guides/model-training-tips/

[giscus[bot]]

guides/model-training-tips/

Find best practices, optimization strategies, and troubleshooting advice for training computer vision models. Improve your model training efficiency and accuracy.

https://docs.ultralytics.com/guides/model-training-tips/

[Taufiq-Ahmed]

Hi I have a question. I am using Yolov8s for trained a dataset that has four class. fire, smoke, person, animal. all the images has areal view to detect forest fire . In my dataset I have forest fire images from drone so after training it detect fire smoke quite accurately. however, I do not have any dataset that has person or animal in forest fire. so I have to collect some other dataset from roboflow that has person and animal and the n I have merged them with my own dataset ( have changed the label of animal and person accordingly so 0 means smoke, 1 means fire, 2 means person and 3 means animal and also updated the data.yaml file) . this time the problem is it could not detect the animal and person properly some times it can detect them. The main reasons seems to me is that the drone is flying in a high altitude so it does not detect person and animal properly because in the dataset of person and animal the view of the drone is not that much high. is there any technique to solve this problem and detect all the objects as many as possible?

[glenn-jocher]

@Taufiq-Ahmed it sounds like the altitude difference between your training and test datasets is causing the detection issues. To address this, consider using data augmentation techniques to simulate higher altitude views in your training data. You can also try multiscale training to help the model generalize better across different scales. Additionally, fine-tuning the model with a mix of high-altitude and low-altitude images might improve its performance. For more detailed strategies, you can refer to our model training tips.

[Imran-ml]

First of all its great to have another powerful model and congratulations. I have fine tuned and trained yolo11x.pt for my custom objects but I'm facing some issues. My images data resolution is around 6000 x 2000 but while trained the model I set the imgsz=(1280, 640). It work well when I adjust the video stream by zoom in-out and detect all my 60 objects well (objects are small) but when there is some zoom or squeezed then it completely dont detect any object or few only.

Please let me know if I'm missing something or I have to do something else?
I have labelled the data on my original images so resizing making any issues? any ratio issues? Please guide me properly so I can use this power yolo11 model effectively.

[glenn-jocher]

@Imran-ml it sounds like the issue might be related to image resizing affecting object detection. Consider enabling multiscale training by adjusting the scale parameter to help the model generalize better across different image sizes. This can improve detection performance when objects appear at varying scales. For more detailed guidance, you can refer to our model training tips at https://docs.ultralytics.com/guides/model-training-tips/.

[VeeranjaneyuluToka]

Hi, I just noticed a thing, looks like the optimizer is getting chosen based on no.of iterations, would you mind telling me the reason for the same? I mean how does iterations matter when we choose optimizer?

[TSHSLU]

What is the minimum Dataset Size required if i use a pretrained model for custom projects. i barely find any sources online which recommend a minimum dataset size.

[glenn-jocher]

There isn’t a fixed “minimum” dataset size when fine-tuning a pretrained model; if your domain is close to the pretrained data and labels are clean, you can often start getting useful signal with ~50–200 images per class (or a few hundred labeled object instances), but for reliably robust results across different lighting/viewpoints/backgrounds you’ll usually want 500–2000+ images per class with good diversity—one practical way to find your true minimum is to train on a small pilot subset, check validation failures, then add data specifically where the model breaks (the fraction workflow in the Model training tips guide helps iterate quickly).

from ultralytics import YOLO

model = YOLO("yolo26n.pt")
model.train(data="data.yaml", epochs=100, fraction=0.1)