Datasets:
Are the training hyper‑parameters for YOLOv8n available?
Hi, I am working on UAV‑based water‑target detection task.
Could you please share the training hyper‑parameters you used for training YOLOv8n on this dataset?
For example: imgsz, batch size, optimizer, learning rate, mosaic augmentation setting, epochs and other related settings.
Thank you very much!
Hi @usca01
Yes absolutely, there would be no point in just releasing the results and data without releasing the code.
Please find the relevant parameters that you are looking for in this config here https://github.com/dronefreak/DetectionBench/blob/main/configs/dataset/seadronessee.yaml
Note that its hard coded to my local PC, you would need to download the SeaDronesSee dataset from hugging face and place it somewhere on your local system and use the config file for YOLO https://huggingface.co/datasets/dronefreak/SeaDronesSee/blob/main/data/detectionbench_data.yaml as your YOLO config (Ultralytics package expects a config like this)
I call it DetectionBench, a small project that I created to benchmark and unify the under-represented object detection datasets such as SeaDronesSee, project link is here https://github.com/dronefreak/DetectionBench
Launching a training is super easy, the documentation is short and compact, its something like
detectionbench-train dataset=seadronessee model.name=yolov8n
It should automatically create an experiments/ folder in your root repo and you should find your model under the same folder. Normally since yolov8n is a very small model, I think I used BS=32 and IMGSZ=640 (standard YOLO) and other standard hyperparams for YOLO.
Please let me know if this helps!