Custom Estimators

FlowMatchingBDT accepts any scikit-learn-compatible regressor via the estimator parameter. Each flow step clones this estimator and wraps it in MultiOutputRegressor to handle multi-dimensional velocity prediction.

from sklearn.datasets import make_moons
data, _ = make_moons(n_samples=500, noise=0.1, random_state=0)

Using Decision Tree

from sklearn.tree import DecisionTreeRegressor
from flowmatching_bdt import FlowMatchingBDT

model = FlowMatchingBDT(
    estimator=DecisionTreeRegressor(),
    n_flow_steps=5,
    n_duplicates=10,
)
model.fit(data)

Using Extra Trees

from sklearn.ensemble import ExtraTreesRegressor
from flowmatching_bdt import FlowMatchingBDT

model = FlowMatchingBDT(
    estimator=ExtraTreesRegressor(n_estimators=10),
    n_flow_steps=5,
    n_duplicates=10,
)
model.fit(data)

Requirements for Custom Estimators

Any estimator you pass must:

1. Implement fit(X, y) — called with noised samples X and velocity targets y
2. Implement predict(X) — called during Euler integration to predict the velocity field
3. Be compatible with sklearn.base.clone — the estimator is cloned once per flow step

By default, the estimator is wrapped in MultiOutputRegressor, so it only needs to support single-output regression. If your estimator natively handles multi-output (e.g. neural networks), set multi_output=True to skip the wrapper:

model = FlowMatchingBDT(
    estimator=DecisionTreeRegressor(),
    multi_output=True,
    n_flow_steps=5,
    n_duplicates=10,
)