我想一个解决方案是用pickle模块单独保存您的自定义属性：

import pickle

# Your custom attribute
custom_attr = ["one", "two", "three"]

# Save the custom attribute to a file
with open("custom_attr.pkl", "wb") as file:
    pickle.dump(custom_attr, file)

字符串
在重新加载模型之后，也加载这个pickle并设置属性setattr(model, custom_attr)

我通过子类化Sequential类并向构造函数添加一个参数来解决这个问题：

class SequentialWithCustomAttr(keras.Sequential):
    def __init__(self, custom_attr=[], layers=None, *args, **kwargs):
        super().__init__(layers=layers, trainable=kwargs["trainable"], name=kwargs["name"])
        self.custom_attr = custom_attr

    def get_config(self):
        config = super().get_config()
        config.update({"custom_attr": self.custom_attr})

        return config

    @classmethod
    def from_config(cls, config, custom_objects=None):
        custom_attr = config.pop("custom_attr", None)

        # Deserialize layers one by one by using their configs.
        layers_confs = config.pop("layers", None)
        layers = list(map(keras.saving.deserialize_keras_object, layers_confs))

        # Create an instance of class SequentialWithCustomAttr.
        model = cls(custom_attr=custom_attr, layers=layers, **config)

        return model

字符串