Implementation:Roboflow Rf detr Evaluate

Knowledge Sources	RF-DETR
Domains	Object_Detection, Evaluation
Last Updated	2026-02-08 15:00 GMT

Overview

Concrete tool for evaluating RF-DETR model performance using COCO metrics and extended precision/recall/F1 analysis.

Description

evaluate() runs the model over a validation DataLoader, computes losses, collects predictions via PostProcess, and passes them to CocoEvaluator for standard COCO metric computation. coco_extended_metrics() then analyzes the raw COCO evaluation data to produce per-class precision, recall, and F1 scores at the optimal confidence threshold.

Usage

Called automatically at the end of each training epoch. Also used for standalone evaluation when resuming from a checkpoint.

Code Reference

Source Location

• Repository: rf-detr
• File: rfdetr/engine.py
• Lines: L363-453 (evaluate), L239-361 (coco_extended_metrics)
• File: rfdetr/datasets/coco_eval.py
• Lines: L37-50 (CocoEvaluator.__init__)

Signature

def evaluate(
    model: torch.nn.Module,
    criterion: torch.nn.Module,
    postprocess: PostProcess,
    data_loader: DataLoader,
    base_ds: COCO,
    device: torch.device,
    args=None,
) -> Tuple[Dict, CocoEvaluator]:
    """
    Evaluate model on validation set.

    Returns:
        test_stats: Dict with coco_eval_bbox (12-element stats),
                    results_json (per-class metrics)
        coco_evaluator: CocoEvaluator instance
    """

def coco_extended_metrics(coco_eval: COCOeval) -> Dict:
    """
    Compute per-class precision/recall/F1 by sweeping confidence thresholds.

    Returns:
        Dict with class_map, map, precision, recall, f1_score
    """

Import

from rfdetr.engine import evaluate, coco_extended_metrics
from rfdetr.datasets.coco_eval import CocoEvaluator

I/O Contract

Inputs

Name	Type	Required	Description
model	torch.nn.Module	Yes	Model to evaluate
criterion	torch.nn.Module	Yes	Loss function (SetCriterion)
postprocess	PostProcess	Yes	Post-processing for converting outputs to detections
data_loader	DataLoader	Yes	Validation data loader
base_ds	COCO	Yes	Ground truth COCO API object
args.fp16_eval	bool	No	Half-precision evaluation (default: False)
args.eval_max_dets	int	No	Max detections for evaluation (default: 500)

Outputs

Name	Type	Description
test_stats	Dict	Contains coco_eval_bbox (12 COCO stats), results_json (per-class metrics)
coco_evaluator	CocoEvaluator	Full evaluation results object

Usage Examples

Evaluation During Training

# Called automatically within Model.train() each epoch:
test_stats, coco_evaluator = evaluate(
    model, criterion, postprocess, data_loader_val, base_ds, device, args=args
)

# Access mAP
map_50_95 = test_stats["coco_eval_bbox"][0]
map_50 = test_stats["coco_eval_bbox"][1]

# Access per-class metrics
per_class = test_stats["results_json"]["class_map"]

Related Pages

Implements Principle

• Principle:Roboflow_Rf_detr_COCO_Evaluation

Requires Environment

• Environment:Roboflow_Rf_detr_Python_GPU_Environment