Loading pretrain weights give bad result

[Sunirban-Sarkar]

Search before asking

• I have searched the RF-DETR issues and found no similar bug report.

Bug

After training model saved two checkpoints files regular and ema. But when I again load those weights back for inference the results are bad or different. Can someone help?

Environment

RF-DETR 1.1.0
OS Ubuntu 22.04 LTS
Python 3.10.18
Pytorch 2.7.1+cu128
Cuda 12.8
GPU RTX A5000

Minimal Reproducible Example

Just load weight and try inference

Additional

No response

Are you willing to submit a PR?

• Yes, I'd like to help by submitting a PR!

[SkalskiP]

Hi @Sunirban-Sarkar 👋🏻 Thanks so much for your interest in RF-DETR. I recommend using the latest version, rfdetr==1.2.1. I’ve just tested training the RF-DETR-Medium model on a custom dataset, benchmarking it, then loading the saved checkpoint from disk and benchmarking it again. In both cases, I got exactly the same results.

For more details, feel free to check out this Google Colab.