tf.Variable 'sequential_1/dense_1/kernel:0' shape=(20, 64) dtype=float32, numpy=
array([[1., 1., 1., ..., 1., 1., 1.],
[1., 1., 1., ..., 1., 1., 1.],
[1., 1., 1., ..., 1., 1., 1.],
...,
[1., 1., 1., ..., 1., 1., 1.],
[1., 1., 1., ..., 1., 1., 1.],
[1., 1., 1., ..., 1., 1., 1.]], dtype=float32)>