TopolearnProduction models must be reloadable reliably, ideally independent of Python pickles.
XGBoost has its own model format, separate from Python pickling:
model.save_model("model.json")writes the trees and parameters as JSON, readable by any XGBoost binding (Python, R, JVM, C) and compatible across XGBoost versions,model.save_model("model.ubj")writes the same content as binary UBJSON: smaller and faster to load,XGBClassifier().load_model("model.json")orxgb.Booster(model_file="model.json")loads it back.
The scikit-learn wrapper stores extra attributes it needs (such as the classes and the best iteration from early stopping) inside the file, so loading into a fresh XGBClassifier restores a usable estimator.
When should you pickle instead? If the model is inside a scikit-learn Pipeline with preprocessing steps, you typically save the whole pipeline with joblib. That works, but it ties the file to your library versions, so record them; and only load pickles you trust. A common compromise is to save the XGBoost model in JSON and the rest of the pipeline separately.
The older binary .deprecated format is gone; always use .json or .ubj extensions. To inspect what a model learned, booster.get_dump() or booster.trees_to_dataframe() list every tree's splits.
Example
import numpy as np
from sklearn.datasets import load_iris
from xgboost import XGBClassifier
X, y = load_iris(return_X_y=True)
model = XGBClassifier(n_estimators=50, max_depth=3).fit(X, y)
model.save_model("iris_xgb.json") # portable, version-independent
loaded = XGBClassifier()
loaded.load_model("iris_xgb.json")
print(np.array_equal(model.predict(X), loaded.predict(X))) # True
print(loaded.classes_) # [0 1 2]
trees = loaded.get_booster().trees_to_dataframe()
print(trees["Tree"].nunique()) # 150 (50 rounds x 3 classes)Remember
save_modelto.jsonor.ubjis portable across versions and languages- Load into a fresh estimator with
load_model - Pickle/joblib only for whole pipelines, with pinned versions
Watch out for
- Pickling a bare XGBoost model and failing to load it after upgrading.