How to see details of NAS models

🤝 Community Help
1
  • Project Type: Object Detection
  • Operating System & Browser: Windows; Google Chrome
  • Project Universe Link or Workspace/Project ID: burrow-id
  • Do you grant Roboflow Support permission to access your Workspace for troubleshooting? (Yes/No): Yes

Hello, I am using the NAS available in Roboflow Core to find the best model for detecting fiddler crab burrow openings (about 1-2 cm in diameter) in drone imagery taken from about 12 feet away (in a low-contrast environment). This is for university-associated research. I have read the Robinson et al. (2026) paper on NAS, but I still have some questions about how to interpret it.

  1. If I like a particular NAS-derived model, how can I tell which model architecture the run uses in Roboflow so I can report it? Or is it more of a black box that prevents interpretation at that level?
  2. Additionally, is it possible to download the model weights for a particular model that I like from the NAS? How can I share those weights with others?
  3. Finally, is there a simple way to look at the accuracy of a particular model from the NAS at the image-level? I want to look at how the accuracy changes depending on the type of land cover the burrows are detected over. I already have a spreadsheet of the image names and their land cover types; I just need to find their individual accuracies. Thanks!
2

Hi @Scarpa ,

Great questions here!

  1. NAS all runs on the same architecture, which also happens to be the architecture that RF-DETR was built on. You can learn more technical detail in the RF-DETR paper or here: RF-DETR: A SOTA Real-Time Object Detection Model but in short it is built on a DINOv2 backbone and is an attention based vision transformer.
  2. NAS model weights are unavailable for download (list of model capabilities here: https://docs.roboflow.com/deploy/supported-models) . The best way to share the model with others would be to invite them to your workspace!
  3. You are absolutely able to test individual models that NAS trains and compare them against each other. Clicking on the model points will also show you general information about the model (F1,Accuracy,Precision,Recall)

Let me know if that answers your questions well!

Best,

Patrick