Change the training dataset

[323-w]

Thank you for your work. I have a few questions I'd like to ask you.

1. May I ask if this weight file is officially provided? https://huggingface.co/gatilin/YOLO-Master-ckpts-v0/tree/main/YOLO-Master-EsMoE-M
2. I want to train my own detection dataset with only one class. May I ask if I only need to modify the following part of the code: /ultralytics/cfg/models/master/v0_1/det/yolo-master-m.yaml, changing nc: 80 # number of classes to nc:1? I also use my own configuration file mydatasets.yaml. Can I directly use the command yolo detect train data=mydatasets.yaml pretrained=yolo-master-m.pt model=/ultralytics/cfg/models/master/v0_1/det/yolo-master-m.yaml epochs=100 to train? When I train with this command, I find that the training shows relevant loss, but all the test metrics are 0. Could you please tell me what might be the reason for this?

[323-w]

I have another question for you. If I'm doing small object detection, what key parameter values ​​would you recommend?

[isLinXu]

Thanks for your interest in YOLO-Master! Let me answer your questions one by one.

1. Are the weights official?

Yes. The official pretrained weights are provided in two places:

• Hugging Face: gatilin/YOLO-Master-ckpts-v0
• GitHub Releases: YOLO-Master-v26.02 — see the Assets section at the bottom of the release page, which contains the weight files (incl. YOLO-Master-EsMoE-M, deployment formats like .engine, etc.) with sha256 checksums.

YOLO-Master-EsMoE-M matches the yolo-master-m (ES-MoE) config, so it's the right pretrained= for yolo-master-m.yaml.

2. Training on your own single-class dataset

The key thing for one-class detection is to add single_cls=True:

yolo detect train data=mydatasets.yaml model=/ultralytics/cfg/models/master/v0_1/det/yolo-master-m.yaml pretrained=yolo-master-m.pt epochs=100 single_cls=True

With single_cls=True, Ultralytics treats all labels as a single class regardless of their original class index. This is the recommended way to do one-class training, and it also avoids the most common cause of your "metrics = 0" problem (see below). You don't strictly need to hand-edit nc in the model YAML — it's overridden from the dataset config / single_cls automatically.

Why loss drops but P/R/mAP are all 0

The loss is computed against your GT labels, so it falling means training "works"; metrics being 0 means predictions never match GT — almost always a label/data issue, not a model issue. Most likely causes, in order:

1. Class-index mismatch (most common). Without single_cls, if nc=1 but your label .txt files still contain a non-zero class id (e.g. the original COCO id), eval matches nothing → 0. Setting single_cls=True collapses everything to class 0 and fixes this. Otherwise, make sure every label id is 0.
2. Val split has no matching labels. Confirm val: in mydatasets.yaml points to images whose label files exist.
3. Wrong label format. Must be class x_center y_center width height, all normalized to [0,1].
4. nc vs len(names) mismatch in the dataset YAML.

Quick diagnosis

1. Add single_cls=True and retrain a few epochs.
2. Check runs/detect/train/train_batch0.jpg — are the GT boxes correct?
3. Validate with a low threshold: yolo detect val model=runs/detect/train/weights/best.pt data=mydatasets.yaml conf=0.001 single_cls=True.

3. Recommended parameters for small object detection

Small objects are exactly where YOLO-Master's Sparse SAHI module helps most.

Use Sparse SAHI (our built-in adaptive slicing). Instead of feeding the full image at low effective resolution, Sparse SAHI adaptively slices the image so small objects appear larger relative to the receptive field, with ~3–5× speedup over naive SAHI. This is the single most impactful thing for small-object recall. The slicing inference assets and usage are in the release: YOLO-Master-v26.02 (see the Assets section).

Training-side parameters:

• imgsz — increase it. 640 → 1024/1280 is usually the biggest single win for small objects (at the cost of memory/speed).
• scale — lower the scale jitter (e.g. 0.5 → 0.2) so objects aren't shrunk too aggressively during training.
• mosaic=1.0 + close_mosaic=10 — keep mosaic for dense small-object scenes, disabling it only in the final epochs; mild copy_paste/mixup can boost rare small instances.
• Stride/heads — small objects are detected on the high-resolution P3 (stride 8) head; for extremely tiny objects, adding a P2 (stride 4) head helps.

Inference-side parameters:

• conf — lower it (e.g. 0.001–0.1 for eval) since small objects produce lower-confidence boxes.
• iou (NMS) — for dense small objects, our Cluster-Weighted NMS avoids over-suppression in crowded scenes.
• SAHI slice size / overlap — smaller slices + ~20% overlap recover objects on tile boundaries.

TL;DR: imgsz=1280, enable Sparse SAHI, lower conf at eval, and rely on Cluster-Weighted NMS for dense scenes. Start there and tune slice size/overlap to your object scale.

If anything is still unclear, feel free to share your mydatasets.yaml, a sample label .txt, and the dataset-scan log header, and we'll dig deeper. 🙌