系统信息

• 是否编写了自定义代码(与使用TensorFlow提供的库存示例脚本相反):是
• OS平台和发行版(例如，Linux Ubuntu 16.04):Ubuntu 18.04.1
• 从哪里安装的TensorFlow(源代码或二进制文件):二进制文件
• TensorFlow版本(请使用以下命令):v2.3.0-54-gfcc4b966f1 2.3.1
• Python版本：3.6.9
• CUDA/cuDNN版本：10.1
• GPU型号和内存：GTX 2080 Ti
  描述当前行为

从包含子层的层中加载权重，但在.weights属性中未正确列出。这不仅会引起困惑，还会破坏对权重的进一步优化或修改。不过它们还是被正确恢复了。
当使用ListWrapper时，问题不会出现。

描述预期行为

当再次加载模型时，List中的子层变量应包含在.weights中

独立代码重现问题

import tensorflow as tf

class TestLayer(tf.keras.layers.Layer):

    def __init__(self, **kwargs):
        super(TestLayer, self).__init__(**kwargs)

        self.static_layer = tf.keras.layers.Dense(128)
        self.my_layers = tf.python.training.tracking.data_structures.List()
        for i in range(4):
            layer = tf.keras.layers.Dense(128)
            self.my_layers.append(layer)

    def call(self, x):
        x = self.static_layer(x)

        for layer in self.my_layers:
            x = layer(x)

        return x

    def get_config(self):
        return super(TestLayer, self).get_config()

model = tf.keras.Sequential([TestLayer()])

x = tf.constant(42.0, shape=[1,1])
y1 = model(x)

model.save('my_test_model', save_format='tf')

model_loaded = tf.keras.models.load_model('my_test_model')
y2 = model_loaded(x)

# output
model.summary()
model_loaded.summary()
print('n vars: ', len(model.weights), ' ', len(model_loaded.weights))
print('diff: ', tf.norm(y1-y2))

输出

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
test_layer (TestLayer)       (None, 128)               66304     
=================================================================
Total params: 66,304
Trainable params: 66,304
Non-trainable params: 0
_________________________________________________________________
Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
test_layer (TestLayer)       (None, 128)               256       
=================================================================
Total params: 256
Trainable params: 256
Non-trainable params: 0
_________________________________________________________________
n vars:  10   2
diff:  tf.Tensor(0.0, shape=(), dtype=float32)