Custom Yolo detection model training

[hardhik-08]

I am working with the YOLOv8m model and would like to extend it with additional custom object classes.

Specifically, I want the model to:

• Retain all the original 80 COCO classes (e.g., person, car, toothbrush, etc.)
• Add new custom classes (e.g., door, window, etc.), continuing the class indexing (e.g., 80: toothbrush, 81: door, 82: window)

I don't want to train the model from scratch with coco + custom object datasets
Any guidance or examples would be greatly appreciated.

[UltralyticsAssistant]

👋 Hello @hardhik-08, thank you for your interest in Ultralytics 🚀! This is an automated response to help get your discussion moving, and an Ultralytics engineer will also assist soon. We recommend a visit to the Docs for new users where you can find many Python and CLI usage examples and where many of the most common questions may already be answered.
Since this is a custom training ❓ question, please provide as much information as possible, including dataset image examples, your dataset YAML, training logs, the exact model you are starting from, and verify you are following our Tips for Best Training Results.
If this turns out to be a 🐛 bug or unexpected behavior, please also provide a minimum reproducible example to help us debug it.
Join the Ultralytics community where it suits you best. For real-time chat, head to Discord 🎧. Prefer in-depth discussions? Check out Discourse. Or dive into threads on our Subreddit to share knowledge with the community.

Upgrade

Upgrade to the latest ultralytics package including all requirements in a Python>=3.8 environment with PyTorch>=1.8 to verify your issue is not already resolved in the latest version:

pip install -U ultralytics

Environments

YOLO may be run in any of the following up-to-date verified environments (with all dependencies including CUDA/CUDNN, Python and PyTorch preinstalled):

• Notebooks with free GPU:
• Google Cloud Deep Learning VM. See GCP Quickstart Guide
• Amazon Deep Learning AMI. See AWS Quickstart Guide
• Docker Image. See Docker Quickstart Guide

Status

If this badge is green, all Ultralytics CI tests are currently passing. CI tests verify correct operation of all YOLO Modes and Tasks on macOS, Windows, and Ubuntu every 24 hours and on every commit.

[glenn-jocher]

Good question — with standard yolov8m.pt, you can fine-tune from pretrained weights, but you can’t just append classes 81+ and keep the original COCO behavior without additional old-class supervision; otherwise the model will usually forget some COCO classes. If you want custom categories without retraining, YOLO-World is the better fit; if you need one fixed COCO+custom detector, you’ll need an incremental/distillation training setup, and the custom trainer guide is the right starting point.

[Y-T-G]

I don't want to train the model from scratch with coco + custom object datasets

That's the only way to stop forgetting. You can try YOLOE with prompts if you want more classes than MSCOCO:

https://docs.ultralytics.com/models/yoloe/#text-prompt

[hardhik-08]

I don't want to train the model from scratch with coco + custom object datasets

That's the only way to stop forgetting. You can try YOLOE with prompts if you want more classes than MSCOCO:

https://docs.ultralytics.com/models/yoloe/#text-prompt

I did try Yoloe and Yoloworld, but they require more resources (AIPU). I have also tried your way of extending detection head, but if I need to include additional objects I need to re-train the head again.

[Y-T-G]

If you export YOLOE, it will be same speed as normal YOLO models. You can set the prompts before export and then export. The prompts get fused into the model, so during inference it's like a normal YOLO model.