How to Disable Attribute Pruning in ONNX Runtime for Performance Optimization of Ensemble Tree Regressor in C++

[uqfam]

When I run a Random Forest Regressor / TreeEnsembleRegressor model in ONNX Runtime in C++, it automatically prunes removable attributes every time I load the model into ORT. This process takes over two minutes, which is by far the biggest chuck of the runtime of the ORT session. Is there any way I can disable the attribute pruning to make my model more efficient?

Here is the corresponding message i get when running the onnx model:

2024-11-14 12:05:55.0911151 [I:onnxruntime:, session_state.cc:260 onnxruntime::SessionState::PruneRemovableAttributes] removed 12 removable attributes for node 'TreeEnsembleRegressor' ('TreeEnsembleRegressor'), among attributes: base_values, nodes_falsenodeids, nodes_featureids, nodes_hitrates, nodes_missing_value_tracks_true, nodes_modes, nodes_nodeids, nodes_treeids, nodes_truenodeids, nodes_values, target_ids, target_treeids, target_nodeids, target_weightsbase_values_as_tensor, nodes_hitrates_as_tensor, nodes_values_as_tensor, class_weights_as_tensor.

[skottmckay]

There's no current way to disable that running.

@xadupre should this be configurable?

[xadupre]

We could make it optional but I'm surprised that takes that much time. Maybe the random forest has any nodes and it takes time to remove the unique array containing many strings (nodes_nodes). That would be worth to profile to understand which part of the code takes time. The array of string will disappears in the next version of onnx. onnxruntime has been updated to support it. We still need to update the converters.