Tensorflow 子类化序贯模型

为了能够编写自定义的

call（）

并处理命名的输入，我想对序列模型进行子类化。然而，对于我来说，

\uuuu init\uuu

函数的微小更改已经导致了一些意想不到的行为。如果我尝试在调用

super（）后向子类添加新成员并初始化它。\uuu init\uuuu（）

模型将无法自动生成

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Conv2D, Activation, MaxPooling2D, Dense, Flatten
import tensorflow as tf
class Sequential2(Sequential):

    def __init__(self):
        super(Sequential2, self).__init__()
        self.custom_member = []

    def get_my_custom_member(self):
        return self.custom_member

model = Sequential2()

if tf.keras.backend.image_data_format() == 'channels_first':
    input_shape = (1, 28, 28)
else:
    assert tf.keras.backend.image_data_format() == 'channels_last'
    input_shape = (28, 28, 1)

layers = [Conv2D(32, (3, 3), input_shape=input_shape)]

for layer in layers:
    model.add(layer)

model.add(Dense(10))
model.add(Activation('relu'))

model.summary()

输出失败：

ValueError:此模型尚未生成。首先通过调用`Build（）`或调用`fit（）`和一些数据来构建模型，或者在第一层中指定一个`input_shape`参数进行自动构建。

但是，如果省略了

self.custom_member=[]

，它将按预期工作

---

我错过了什么？（使用Tensorflow 1.14进行测试）

此问题已在TF 2.2中修复。您可以参考如下所示的工作代码

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Conv2D, Activation, MaxPooling2D, Dense, Flatten

import tensorflow as tf
print(tf.__version__)

class Sequential2(Sequential):

    def __init__(self):
        super(Sequential2, self).__init__()
        self.custom_member = []

    def get_my_custom_member(self):
        return self.custom_member

model = Sequential2()

if tf.keras.backend.image_data_format() == 'channels_first':
    input_shape = (1, 28, 28)
else:
    assert tf.keras.backend.image_data_format() == 'channels_last'
    input_shape = (28, 28, 1)

layers = [Conv2D(32, (3, 3), input_shape=input_shape)]

for layer in layers:
    model.add(layer)

model.add(Dense(10))
model.add(Activation('relu'))

model.summary()

输出：

2.2.0
Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d (Conv2D)              (None, 26, 26, 32)        320       
_________________________________________________________________
dense (Dense)                (None, 26, 26, 10)        330       
_________________________________________________________________
activation (Activation)      (None, 26, 26, 10)        0         
=================================================================
Total params: 650
Trainable params: 650
Non-trainable params: 0
_________________________________________________________________

甚至不要使用TF2.2（我也遇到了类似的问题）。[背景]