Skip to content

Latest commit

 

History

History
192 lines (142 loc) · 7.74 KB

4.2_parameters.md

File metadata and controls

192 lines (142 loc) · 7.74 KB
Raw

模型参数的访问、初始化和共享

在 3.3 节（“线性回归的简洁实现”）一节中，我们通过init模块来初始化模型的全部参数。我们也介绍了访问模型参数的简单方法。本节将深入讲解如何访问和初始化模型参数，以及如何在多个层之间共享同一份模型参数。

我们先定义一个与上一节中相同的含单隐藏层的多层感知机。我们依然使用默认方式初始化它的参数，并做一次前向计算。

import tensorflow as tf
import numpy as np
print(tf.__version__)
2.0.0

net = tf.keras.models.Sequential()
net.add(tf.keras.layers.Flatten())
net.add(tf.keras.layers.Dense(256,activation=tf.nn.relu))
net.add(tf.keras.layers.Dense(10))

X = tf.random.uniform((2,20))
Y = net(X)
Y
<tf.Tensor: id=62, shape=(2, 10), dtype=float32, numpy=
array([[ 0.15294254,  0.0355227 ,  0.05113338,  0.06625789,  0.12223213,
        -0.5954561 ,  0.38035268, -0.17244355,  0.6725004 ,  0.00750941],
       [ 0.12288147, -0.2162356 , -0.02103446,  0.14871466,  0.10256162,
        -0.57710034,  0.22278625, -0.21283135,  0.52407515, -0.1426214 ]],
      dtype=float32)>

4.2.1 access model parameters

对于使用Sequential类构造的神经网络，我们可以通过weights属性来访问网络任一层的权重。回忆一下上一节中提到的Sequential类与tf.keras.Model类的继承关系。对于Sequential实例中含模型参数的层，我们可以通过tf.keras.Model类的weights属性来访问该层包含的所有参数。下面，访问多层感知机net中隐藏层的所有参数。索引0表示隐藏层为Sequential实例最先添加的层。

net.weights[0], type(net.weights[0])
(<tf.Variable 'sequential/dense/kernel:0' shape=(20, 256) dtype=float32, numpy=
 array([[-0.07852519, -0.03260126,  0.12601742, ...,  0.11949158,
          0.10042094, -0.10598273],
        [ 0.03567271, -0.11624913,  0.04699135, ..., -0.12115637,
          0.07733515,  0.13183317],
        [ 0.03837337, -0.11566538, -0.03314627, ..., -0.10877015,
          0.09273799, -0.07031895],
        ...,
        [-0.03430544, -0.00946991, -0.02949082, ..., -0.0956497 ,
         -0.13907745,  0.10703176],
        [ 0.00447187, -0.07251608,  0.08081181, ...,  0.02697623,
          0.05394638, -0.01623751],
        [-0.01946831, -0.00950103, -0.14190955, ..., -0.09374787,
          0.08714674,  0.12475103]], dtype=float32)>,
 tensorflow.python.ops.resource_variable_ops.ResourceVariable)

4.2.2 initialize params

我们在[“数值稳定性和模型初始化”]一节中描述了模型的默认初始化方法：权重参数元素为[-0.07, 0.07]之间均匀分布的随机数，偏差参数则全为0。但我们经常需要使用其他方法来初始化权重。在下面的例子中，我们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，并依然将偏差参数清零。

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

net = Linear()
net(X)
net.get_weights()
[array([[-0.00306494,  0.01149799,  0.00900665, -0.00952527, -0.00651997,
      0.00010531,  0.00802666, -0.01102469,  0.01838934,  0.00915548],
    [ 0.00401672,  0.01788972, -0.00245794, -0.01051202,  0.02268461,
     -0.00271502, -0.00447782,  0.00636486,  0.00408998, -0.01373187],
    [-0.00468962, -0.00180526, -0.0117501 ,  0.01840584,  0.00044537,
     -0.00745311,  0.01155732, -0.00615015, -0.00942082, -0.00023081],
    [-0.01116156, -0.00614527, -0.00119119, -0.00843481,  0.01192368,
      0.00889105, -0.01000126, -0.0017869 , -0.00833272,  0.0019026 ],
    [ 0.0183291 , -0.00640716,  0.00936602,  0.01040828, -0.00140882,
     -0.00143817,  0.00126366,  0.01094474,  0.0132029 ,  0.00405393],
    [-0.00548183, -0.00489746, -0.01264372, -0.00501967,  0.00602909,
      0.00439432,  0.02449438,  0.00426046, -0.0017243 , -0.00319188],
    [-0.00034199, -0.00648715, -0.00694025, -0.00984227,  0.02798587,
     -0.01283635, -0.01735584, -0.00181439,  0.01585936,  0.00348289],
    [ 0.00181157, -0.00343991,  0.01415697, -0.00160312,  0.0018713 ,
     -0.00968461, -0.00268579,  0.01320006, -0.00041133, -0.01282531],
    [-0.0145638 ,  0.0096653 , -0.00787722, -0.00073892, -0.00222261,
      0.0031008 , -0.01858314,  0.00559973,  0.00439452, -0.02467434],
    [-0.00303086,  0.0015006 , -0.00920389,  0.01035136, -0.00040001,
     -0.00945453, -0.00506378,  0.00816534,  0.00347233,  0.01201165],
    [ 0.01979353,  0.00881971, -0.00060045, -0.00671935,  0.02482731,
     -0.0039808 ,  0.01195751, -0.00499541, -0.01421177,  0.00125722],
    [-0.00206965,  0.00737946,  0.02711954, -0.00566722, -0.01916223,
      0.00635906, -0.00112362,  0.00351852,  0.0027598 ,  0.00804986],
    [ 0.00190901,  0.00799948, -0.01007551, -0.00751526,  0.0027352 ,
     -0.00126002,  0.00079498, -0.00190032, -0.00912007,  0.00432031],
    [-0.00574654,  0.00703932,  0.00375365,  0.01700558, -0.00392553,
      0.00246399,  0.00686003, -0.00327425, -0.00158563,  0.01139532],
    [-0.010441  , -0.01566261,  0.01807244, -0.01265192, -0.00422926,
     -0.00729915, -0.00717674, -0.00036729,  0.00728995,  0.0034066 ],
    [-0.00497032, -0.01395558, -0.00276683,  0.0114197 , -0.01044411,
     -0.01518542,  0.00793149, -0.00169621, -0.008745  , -0.00825851],
    [-0.00098009, -0.00765272, -0.01993775,  0.0207908 , -0.0088134 ,
      0.01211826,  0.0033179 ,  0.0064116 ,  0.00399073,  0.00067746],
    [ 0.00282402,  0.00589997,  0.00674444, -0.01209166, -0.00875635,
      0.01789016, -0.00037993,  0.00392861,  0.02248183, -0.00427692],
    [-0.00629026, -0.01388059,  0.0160582 ,  0.00855581,  0.00170209,
      0.00430258,  0.0092911 ,  0.00232163,  0.00591121,  0.02038265],
    [-0.00792203, -0.00259904, -0.00109487, -0.00959524, -0.00030968,
     -0.01322429,  0.00489308,  0.00503101,  0.01801165,  0.00972504]],
   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)]

4.2.3 define initializer

可以使用tf.keras.initializers类中的方法实现自定义初始化。

def my_init():
    return tf.keras.initializers.Ones()

model = tf.keras.models.Sequential()
model.add(tf.keras.layers.Dense(64, kernel_initializer=my_init()))

Y = model(X)
model.weights[0]
<tf.Variable 'sequential_1/dense_4/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)>

注：本节除了代码之外与原书基本相同，原书传送门