keras.layers.Layer中無(wú)法定義name問(wèn)題

在使用keras之中定義層的時(shí)候，如果按照以下的方法直接在keras.layers.Layer中定義相應(yīng)的self.name

會(huì)發(fā)生相應(yīng)的報(bào)錯(cuò)

import params as pp
import params_flow as pf
import tensorflow as tf
import tensorflow.keras as keras
class MyParams(tf.keras.layers.Layer):   
    def __init__(self):
        super(MyParams,self).__init__()
        self.name = 'hello'
    def build(self,input_shape):
        self.dense0 = keras.layers.Dense(units = 25,
                                #kernel_initializer = self.create_initializer(),
                                name = "dense0")
    def call(self,inputs):
        results = self.dense0(inputs)
data = MyParams()
print(data.name)

此時(shí)會(huì)相應(yīng)的報(bào)錯(cuò)

AttributeError: Can't set the attribute "name", likely because it conflicts with an existing read-only @property of the object. Please choose a different name.

也就是說(shuō)，在keras之中的"self.name"為只讀屬性，不能夠被定義，此時(shí)需要更換另外一個(gè)名字。

但是只有name這個(gè)屬性能夠?qū)?yīng)到self.weights的名稱之中，那么這里我們?cè)撊绾味x呢

此時(shí)我發(fā)現(xiàn)，直接定義在__init__函數(shù)之前，可以對(duì)keras的名稱完成相應(yīng)的定義

import params as pp
import params_flow as pf
import tensorflow as tf
import tensorflow.keras as keras
class MyParams(tf.keras.layers.Layer):   
    name = 'hello'
    def __init__(self):
        super(MyParams,self).__init__()
        
    def build(self,input_shape):
        self.dense0 = keras.layers.Dense(units = 25,
                                #kernel_initializer = self.create_initializer(),
                                name = "dense0")
    def call(self,inputs):
        results = self.dense0(inputs)
data = MyParams()
print(data.name)

同時(shí)我們傳入一個(gè)tensor的input_ids類型進(jìn)行相應(yīng)的輸入，發(fā)現(xiàn)已經(jīng)能夠?qū)eights的權(quán)重名稱進(jìn)行改變了

input_ids = keras.layers.Input(shape=(50,), dtype='int32', name="input_ids")
outputs = data(input_ids)
data.weights

對(duì)應(yīng)的輸出內(nèi)容如下

[<tf.Variable 'hello/dense0/kernel:0' shape=(50, 25) dtype=float32, numpy=
 array([[-0.10505645,  0.09756875,  0.14427656, ..., -0.17254017,
         -0.18592533, -0.13920134],
        [-0.10033116, -0.17831415, -0.03435555, ..., -0.02460951,
          0.13194972,  0.21918347],
        [ 0.15699485, -0.24836   ,  0.01044622, ...,  0.04577217,
          0.23334488,  0.09155059],
        ...,
        [-0.22210473,  0.14221036,  0.07721925, ...,  0.03358698,
          0.08100349,  0.15415356],
        [-0.1433322 , -0.00878078, -0.0760702 , ..., -0.06091703,
          0.18796855, -0.19009456],
        [-0.0446853 ,  0.14639893,  0.1729418 , ..., -0.04699725,
          0.12940568, -0.24003454]], dtype=float32)>,
 <tf.Variable 'hello/dense0/bias:0' shape=(25,) dtype=float32, numpy=
 array([0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0.], dtype=float32)>]

可以發(fā)現(xiàn)這里的權(quán)重內(nèi)容已經(jīng)變成hello打頭的內(nèi)容了，然而此時(shí)又一個(gè)對(duì)應(yīng)的問(wèn)題出現(xiàn)了，這得益于__init__函數(shù)前面內(nèi)容的特殊性：

如果我們將對(duì)應(yīng)的內(nèi)容改為如下

import params as pp
import params_flow as pf
import tensorflow as tf
import tensorflow.keras as keras
class MyParams(tf.keras.layers.Layer):   
    print('begining')
    name = 'hello'
    def __init__(self):
        super(MyParams,self).__init__()
        
    def build(self,input_shape):
        self.dense0 = keras.layers.Dense(units = 25,
                                #kernel_initializer = self.create_initializer(),
                                name = "dense0")
    def call(self,inputs):
        results = self.dense0(inputs)
data = MyParams()
print(data.name)

此時(shí)運(yùn)行的時(shí)候會(huì)運(yùn)行一次print(‘begining’)，輸出對(duì)應(yīng)的begining的內(nèi)容，

也就是說(shuō)__init__函數(shù)之前的內(nèi)容會(huì)在定義MyParams這個(gè)類的時(shí)候就調(diào)用，而MyParams這個(gè)類只會(huì)被定義一次，也就是說(shuō)__init__函數(shù)之前的內(nèi)容只會(huì)被調(diào)用一次。

這樣就帶來(lái)了一個(gè)問(wèn)題，也就是name = 'hello’這里的name沒(méi)有辦法修改，也就是說(shuō)我們需要想一種辦法將__init__函數(shù)前面的name='hello’修改一次，這里究竟該如何修改呢？

此時(shí)我們只需要在定義之后每次使用的時(shí)候重新對(duì)name的值進(jìn)行定義即可

具體的例子如下

import tensorflow as tf
import tensorflow.keras as keras
class MyParams(tf.keras.layers.Layer):   
    #print('begining')
    name = 'hello'
    def __init__(self):
        super(MyParams,self).__init__()
    def build(self,input_shape):
        self.dense0 = keras.layers.Dense(units = 25,
                                #kernel_initializer = self.create_initializer(),
                                name = "dense0")
    def call(self,inputs):
        results = self.dense0(inputs)
data = MyParams()
print(data.name)
data.name = 'hello10000'
print(data.name)

輸出的內(nèi)容為

進(jìn)一步查看一下對(duì)應(yīng)的weight的屬性

可以看出這里weights之中的屬性名稱已經(jīng)成功地被我們定義了

總結(jié)

由于name的屬性的特殊性，如果在__init__函數(shù)之中直接定義可讀變量會(huì)造成報(bào)錯(cuò)，此時(shí)我們需要調(diào)整思路，通過(guò)__init__函數(shù)之前定義name變量實(shí)現(xiàn)對(duì)于name的修改

以上為個(gè)人經(jīng)驗(yàn)，希望能給大家一個(gè)參考，也希望大家多多支持腳本之家。