resume/retrain yolov11 with only 3 classes

[ardeal]

Hi,

I would like to resume training yolov11 model(yolo11m.pt) for only car, bicycle and person.
How should I do?

The purpose of resuming training is that I would like to train yolov11 with the initialization of yolov11m.pt model and add my new images(care/bicycle/person).

[UltralyticsAssistant]

👋 Hello @ardeal, thank you for your interest in Ultralytics 🚀! To help answer your question more effectively, I recommend providing some additional details such as the dataset structure, examples of your dataset labels (for car, bicycle, and person), and your training configuration. This will allow us to better understand your setup.

If this is a 🐛 Bug Report, please ensure you include a minimum reproducible example. This will significantly help in debugging the issue.

If this is a custom training ❓ Question, please provide logs or outputs from your training process and specify any adjustments you've made to the model or training hyperparameters. Also, verify that you're following best practices for training and data preparation to ensure quality results.

It’s a good idea to upgrade to the latest version of the ultralytics package, along with all its dependencies, to ensure compatibility and access to the latest features and fixes. You can do so using the appropriate package manager for your environment.

For development or training, confirmed working environments include notebooks with free GPU resources, preconfigured cloud environments such as GCP or AWS, and official Docker containers.

This is an automated response to guide you, but an Ultralytics engineer will review your question and assist you further soon 🚀!

[Y-T-G]

You can't do that. You can train on your dataset, but it's performance will get worse on other classes that are not in your dataset. And it will also get worse for the classes that are in your dataset if you only have few images due to overfitting.

[ardeal]

@Y-T-G
Thank you for your reply!
I would like to re-explain my questions:
for default yolov11m.pt, there are 80 classes.
for these 3 classes such as person, bicycle and car. I have already prepared more images for person, bicycle and car with correct label, and I would like to resume/retrain yolov11 network with yolov11m.pt initialization. The purpose of my training is to make yolov11 preform better.

[Y-T-G]

You can train normally. It's not resuming. It's starting from YOLO11 weights. But it will forget the old dataset it trained on if you don't include them in your dataset.

[ardeal]

@Y-T-G
Thank you for your reply!
The problem I am experiencing is that the accuracy gets worse after I train yolov11 with my those 3 classes images.
My training is initialized from the yolov11m.pt model.

My question:
I add more images to continue the training. Normally, the accuracy should get better, why the accuracy get worse after my training?

[glenn-jocher]

You're welcome! If you have any further questions about YOLO or training configurations, feel free to ask.

[sayyaradonis1]

Hi dear glenn,
I would like to continue training yolov11 model that is trained by coco dataset for example yolo11s.pt for only person detection by some custom person dataset.
How should I do?

[glenn-jocher]

To continue training yolo11s.pt for person detection, ensure your dataset is in YOLO format with only the "person" class labeled. Then, use the following command: yolo detect train data=your_dataset.yaml model=yolo11s.pt epochs=50 imgsz=640. Adjust parameters as needed. For more details, refer to the Train documentation.

[sayyaradonis1]

Thank you for your guidance.
We used RoboFlow to label the car, and all the labels in the annotation file were set to zero, while in Coco, the car class is 3. please help.

[glenn-jocher]

@sayyaradonis1 it seems your dataset annotations are misaligned with the COCO class indices. Ensure that the class indices in your annotation files match the COCO format. You can use a script to update the indices or reconfigure them in RoboFlow. For details on annotation formats, refer to Ultralytics Dataset Guide.

[ardeal]

@glenn-jocher

I use the following code to re-train yolov11.
The questions are:

1. the weights are initialized by yolov11m.pt, right?
2. are those data augments enabled in default config?

model_path = "yolo11m.pt"
model = YOLO(model_path)

device = torch.cuda.device(0)

train_results = model.train(
	data=r'cfg\datasets\person_ebike_bike_debug.yaml',
	epochs=100,  # number of training epochs
	imgsz=640,  # training image size
	device='0',  # device to run on, i.e. device=0 or device=0,1,2,3 or device=cpu
)

metrics = model.val()

[glenn-jocher]

@ardeal yes, the weights are initialized by yolo11m.pt when you load the model using YOLO(model_path). Regarding data augmentation, standard augmentations (e.g., mosaic, mixup, color, etc.) are enabled by default in the training configuration. You can adjust or disable these augmentations by modifying the augmentations section in your dataset configuration or passing specific arguments during training. For more details, check the Train Documentation. Let us know if you have further questions!

[ardeal]

@glenn-jocher
Thank you for your reply!

I did the following experiment:

step1: choose all those COCO 2017images which contains person, or bycycle, or motobicycle. I changed label of person, motobicycle and bike to 0, 1, and 2.

image  count == 65382
object count == 272962
person count == 257252
ebike  count == 8654
bike   count == 7056

step2: I prepare my own dataset:

image  count == 18798
object count == 38247
person count == 17126
ebike  count == 15252
bike   count == 5869

step3: I put the above 2 datasets in one folder.

step4: I modified yolov11 configuration: nc=3, warmup_epochs = 0, lr0=0.01/10

nc: 3
names:
  0: person
  1: ebike
  2: bike
  

step5: train yolov11 with the following code:

    model_path = r"yolo11m.pt"
    model = YOLO(model_path)
    device = torch.cuda.device(0)
    train_results = model.train(
        data=r'cfg\datasets\person_ebike_bike.yaml',  # path to dataset YAML
        epochs=100,  # number of training epochs
        imgsz=640,  # training image size
        device='0',  # device to run on, i.e. device=0 or device=0,1,2,3 or device=cpu
    )
    metrics = model.val()

results: the bad result is: the trained model usually recognize table legs as person :)

[glenn-jocher]

@ardeal it seems like the model is overfitting or encountering label-related confusion. Ensure your dataset is well-balanced and accurately labeled. Additionally, you may want to apply data augmentation techniques (e.g., flipping, scaling) to increase diversity and reduce overfitting. If the issue persists, consider fine-tuning learning rate, adding more training data, or increasing warmup_epochs. For further guidance, refer to the YOLO11 training documentation.