Implementation:Rapidsai Cuml Classification Metrics

Knowledge Sources	Rapidsai_Cuml
Domains	Machine_Learning, Classification, Model_Evaluation
Last Updated	2026-02-08 12:00 GMT

Overview

Provides GPU-accelerated classification evaluation metrics including accuracy score and log loss (cross-entropy loss).

Description

The _classification.py module implements two core classification metrics on the GPU:

• accuracy_score -- Computes the fraction (or count) of correctly classified samples. Supports optional sample weights and a normalize flag that controls whether the result is a fraction or an absolute count. It handles both CuPy arrays and cuDF Series inputs, including categorical dtype alignment.

• log_loss -- Computes the negative log-likelihood (cross-entropy) loss, the standard loss function for logistic regression and neural networks. Supports binary and multiclass problems, probability clipping via an eps parameter, and optional sample weights with a normalize toggle.

An internal helper _input_to_cupy_or_cudf_series coerces heterogeneous inputs (cuDF Series, CuPy, NumPy, Numba) into a uniform 1-D representation, handling edge cases like string labels (via cuDF) and float16 (via CuPy).

Usage

Use these functions to evaluate classification model quality on GPU data. They serve as drop-in replacements for the corresponding scikit-learn metrics, accepting device arrays and host arrays alike. Typical usage is after model prediction to measure predictive accuracy or calibration quality.

Code Reference

Source Location

• Repository: Rapidsai_Cuml
• File: python/cuml/cuml/metrics/_classification.py

Signature

def accuracy_score(y_true, y_pred, *, sample_weight=None, normalize=True)

def log_loss(
    y_true, y_pred, eps=1e-15, normalize=True, sample_weight=None
) -> float

Import

from cuml.metrics import accuracy_score
from cuml.metrics import log_loss

I/O Contract

Inputs -- accuracy_score

Name	Type	Required	Description
y_true	array-like of shape (n_samples,)	Yes	Ground truth (correct) labels.
y_pred	array-like of shape (n_samples,)	Yes	Predicted labels, as returned by a classifier.
sample_weight	array-like of shape (n_samples,)	No	Per-sample weights. Default None.
normalize	bool	No	If True (default), return fraction of correct predictions; if False, return the count.

Inputs -- log_loss

Name	Type	Required	Description
y_true	array-like of shape (n_samples,)	Yes	True labels (integer-valued, non-negative).
y_pred	array-like of shape (n_samples, n_classes) or (n_samples,)	Yes	Predicted probabilities or decision values.
eps	float	No	Clipping bound for probabilities to avoid log(0). Default 1e-15.
normalize	bool	No	If True (default), return mean loss per sample; otherwise return total sum.
sample_weight	array-like of shape (n_samples,)	No	Per-sample weights. Default None.

Outputs

Name	Type	Description
score (accuracy_score)	float	Fraction of correctly classified samples, or count if normalize=False.
loss (log_loss)	float	The (mean or summed) cross-entropy loss.

Usage Examples

import cupy as cp
from cuml.metrics import accuracy_score, log_loss

# Accuracy score
y_true = cp.array([0, 1, 2, 3])
y_pred = cp.array([0, 2, 1, 3])
print(accuracy_score(y_true, y_pred))  # 0.5

# Log loss (binary classification with probability predictions)
y_true = cp.array([1, 0, 0, 1])
y_pred_proba = cp.array([[0.1, 0.9], [0.9, 0.1], [0.8, 0.2], [0.35, 0.65]])
print(log_loss(y_true, y_pred_proba))  # ~0.2162

Related Pages

• Environment:Rapidsai_Cuml_CUDA_GPU
• Environment:Rapidsai_Cuml_Python_RAPIDS_Stack