LightGBM's predict(X, pred_leaf=True) has no ONNX equivalent — the converter had no mechanism to output which leaf node each sample reached per tree.

Approach

Adds a decision_leaf=False parameter to the converter. When True, a second TreeEnsembleRegressor node is generated alongside the main prediction node, configured to output the sequential leaf index per tree (cast to int64). This avoids modifying the existing prediction graph and keeps the two concerns independent.

Output shape: [N, n_trees], dtype int64.

Changes

• convert/main.py + lightgbm/convert.py: Expose decision_leaf on the public convert_lightgbm() API and thread it through to parse_lightgbm()
• lightgbm/_parse.py: Propagate decision_leaf through parse functions; attach it to the operator and append a leaf_indices output variable when enabled; fix LgbmZipMap to only receive the first two outputs (label + probs) as inputs
• lightgbm/operator_converters/LightGbm.py: Add _add_decision_leaf_output() — builds target_* attrs from class_treeids/class_nodeids, assigns one output column per tree (n_targets = n_trees), and casts float output to int64; fix regressor branch to use operator.outputs[0].full_name instead of operator.output_full_names (which would incorrectly include leaf_indices)
• Shape calculators (Classifier.py, Regressor.py, Ranker.py): Accept the optional extra output; set Int64TensorType([N, n_trees]) when decision_leaf=True
• tests/lightgbm/test_lightgbm_decision_leaf.py: Tests for binary classifier, multiclass classifier, regressor, and the default no-leaf case

Usage

from onnxmltools.convert import convert_lightgbm
from onnxmltools.convert.common.data_types import FloatTensorType

onx = convert_lightgbm(
    model,
    initial_types=[("X", FloatTensorType([None, n_features]))],
    zipmap=False,
    decision_leaf=True,
)
# onx.graph.output → [label, probabilities, leaf_indices]
# leaf_indices: int64[N, n_trees]

sess = onnxruntime.InferenceSession(onx.SerializeToString())
label, proba, leaf_indices = sess.run(None, {"X": X})

• Fixes LightGBM: Support for outputting decision path and leaf #710

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.