Switching from XGBoost

The estimator layer is built so that the common XGBoost script runs with the class name swapped and nothing else. XGBRegressor becomes BonsaiRegressor, XGBClassifier becomes BonsaiClassifier, and the constructor arguments, fit shapes, and post-fit attributes you are likely using keep their spellings.

A canonical XGBoost script, unchanged except for the class name:

import bonsai
import numpy as np

rng = np.random.default_rng(0)
X = rng.random((600, 8), dtype=np.float32)
y = (X[:, 0] * 2 + np.sin(X[:, 1] * 6) + rng.normal(0, 0.2, 600)).astype(np.float32)
X_train, y_train, X_valid, y_valid = X[:400], y[:400], X[400:], y[400:]

est = bonsai.BonsaiRegressor(
    n_estimators=200,
    learning_rate=0.1,
    max_depth=5,
    subsample=0.8,
    colsample_bytree=0.8,
    min_child_weight=1.0,
    reg_lambda=1.0,
    objective="reg:squarederror",
    random_state=0,
    early_stopping_rounds=20,
)
est.fit(X_train, y_train, eval_set=[(X_valid, y_valid)], verbose=False)

print("best_iteration:", est.best_iteration)
print("best rmse:", round(est.best_score, 4))
curve = est.evals_result()["validation_0"]["rmse"]
print("eval curve, first and best:", round(curve[0], 4), round(min(curve), 4))
print("top feature:", int(est.feature_importances_.argmax()))

What maps automatically

XGBoost spelling	bonsai meaning
n_estimators, learning_rate, max_depth, max_leaves	same knobs (n_estimators wins over n_iters when both are set)
random_state, n_jobs	booster.random_seed, parallel.n_threads
reg_lambda, reg_alpha, gamma	tree.lambda_l2, tree.lambda_l1, tree.min_gain_to_split
min_child_weight	tree.min_child_hess: the same minimum hessian mass per child, same default of 1.0 (a row count under squared error)
subsample	row sampling; switches the sampler to bernoulli when sampler is at its default
colsample_bytree, max_bin, min_child_samples	tree.feature_fraction, bin_mapper.max_bin, tree.min_data_in_leaf
device="cuda"	the CUDA grower matching your grower choice (cuda_depthwise by default, cuda_oblivious for grower="oblivious")
objective="reg:squarederror" etc.	mse, reg:absoluteerror is mae, reg:quantileerror (+ quantile_alpha) is quantile, count:poisson is poisson; classifiers accept binary:logistic / multi:softprob / multi:softmax and derive the real objective from the class count
eval_set=[(X, y)]	the list form is native; with several entries the last one drives the eval history and early stopping, XGBoost's own convention
evals_result(), best_iteration, best_score	same shapes; the squared-error eval is presented as rmse (the exact root, so early stopping saw the same ordering)
save_model / load_model / apply / iteration_range	save / from_file / predict_leaf / num_iteration under their XGBoost names

What is deliberately different

Honest differences rather than missing spellings, so a silent behavior change never hides behind a familiar name.

• early_stopping_rounds lives in the constructor, which matches XGBoost 2.x and later; there is no fit(early_stopping_rounds=...).
• reg:pseudohubererror maps to bonsai's huber, which is the exact Huber loss, not the pseudo-Huber approximation; the huber_delta knob is params={"objective.huber_delta": ...}.
• iteration_range must start at 0: a boosted sum has no meaning without its head.
• verbose is accepted and ignored; bonsai prints one line on early stop and nothing per round.
• Categorical features go through OrderedTargetEncoder rather than an enable_categorical flag; the measurement behind that choice is decision 58.
• The native layer (train, Dataset, Model) is bonsai's own explicit API, not a DMatrix clone; callbacks, dask, and spark integrations are out of scope.
• Loading a saved classifier restores encoded 0..K-1 class ids (XGBoost's load_model convention); pickle the estimator to preserve original labels.

Every knob that has no alias is reachable as a dotted config key through params=; Parameters lists them all.