unable to use convert_sklearn for xgboost calibratedclassifier #1173
Description
met following error when try to use convert_sklearn
to convert calibrated xgboost classifer to onnx_model with skl2onnx
<class 'xgboost.sklearn.XGBClassifier'>
File <command-3767976754122279>, line 51
49 # Convert extracted XGBoost model to ONNX
50 initial_type = [("float_input", FloatTensorType([None, X_train.shape[1]]))]
---> 51 onnx_model = convert_sklearn(calibrated_xgb, initial_types=initial_type)
53 # Save ONNX model
54 onnx_filename = "xgb_extracted_model.onnx"
File /local_disk0/.ephemeral_nfs/envs/pythonEnv-7ca10a6f-be3f-434e-afeb-993e59ebdc60/lib/python3.10/site-packages/skl2onnx/convert.py:210, in convert_sklearn(model, name, initial_types, doc_string, target_opset, custom_conversion_functions, custom_shape_calculators, custom_parsers, options, intermediate, white_op, black_op, final_types, dtype, naming, model_optim, verbose)
208 if verbose >= 1:
209 print("[convert_sklearn] convert_topology")
--> 210 onnx_model = convert_topology(
211 topology,
212 name,
213 doc_string,
214 target_opset,
215 options=options,
216 remove_identity=model_optim and not intermediate,
217 verbose=verbose,
218 )
Also after updated the code toonnx_model = convert_sklearn(xgb_model_extracted, initial_types=initial_type) with xgb_model_extracted = calibrated_xgb.base_estimator still has following issue.
MissingShapeCalculator: Unable to find a shape calculator for type '<class 'xgboost.sklearn.XGBClassifier'>'.
It usually means the pipeline being converted contains a
transformer or a predictor with no corresponding converter
implemented in sklearn-onnx. If the converted is implemented
in another library, you need to register
the converted so that it can be used by sklearn-onnx (function
update_registered_converter). If the model is not yet covered
by sklearn-onnx, you may raise an issue to
https://github.com/onnx/sklearn-onnx/issues
to get the converter implemented or even contribute to the
project. If the model is a custom model, a new converter must
be implemented. Examples can be found in the gallery.
File <command-3767976754122279>, line 54
Is this a bug? and how can I best transform the calibratedclassifer xgboost object to the onnx format? ty!
below is the example code to regenerate the error:
%pip install onnx skl2onnx onnxruntime
import numpy as np
import xgboost as xgb
import joblib
import onnx
import skl2onnx
import onnxruntime as ort
from skl2onnx import convert_sklearn
from skl2onnx.common.data_types import FloatTensorType
from sklearn.calibration import CalibratedClassifierCV
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score, log_loss
from tensorflow.keras.datasets import mnist
# Load Keras dataset (example: MNIST)
(X_train, y_train), (X_test, y_test) = mnist.load_data()
# Preprocess data: Flatten images and normalize
X_train = X_train.reshape(X_train.shape[0], -1) / 255.0
X_test = X_test.reshape(X_test.shape[0], -1) / 255.0
# Further split training data for calibration
X_train, X_calib, y_train, y_calib = train_test_split(X_train, y_train, test_size=0.2, random_state=42)
# Define XGBoost model
xgb_model = xgb.XGBClassifier(
objective="multi:softmax", # Multi-class classification
num_class=10, # 10 classes for MNIST
use_label_encoder=False,
eval_metric="mlogloss"
)
# Train XGBoost model
xgb_model.fit(X_train, y_train)
# Apply sigmoid calibration
calibrated_xgb = CalibratedClassifierCV(xgb_model, method="sigmoid", cv="prefit")
calibrated_xgb.fit(X_calib, y_calib)
# Save calibrated model to joblib format
joblib.dump(calibrated_xgb, "calibrated_xgb_model.joblib")
# Load the joblib model
calibrated_xgb = joblib.load("calibrated_xgb_model.joblib")
# Convert class labels to strings
calibrated_xgb.classes_ = calibrated_xgb.classes_.astype(str)
xgb_model_extracted = calibrated_xgb.base_estimator
# Convert extracted XGBoost model to ONNX
initial_type = [("float_input", FloatTensorType([None, X_train.shape[1]]))]
onnx_model = convert_sklearn(xgb_model_extracted, initial_types=initial_type)
#onnx_model = convert_sklearn(calibrated_xgb, initial_types=initial_type)
# Save ONNX model
onnx_filename = "xgb_extracted_model.onnx"
with open(onnx_filename, "wb") as f:
f.write(onnx_model.SerializeToString())
print(f"Extracted XGBoost Model successfully converted to ONNX and saved as {onnx_filename}")
# Load ONNX model and make raw predictions
ort_session = ort.InferenceSession(onnx_filename)
onnx_inputs = {ort_session.get_inputs()[0].name: X_test.astype(np.float32)}
onnx_raw_prob = ort_session.run(None, onnx_inputs)[0] # Probabilities before calibration
# Apply sigmoid calibration manually using the trained calibration model
calibrated_prob = calibrated_xgb.calibrators_[0].predict_proba(onnx_raw_prob) # Apply calibration layer
onnx_pred = np.argmax(calibrated_prob, axis=1) # Convert to class labels
# Evaluate ONNX model after applying calibration
accuracy_onnx_calibrated = accuracy_score(y_test, onnx_pred)
logloss_onnx_calibrated = log_loss(y_test, calibrated_prob)
print(f"Calibrated XGBoost Accuracy (ONNX, after calibration): {accuracy_onnx_calibrated:.4f}")
print(f"Calibrated XGBoost Log Loss (ONNX, after calibration): {logloss_onnx_calibrated:.4f}")
# Compare probability distributions
prob_diff = np.mean(np.abs(y_prob_joblib - calibrated_prob))
print(f"Average absolute probability difference between Joblib and ONNX-calibrated models: {prob_diff:.6f}")
I also tried to_onnx, didn't work as well...AttributeError: 'numpy.uint8' object has no attribute 'encode' File <command-3767976754122452>, line 47 45 # Convert to ONNX format 46 initial_type = [("float_input", FloatTensorType([None, X_train.shape[1]]))] ---> 47 onnx_model = to_onnx(calibrated_xgb, initial_types=initial_type) 49 # Save ONNX model 50 onnx_filename = "calibrated_xgb_model.onnx"