Roboflow Leaderboard: Optimal Validation parameters

[LinasKo]

💡 Your Question

Hello from Roboflow!

We're building a model leaderboard comparing the most popular models out there on a single page across multiple metrics. We're using COCO 2017 validation set, and aim to reproduce optimal model results, as found in model author's benchmarks.

I've had a bit of trouble finding which parameters (iou, conf, max_det) should best be used to validate Yolo-NAS. (PR #27)

The latest parameter set we landed on was the following, yet the results are slightly worse than I'd expect.

model_params = dict(
    conf=0.01,
    iou=0.7,
    nms_top_k=1000,
    max_predictions=300,
)

Could you please provide a set of model parameters that were used to benchmark the model originally?

Versions

No response