build_artifact

import numpy as np
from sklearn.metrics import roc_auc_score
 
from _setup import (
    REASONS,
    X_test,
    X_train,
    bands,
    feature_names,
    latent,
    teacher_test_scores,
    whitebox,
    y_test,
    y_train,
)
 
from compileml.artifact import build_artifact, save_artifact
from compileml.runtime import decide, load_artifact
 
artifact = build_artifact(
    whitebox,
    feature_names,
    baseline=np.median(X_train, axis=0),  # -> features.baseline, the attribution reference
    band_edges=bands,                     # -> bands.edges_int
    calibration_latent=latent,            # -> calibration.f_micro
    calibration_y=y_train,                # -> calibration.pd_ppm
    reasons=REASONS,                      # -> reasons{}
    missing_policy="baseline",            # -> features.missing_policy
)
 
save_artifact(artifact, "decision.json")
 
print("schema version ", artifact["schema_version"])
print("hash           ", artifact["artifact_hash"])
print("features       ", len(artifact["features"]["names"]))
print("trees          ", len(artifact["model"]["trees"]))
print("bands          ", len(artifact["bands"]["labels"]))
print("missing policy ", artifact["features"]["missing_policy"])
 
# What the compilation cost, on this data. Retention is a property of a
# dataset, a teacher and a depth budget — not of the library — so it is
# reported here rather than quoted as a constant anywhere.
gini = lambda y, s: 2 * roc_auc_score(y, s) - 1  # noqa: E731
 
compiled = load_artifact("decision.json")
scores = [decide(compiled, row.tolist(), explain=False)["latent_int"] for row in X_test]
 
teacher_gini = gini(y_test, teacher_test_scores)
artifact_gini = gini(y_test, scores)
 
print()
print(f"teacher gini    {teacher_gini:.3f}")
print(f"artifact gini   {artifact_gini:.3f}")
print(f"retained        {100 * artifact_gini / teacher_gini:.1f}%")