torus-ml/docs/references/api/torus.ml.sklearn.tree_to_numpy.md

module torus.ml.sklearn.tree_to_numpy

Implements the conversion of a tree model to a numpy function.

Global Variables

• MAX_BITWIDTH_BACKWARD_COMPATIBLE
• OPSET_VERSION_FOR_ONNX_EXPORT
• MSB_TO_KEEP_FOR_TREES
• MIN_CIRCUIT_THRESHOLD_FOR_TREES

---

function get_onnx_model

get_onnx_model(model, x: ndarray, framework: str) → ModelProto

Create ONNX model with Hummingbird convert method.

Args:

• model: The tree model to convert.
• x (numpy.ndarray): Dataset used to trace the tree inference and convert the model to ONNX.
• framework (str): The framework from which the ONNX model is generated.
• (options: 'xgboost', 'sklearn')

Returns:

• onnx.ModelProto: The ONNX model.

---

function workaround_squeeze_node_xgboost

workaround_squeeze_node_xgboost(onnx_model: ModelProto)

Workaround to fix torch issue that does not export the proper axis in the ONNX squeeze node.

FIXME: https://github.com/luxfi/torus-ml-internal/issues/2778 The squeeze ops does not have the proper dimensions. remove the following workaround when the issue is fixed Add the axis attribute to the Squeeze node

Args:

• onnx_model (onnx.ModelProto): The ONNX model.

---

function assert_add_node_and_constant_in_xgboost_regressor_graph

assert_add_node_and_constant_in_xgboost_regressor_graph(onnx_model: ModelProto)

Assert if an Add node with a specific constant exists in the ONNX graph.

Args:

• onnx_model (onnx.ModelProto): The ONNX model.

---

function add_transpose_after_last_node

add_transpose_after_last_node(
    onnx_model: ModelProto,
    fhe_ensembling: bool = False
)

Add transpose after last node.

Args:

• onnx_model (onnx.ModelProto): The ONNX model.
• fhe_ensembling (bool): Determines whether the sum of the trees' outputs is computed in FHE. Default to False.

---

function preprocess_tree_predictions

preprocess_tree_predictions(
    init_tensor: ndarray,
    output_n_bits: int
) → QuantizedArray

Apply post-processing from the graph.

Args:

• init_tensor (numpy.ndarray): Model parameters to be pre-processed.
• output_n_bits (int): The number of bits of the output.

Returns:

• QuantizedArray: Quantizer for the tree predictions.

---

function tree_onnx_graph_preprocessing

tree_onnx_graph_preprocessing(
    onnx_model: ModelProto,
    framework: str,
    expected_number_of_outputs: int,
    fhe_ensembling: bool = False
)

Apply pre-processing onto the ONNX graph.

Args:

• onnx_model (onnx.ModelProto): The ONNX model.
• framework (str): The framework from which the ONNX model is generated.
• (options: 'xgboost', 'sklearn')
• expected_number_of_outputs (int): The expected number of outputs in the ONNX model.
• fhe_ensembling (bool): Determines whether the sum of the trees' outputs is computed in FHE. Default to False.

---

function tree_values_preprocessing

tree_values_preprocessing(
    onnx_model: ModelProto,
    framework: str,
    output_n_bits: int
) → QuantizedArray

Pre-process tree values.

Args:

• onnx_model (onnx.ModelProto): The ONNX model.
• framework (str): The framework from which the ONNX model is generated.
• (options: 'xgboost', 'sklearn')
• output_n_bits (int): The number of bits of the output.

Returns:

• QuantizedArray: Quantizer for the tree predictions.

---

function tree_to_numpy

tree_to_numpy(
    model: BaseEstimator,
    x: ndarray,
    framework: str,
    use_rounding: bool = True,
    fhe_ensembling: bool = False,
    output_n_bits: int = 8
) → Tuple[Callable, List[UniformQuantizer], ModelProto]

Convert the tree inference to a numpy functions using Hummingbird.

Args:

• model (Callable): The tree model to convert.
• x (numpy.ndarray): The input data.
• use_rounding (bool): Determines whether the rounding feature is enabled or disabled. Default to True.
• fhe_ensembling (bool): Determines whether the sum of the trees' outputs is computed in FHE. Default to False.
• framework (str): The framework from which the ONNX model is generated.
• (options: 'xgboost', 'sklearn')
• output_n_bits (int): The number of bits of the output. Default to 8.

Returns:

• Tuple[Callable, List[QuantizedArray], onnx.ModelProto]: A tuple with a function that takes a numpy array and returns a numpy array, QuantizedArray object to quantize and de-quantize the output of the tree, and the ONNX model.

---

function onnx_fp32_model_to_quantized_model

onnx_fp32_model_to_quantized_model(
    onnx_model: ModelProto,
    n_bits: int,
    framework: str,
    expected_number_of_outputs: int,
    n_features: int,
    model_inputs: Optional[ndarray] = None
)

Build a FHE-compliant onnx-model using a fitted scikit-learn model.

Args:

• onnx_model (onnx.ModelProto): The fitted scikit-learn as a Hummingbird onnx model to convert
• n_bits (int): Number of bits to quantize the model. If an int is passed for n_bits, the value will be used for quantizing inputs and weights. If a dict is passed, then it should contain "op_inputs" and "op_weights" as keys with corresponding number of quantization bits so that: - op_inputs : number of bits to quantize the input values - op_weights: number of bits to quantize the learned parameters
• framework (str): either sklearn or xgboost
• expected_number_of_outputs (int): expected number of outputs
• n_features (int): number of features as inputs of the model
• model_inputs (Optional[numpy.ndarray]): optional dataset to use for quantization

Returns:

• onnx.ModelProto: The converted onnx model
• Optional[Tuple[int, int]]: Least significant bits to remove
• list[UniformQuantizer]: inputs quantizers
• list[UniformQuantizer]: outputs quantizers