TensorFlow加载Keras模型出现不一致引用警告，是否需处理？

加载通过TFX的Pusher组件保存的、包含自定义层的Keras模型时，出现多条如下警告：

WARNING:tensorflow:Inconsistent references when loading the checkpoint into this object graph. For example, in the saved checkpoint object, `model.layer.weight` and `model.layer_copy.weight` reference the same variable, while in the current object these are two different variables. The referenced variables are:(<keras.layers.core.dense.Dense object at 0x7f1d0c557190> and <keras.engine.functional.Functional object at 0x7f1d0c557f10>).
WARNING:tensorflow:Inconsistent references when loading the checkpoint into this object graph. For example, in the saved checkpoint object, `model.layer.weight` and `model.layer_copy.weight` reference the same variable, while in the current object these are two different variables. The referenced variables are:(<keras.layers.core.dense.Dense object at 0x7f1d0c5577c0> and <Layers.SingleOutputWithName.SingleOutputWithName object at 0x7f1d0c4cde80>).
WARNING:tensorflow:Inconsistent references when loading the checkpoint into this object graph. For example, in the saved checkpoint object, `model.layer.weight` and `model.layer_copy.weight` reference the same variable, while in the current object these are two different variables. The referenced variables are:(<keras.engine.functional.Functional object at 0x7f1d0c4cd7c0> and <keras.saving.legacy.saved_model.load.TensorFlowTransform>TransformFeaturesLayer object at 0x7f1d0c557e20>).
WARNING:tensorflow:Inconsistent references when loading the checkpoint into this object graph. For example, in the saved checkpoint object, `model.layer.weight` and `model.layer_copy.weight` reference the same variable, while in the current object these are two different variables. The referenced variables are:(<keras.saving.legacy.saved_model.load.TensorFlowTransform>TransformFeaturesLayer object at 0x7f1d0c557e20> and <keras.engine.input_layer.InputLayer object at 0x7f1d0d5e1a00>).
WARNING:tensorflow:Inconsistent references when loading the checkpoint into this object graph. For example, in the saved checkpoint object, `model.layer.weight` and `model.layer_copy.weight` reference the same variable, while in the current object these are two different variables. The referenced variables are:(<keras.engine.functional.Functional object at 0x7f1d0c4cd7c0> and <keras.saving.legacy.saved_model.load.TensorFlowTransform>TransformFeaturesLayer object at 0x7f1d0c557e20>).

目前模型能正常加载运行，但不确定这些警告是否会影响模型输出，想咨询这些警告是否重要，是否需要进行处理。

我的模型结构简单，未使用权重共享等复杂机制，仅包含如下自定义层：

import tensorflow as tf
from tensorflow.keras import layers


@tf.keras.utils.register_keras_serializable()
class SingleOutputWithName(layers.Layer):
    def __init__(self, label_name: str, **kwargs):
        super(SingleOutputWithName, self).__init__()
        self.label_name = label_name

    def call(self, x):
        return {self.label_name: x,
                'dummy': x}

    def get_config(self):
        config = super().get_config()
        config['label_name'] = self.label_name
        return config

这些警告暂时不会影响模型的输出正确性，但属于TensorFlow加载模型时的内部引用不一致问题，具体分析和处理建议如下：

• 警告本质：模型保存时的对象图中，某些变量被多个层共享引用，但加载后这些引用指向了不同的对象实例。你的模型没有手动设置权重共享，大概率是TFX或Keras在保存/加载流程中自动生成的临时层（比如TransformFeaturesLayer这类TFX相关预处理层）导致的内部引用差异。
• 影响判断：既然模型能正常加载运行，且你的自定义层SingleOutputWithName本身没有可训练权重（仅做输出格式转换），这些引用不一致不会改变模型的计算逻辑和权重值，输出结果是可靠的。

如果想要消除警告，可以尝试以下方法：

• 确认加载模型时自定义层的注册正确（你已经用@tf.keras.utils.register_keras_serializable()完成注册，这步没问题）；
• 用tf.keras.models.load_model()加载时，指定compile=False（如果不需要重新编译模型），减少加载时的额外检查；
• 检查TFX与TensorFlow的版本兼容性，部分版本组合会存在这类小问题，升级到稳定版可能解决。

总结：如果当前模型运行正常、输出符合预期，可以暂时不处理这些警告；若追求无警告的运行环境，再尝试上述优化方案。

内容的提问来源于stack exchange，提问作者Mehran