如何将自定义树添加到自定义 Keras 层？

Question

我正在尝试汇总以下教程：

1. Creating decision tree by hand
2. Custom layers via subclassing
3. Composing Decision Forest and Neural Network models

目标是 1. 创建自定义树，2. 将其嵌入自定义层，3. 将其与其他层结合到模型中。

问题在于，在第 1 步中，通过使用 RandomForestBuilder，模型被序列化和反序列化，导致对象类型为 keras.saving.saved_model.load.C​​oreModel

但是，第 3 步中的教程通过 tfdf.keras.RandomForestModel 嵌入树层

理想情况下，自定义层将通过在其构造函数中调用 RandomForestBuilder 来创建自定义树，但是，考虑到模型的导出和加载，这并不简单。

下面给出了输入层结构的错误，如果省略了前者，则会给出从 SavedModel 加载的没有匹配的具体函数调用的错误：

第 1 步：

builder = tfdf.builder.RandomForestBuilder(
    path="/tmp/manual_model",
    objective = tfdf.py_tree.objective.RegressionObjective(label='tree_result')
)

Tree = tfdf.py_tree.tree.Tree
SimpleColumnSpec = tfdf.py_tree.dataspec.SimpleColumnSpec
ColumnType = tfdf.py_tree.dataspec.ColumnType
RegressionValue = tfdf.py_tree.value.RegressionValue

NonLeafNode = tfdf.py_tree.node.NonLeafNode
LeafNode = tfdf.py_tree.node.LeafNode
NumericalHigherThanCondition = tfdf.py_tree.condition.NumericalHigherThanCondition
CategoricalIsInCondition = tfdf.py_tree.condition.CategoricalIsInCondition

tree = Tree(
    NonLeafNode(
        condition=CategoricalIsInCondition(
            feature=SimpleColumnSpec(name='feature_name', type=ColumnType.CATEGORICAL),
            mask=['class_1'],
            missing_evaluation=False
        ),
        pos_child = LeafNode(value=RegressionValue(value=0.5)),
        neg_child = LeafNode(value=RegressionValue(value=0.6))
    )
)

builder.add_tree(tree)
builder.close()
custom_tree = tf.keras.models.load_model("/tmp/manual_model")

第 2 步：

class CustomTree(tf.keras.layers.Layer):
  def __init__(self, custom_tree):
    super(CustomTree, self).__init__()
    self.custom_tree = custom_tree

  def call(self, inputs):
    return self.custom_tree(inputs)


input_layer = tf.keras.layers.Input(shape=(None,), name='feature_name', dtype=tf.string)
output_layer = CustomTree(custom_tree)(input_layer)

model = tf.keras.models.Model(input_layer, output_layer, name='SomeModel')

model.predict(tf.data.Dataset.from_tensor_slices(
    {'feature_name': ['class_1','class_2']}
).batch(1))

Answer 1

找到以下解决方案：

1. 以 yggdrasil 格式导出自定义树：

    builder = tfdf.builder.RandomForestBuilder(
        model_format = tfdf.builder.ModelFormat.YGGDRASIL_DECISION_FOREST,
        path=self._temp_directory,
        objective = objective)
   

2. 通过关闭构建器保存模型后，将树加载为 CoreModel：

    builder.close()
   
    inspector_lib = tfdf.component.inspector.inspector
    inspector = inspector_lib.make_inspector(self._temp_directory, file_prefix=None)
   
    custom_tree = tfdf.keras.CoreModel(
        task=objective.task,
        learner='MANUAL',
        ranking_group=None,
        temp_directory=self._temp_directory)
   
    custom_tree._set_from_yggdrasil_model(
        inspector,
        self._temp_directory,
        file_prefix=None, 
        input_model_signature_fn=builder._input_model_signature_fn)
   

3. 使用自定义拟合函数创建 tfdf.keras.CartModel 的模型子类，将自定义树结构分配给模型：

   class CustomDT(tfdf.keras.CartModel):
       def __init__(self, name=None, **kwargs):
           super(CustomDT, self).__init__(name=name, **kwargs)
   
       def fit(self):
           custom_tree = self.create_custom_tree()
           self._model = custom_tree._model
           self._is_trained.assign(True)
           self._semantics = custom_tree._semantics