`head` and `backbone` keys missing

[fromtheeast710]

I'm using this package on a custom dataset by me. When I run python src/model/train.py --batch 64 --data coco.yaml --device 0 the training will fail with error:

KeyError: 'backbone'

When I check the generated dataset.yaml, there's no head nor backbone key. The generated yaml file is as follow

path: /home/theeast/Documents/Education/WSU/CS3/ProfessionalExperience/final-system-pa2524/data/train/color/YOLODataset
train: images/train
val: images/val
test:

names:
    0: wombat

This is the training script I'm using

from ultralytics import YOLO

wombot_data = "data/train/color/YOLODataset/dataset.yaml"

model = YOLO(wombot_data, task="segment")
model = YOLO("asset/model/yolo11x-seg.pt")
model = YOLO(wombot_data).load("asset/model/yolo11x.pt")

results = model.train(
    data=wombot_data, epochs=100, imgsz=640
)

[GreatV]

The error KeyError: 'backbone' occurs because your generated dataset.yaml file is missing the model configuration keys head and backbone. These keys are not required for a standard YOLO dataset YAML file; instead, they are found in YOLO model configuration YAMLs (e.g., yolov8n.yaml, yolov8x-seg.yaml, etc.).

Here's what's happening based on your process and the documents:

1. The dataset.yaml file

Your generated dataset.yaml looks correct for a YOLO dataset file. It contains paths for training/validation images and class names. YOLO expects this format for the --data argument:

path: /path/to/dataset
train: images/train
val: images/val
test:
names:
    0: wombat

This file should not have head or backbone keys. Those keys are for model config YAMLs, not dataset YAMLs.

2. The error source

Based on how you're initializing your model:

model = YOLO(wombot_data, task="segment")
model = YOLO("asset/model/yolo11x-seg.pt")
model = YOLO(wombot_data).load("asset/model/yolo11x.pt")

It appears that you may be accidentally passing the dataset YAML to functions that expect a model YAML, especially when you use YOLO(wombot_data). This would cause YOLO to look for model configuration keys (like backbone), leading to your error.

3. Correct usage

• To load a model: pass a model weights file or a model config YAML.
• To train a model: pass the dataset YAML through the data argument.

Example (for segmentation):

from ultralytics import YOLO

# Load a model by weights file
model = YOLO("asset/model/yolo11x-seg.pt")  # Load pre-trained segmentation model

# Train the model with your dataset
results = model.train(
    data=wombot_data,  # path to your dataset.yaml
    epochs=100,
    imgsz=640,
)

Do not use YOLO(wombot_data) by itself, since that tells Ultralytics YOLO to load a model, not data. That will trigger the code that looks for head or backbone in YAML and cause the KeyError.

4. References

This is a common confusion point with Ultralytics YOLO. The dataset YAML and the model YAML have different formats and can't be used interchangeably.

• Official guidance: YOLO Dataset YAML Format
• Official guidance: YOLO Model YAML Format

---

Summary:
You do NOT need head or backbone in dataset.yaml. Instead, ensure you only use the dataset YAML as the data parameter in model.train(...) and always load the model from weights or a model config YAML. The error is because you are passing the dataset YAML to the YOLO() call, which expects a model definition.

If you fix this by removing YOLO(wombot_data) and only calling YOLO() with a model file (like .pt), your problem should be solved.

---

Response generated by 🤖 feifei-bot | gpt-4.1

⚠️ This is an AI-generated response. Please verify any critical information.