Implementation:Rapidsai Cuml Confusion Matrix

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

Overview

Computes a GPU-accelerated confusion matrix to evaluate the accuracy of a classification model.

Description

The confusion_matrix function builds a square confusion matrix C of shape (n_classes, n_classes) where entry C[i, j] is the number of observations known to be in class i and predicted to be in class j. It uses CuPy sparse COO matrices (cupyx.scipy.sparse.coo_matrix) for efficient construction, then converts to a dense array.

Key features:

• Label handling -- When labels is None, labels are inferred from the sorted union of y_true and y_pred. A custom labels array can be passed to reorder or subset the matrix.
• Sample weights -- An optional sample_weight array replaces the default unit counts.
• Normalization -- The normalize parameter supports "true" (row-normalize), "pred" (column-normalize), "all" (total-normalize), or None (raw counts).
• Label monotonization -- Internally uses cuml.prims.label.make_monotonic to map arbitrary integer labels to contiguous indices.

Usage

Use this function after classification to understand which classes are being confused. The confusion matrix is a foundation for computing precision, recall, F1-score, and other per-class metrics. It is commonly visualized as a heatmap.

Code Reference

Source Location

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

Signature

def confusion_matrix(
    y_true,
    y_pred,
    labels=None,
    sample_weight=None,
    normalize=None,
    convert_dtype=False,
) -> CumlArray

Import

from cuml.metrics import confusion_matrix

I/O Contract

Inputs

Name	Type	Required	Description
y_true	array-like (device or host) of shape (n_samples,)	Yes	Ground truth (correct) target labels. Must be integer type (int32 or int64).
y_pred	array-like (device or host) of shape (n_samples,)	Yes	Predicted target labels. Must be integer type (int32 or int64).
labels	array-like of shape (n_classes,)	No	List of label indices for the matrix. If None, inferred from the union of y_true and y_pred.
sample_weight	array-like of shape (n_samples,)	No	Per-sample weights. Default is uniform (1.0).
normalize	str in ['true', 'pred', 'all'] or None	No	Normalization mode. "true" normalizes over rows, "pred" over columns, "all" over the entire matrix. Default None (raw counts).
convert_dtype	bool	No	When True, automatically convert inputs to int32. Default False.

Outputs

Name	Type	Description
C	CumlArray of shape (n_classes, n_classes)	The confusion matrix. Entry C[i, j] is the count (or normalized proportion) of samples with true label i predicted as label j.

Usage Examples

import cupy as cp
from cuml.metrics import confusion_matrix

y_true = cp.array([2, 0, 2, 2, 0, 1])
y_pred = cp.array([0, 0, 2, 2, 0, 2])

# Basic confusion matrix
cm = confusion_matrix(y_true, y_pred)
print(cm)
# [[2, 0, 0],
#  [0, 0, 1],
#  [1, 0, 2]]

# Normalized over true labels (rows)
cm_norm = confusion_matrix(y_true, y_pred, normalize='true')
print(cm_norm)

Related Pages

• Environment:Rapidsai_Cuml_CUDA_GPU
• Environment:Rapidsai_Cuml_Python_RAPIDS_Stack