Support training with multi-scale

[dFuong]

Does Yolo-nas support training with multi-scale ?

[dagshub[bot]]

Join the discussion on DagsHub!

[shaydeci]

Training in SG generally supports multi-scale. For YoloNAS, you will need to pass DetectionMultiscalePrePredictionCallback through the "pre_prediction_callback" training parameter. See its docs here.

[BloodAxe]

To add to @shaydeci answer, you can use collate function from PPYolo-E training recipe that does multi-scale training on batch level - https://github.com/Deci-AI/super-gradients/blob/d775481084d8e8fc57324ebd34b33a0644de2817/src/super_gradients/recipes/dataset_params/coco_detection_ppyoloe_dataset_params.yaml#L57 What you need to do is replace collate_fn in the dataset config YAML for YOLO-NAS.