Implementation:Roboflow Rf detr RFDETR Train Config

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

Overview

Concrete tool for configuring and launching RF-DETR fine-tuning provided by the RF-DETR library.

Description

RFDETR.train() accepts keyword arguments, validates them through TrainConfig (a Pydantic model), loads class names from the dataset, reinitializes the detection head for the correct class count, registers metric callbacks, and delegates to Model.train(). The TrainConfig class defines all training hyperparameters with sensible defaults for fine-tuning.

Usage

Call model.train(dataset_dir=...) on any RFDETR size variant to start fine-tuning.

Code Reference

Source Location

Repository: rf-detr
File: rfdetr/detr.py
Lines: L89-94 (RFDETR.train), L167-238 (train_from_config)
File: rfdetr/config.py
Lines: L242-301 (TrainConfig, SegmentationTrainConfig)

Signature

class RFDETR:
    def train(self, **kwargs) -> None:
        """
        Train an RF-DETR model.
        All training parameters are passed as keyword arguments.
        """

class TrainConfig(BaseModel):
    lr: float = 1e-4
    lr_encoder: float = 1.5e-4
    batch_size: int = 4
    grad_accum_steps: int = 4
    epochs: int = 100
    dataset_dir: str  # Required
    dataset_file: Literal["coco", "o365", "roboflow", "yolo"] = "roboflow"
    output_dir: str = "output"
    use_ema: bool = True
    ema_decay: float = 0.993
    early_stopping: bool = False
    early_stopping_patience: int = 10
    multi_scale: bool = True
    tensorboard: bool = True
    wandb: bool = False
    weight_decay: float = 1e-4
    warmup_epochs: float = 0.0
    ...

Import

from rfdetr import RFDETRBase
from rfdetr.config import TrainConfig

I/O Contract

Inputs

Name	Type	Required	Description
dataset_dir	str	Yes	Path to dataset root directory
epochs	int	No	Number of training epochs (default: 100)
batch_size	int	No	Batch size per device (default: 4)
lr	float	No	Base learning rate (default: 1e-4)
lr_encoder	float	No	Backbone learning rate (default: 1.5e-4)
output_dir	str	No	Directory for checkpoints and logs (default: "output")
use_ema	bool	No	Enable EMA model tracking (default: True)

Outputs

Name	Type	Description
Trained model	RFDETR	Model with updated weights (best EMA or regular)
Checkpoints	Files	checkpoint_best_total.pth, checkpoint_best_regular.pth, checkpoint_best_ema.pth
results.json	File	Per-class mAP, precision, recall, F1 scores

Usage Examples

Basic Fine-tuning

from rfdetr import RFDETRBase

model = RFDETRBase()
model.train(
    dataset_dir="/path/to/dataset",
    epochs=50,
    batch_size=8,
    output_dir="./training_output",
)

Advanced Configuration

from rfdetr import RFDETRMedium

model = RFDETRMedium()
model.train(
    dataset_dir="/path/to/dataset",
    epochs=100,
    batch_size=4,
    grad_accum_steps=8,
    lr=5e-5,
    lr_encoder=7.5e-5,
    use_ema=True,
    early_stopping=True,
    early_stopping_patience=15,
    wandb=True,
    project="my-detection-project",
)

Related Pages

Implements Principle

Principle:Roboflow_Rf_detr_Training_Configuration

Requires Environment

Environment:Roboflow_Rf_detr_Python_GPU_Environment

Uses Heuristic

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment