Implementation:Roboflow Rf detr RFDETR Size Variants

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

Overview

Concrete tool for selecting object detection model size variants provided by the RF-DETR library.

Description

RF-DETR provides five detection model size classes (RFDETRNano, RFDETRSmall, RFDETRBase, RFDETRMedium, RFDETRLarge) and seven segmentation variants. Each subclass overrides get_model_config() to return a size-specific ModelConfig with architecture parameters. All inherit from the base RFDETR class and share the same API surface.

Usage

Import the appropriate size variant class when instantiating a model for inference or training. Choose based on your accuracy vs. speed requirements.

Code Reference

Source Location

• Repository: rf-detr
• File: rfdetr/detr.py
• Lines: L462-567 (size variant classes)
• Config File: rfdetr/config.py L57-141 (config classes)

Signature

class RFDETRBase(RFDETR):
    """Train an RF-DETR Base model (29M parameters)."""
    size = "rfdetr-base"
    def get_model_config(self, **kwargs) -> RFDETRBaseConfig: ...

class RFDETRNano(RFDETR):
    """Train an RF-DETR Nano model."""
    size = "rfdetr-nano"
    def get_model_config(self, **kwargs) -> RFDETRNanoConfig: ...

class RFDETRSmall(RFDETR):
    """Train an RF-DETR Small model."""
    size = "rfdetr-small"
    def get_model_config(self, **kwargs) -> RFDETRSmallConfig: ...

class RFDETRMedium(RFDETR):
    """Train an RF-DETR Medium model."""
    size = "rfdetr-medium"
    def get_model_config(self, **kwargs) -> RFDETRMediumConfig: ...

class RFDETRLarge(RFDETR):
    """Train an RF-DETR Large model."""
    size = "rfdetr-large"
    def get_model_config(self, **kwargs) -> RFDETRLargeConfig: ...

Import

from rfdetr import RFDETRBase, RFDETRNano, RFDETRSmall, RFDETRMedium, RFDETRLarge

I/O Contract

Inputs

Name	Type	Required	Description
pretrain_weights	Optional[str]	No	Path or hosted key for pretrained weights (e.g. "rf-detr-base.pth")
resolution	int	No	Input resolution (default varies by size: Nano=384, Small=512, Base=560, Medium=576, Large=704)
num_classes	int	No	Number of object classes (default: 90 for COCO)
device	str	No	Compute device: "cuda", "cpu", or "mps"

Outputs

Name	Type	Description
instance	RFDETR subclass	Initialized model with appropriate ModelConfig, ready for predict/train/export

Usage Examples

Select Base Model

from rfdetr import RFDETRBase

# Default COCO-pretrained Base model
model = RFDETRBase()

# Custom weights and resolution
model = RFDETRBase(pretrain_weights="path/to/weights.pth", resolution=640)

Compare Model Sizes

from rfdetr import RFDETRNano, RFDETRBase, RFDETRLarge

# Fast inference on edge devices
nano_model = RFDETRNano()  # resolution=384, 2 decoder layers

# Balanced performance
base_model = RFDETRBase()  # resolution=560, 3 decoder layers

# Maximum accuracy
large_model = RFDETRLarge()  # resolution=704, 4 decoder layers

Model Configurations

Size	Resolution	Dec Layers	Patch Size	Num Windows	Pretrained Weights
Nano	384	2	16	2	rf-detr-nano.pth
Small	512	3	16	2	rf-detr-small.pth
Base	560	3	14	4	rf-detr-base.pth
Medium	576	4	16	2	rf-detr-medium.pth
Large	704	4	16	2	rf-detr-large-2026.pth

Related Pages

Implements Principle

• Principle:Roboflow_Rf_detr_Model_Size_Selection

Requires Environment

• Environment:Roboflow_Rf_detr_Python_GPU_Environment