Implementation:Roboflow Rf detr RFDETR Init Finetuned

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

Overview

Concrete tool for loading fine-tuned RF-DETR checkpoints with custom class names for domain-specific inference.

Description

Uses the same RFDETR.__init__ pathway as standard initialization but with pretrain_weights pointing to a fine-tuned checkpoint. The Model.__init__ method detects the checkpoint's class count from the class_embed.bias shape, reinitializes the detection head if needed, and loads custom class names from checkpoint['args'].class_names. The model's class_names property then returns the fine-tuned class names instead of COCO classes.

Usage

Instantiate any RFDETR size variant with the pretrain_weights parameter set to the path of a fine-tuned checkpoint file.

Code Reference

Source Location

• Repository: rf-detr
• File: rfdetr/detr.py
• Lines: L63-76 (RFDETR.__init__), L252-264 (class_names property)
• File: rfdetr/main.py
• Lines: L94-166 (Model.__init__ - weight loading with class mismatch handling)

Signature

# Same API as standard initialization
model = RFDETRBase(pretrain_weights="output/checkpoint_best_total.pth")

# Access custom class names
model.class_names  # -> {1: "cat", 2: "dog", ...}

Import

from rfdetr import RFDETRBase  # or matching size variant used during training

I/O Contract

Inputs

Name	Type	Required	Description
pretrain_weights	str	Yes	Path to fine-tuned checkpoint (e.g. "output/checkpoint_best_total.pth")

Outputs

Name	Type	Description
model	RFDETR	Model loaded with fine-tuned weights and custom class names
model.class_names	Dict[int, str]	Custom class name mapping from checkpoint

Usage Examples

Load and Run Fine-tuned Model

from rfdetr import RFDETRBase

# Load fine-tuned model
model = RFDETRBase(pretrain_weights="output/checkpoint_best_total.pth")

# Verify custom classes loaded
print(model.class_names)  # {1: "hardhat", 2: "vest", 3: "person"}

# Run inference
detections = model.predict("construction_site.jpg", threshold=0.5)

# Use custom class names for labels
labels = [
    f"{model.class_names[cid]} {conf:.2f}"
    for cid, conf in zip(detections.class_id, detections.confidence)
]

Related Pages

Implements Principle

• Principle:Roboflow_Rf_detr_Finetuned_Model_Loading

Requires Environment

• Environment:Roboflow_Rf_detr_Python_GPU_Environment