cosine_scheduler.rf-detr-nano

RF-DETR with cosine learning rate scheduling and optimized hyperparameters

We fine-tuned RF-DETR using a cosine learning rate scheduler to provide smoother optimization and improved late-stage refinement. The model demonstrates stable convergence with balanced reductions across classification, bounding box regression, and GIoU losses. EMA weights consistently outperform raw parameters, confirming reduced variance during training. The final model achieves 0.555 mAP@50:95, with strong performance on well-represented vehicle classes such as two-wheelers, buses, and trucks. Remaining challenges are concentrated in visually ambiguous, low-frequency classes such as minibuses and vans.

Their is almost ~similar or no identical differences with step vs cosine lr_scheduling, it might be also because models are only finetuned for 2 epochs due to computational cost and disconnection in T4 GPU.

Future goals

Retrain the models for around ~50 epochs one with cosine and one with step lr_scheduling to see the varying difference in smooth training and better convergence.