将TensorFlow决策树的RandomForest模型转TFlite时遇错求助

尝试将tensor_flow_decision_tree库中的RandomForest模型转换为TFlite模型时，出现如下ConverterError错误：

ConverterError: :0: error:
loc(fused["SimpleMLCreateModelResource:",
"SimpleMLCreateModelResource"]): 'tf.SimpleMLCreateModelResource' op
is neither a custom op nor a flex op :0: note:
loc(fused["SimpleMLCreateModelResource:",
"SimpleMLCreateModelResource"]): Error code: ERROR_NEEDS_CUSTOM_OPS
/usr/local/lib/python3.8/dist-packages/tensorflow/python/framework/func_graph.py:749:0:
error: 'tf.SimpleMLInferenceOpWithHandle' op is neither a custom op
nor a flex op :0: note: loc(fused["StatefulPartitionedCall:",
"StatefulPartitionedCall"]): called from
/usr/local/lib/python3.8/dist-packages/tensorflow/python/framework/func_graph.py:749:0:
note: Error code: ERROR_NEEDS_CUSTOM_OPS :0: error: failed
while converting: 'main': Some ops in the model are custom ops, See
instructions to implement custom ops:
Custom ops:
SimpleMLCreateModelResource, SimpleMLInferenceOpWithHandle Details:
tf.SimpleMLCreateModelResource() -> (tensor<!tf_type.resource>) :
{container = "", device = "", shared_name =
"simple_ml_model_243450c3-97d3-43ae-bd29-ea9628cd031d"}
tf.SimpleMLInferenceOpWithHandle(tensor<?x12xf32>, tensor<0x0xf32>,
tensor<0x0xi32>, tensor<0xi32>, tensor<1xi64>, tensor<1xi64>,
tensor<!tf_type.resource>) -> (tensor<?x2xf32>,
tensor<2x!tf_type.string>) : {dense_output_dim = 2 : i64, device = ""}

已查阅官方自定义算子指南但未找到有效解决方法，求处理方案。

• 使用TensorFlow Decision Forests（TF-DF）官方提供的专用TFLite转换API，不要用标准的tf.lite.TFLiteConverter。TF-DF的模型依赖专属状态算子，标准转换器无法处理，专用API会自动处理这些算子的序列化：

  import tensorflow_decision_forests as tfdf
  import tensorflow as tf
  
  # 加载训练好的TF-DF随机森林模型
  model = tfdf.keras.load_model("your_model_path")
  
  # 获取TF-DF专属转换器
  converter = tfdf.lite.convert.get_converter(model)
  # 执行转换
  tflite_model = converter.convert()
  
  # 保存转换后的模型
  with open("converted_model.tflite", "wb") as f:
      f.write(tflite_model)
  

• 确保TF-DF与TensorFlow版本严格兼容。TF-DF对TensorFlow版本有明确依赖，版本不匹配会导致算子识别失败。可以通过以下命令查看当前版本：

  pip show tensorflow_decision_forests
  pip show tensorflow
  

  根据TF-DF官方说明调整到兼容版本。

• 尝试导出无状态TFLite模型。TF-DF模型默认是有状态的，依赖资源算子，可通过转换参数强制转为无状态模式，避开资源算子问题：

  converter = tfdf.lite.convert.get_converter(model, use_training_model=False)
  tflite_model = converter.convert()
  

  注意：无状态模式会失去模型在线更新能力，仅适合推理场景。

• 固定模型输入形状。如果模型输入是动态维度，可能导致转换失败，可在转换前固定输入形状：

  # 假设输入特征维度为12
  model.build(input_shape=[None, 12])
  

内容的提问来源于stack exchange，提问作者WinterSolstice