yolo-world export onnx

[Egrt]

Search before asking

• I have searched the YOLOv8 issues and discussions and found no similar questions.

Question

I used yolo-worldv2 to train the custom model with only one category, when he exported the onnx, the onnx has only one output head which is inconsistent with the structure of the official yolo-worldv2.onnx, how should I modify it in order to get the same structure as the official yolo-worldv2.onnx?
The onnx model I exported:

The official onnx mode:

Additional

No response

[glenn-jocher]

Hello! It looks like your custom-trained YOLO-World model has a different output structure possibly due to the implementation differences aimed for a single class as opposed to multiple classes as in the official model.

A quick suggestion would be to check if any output layer specifications set during your training process are causing this variation. You may want to match the output layers' configuration to that of the official model. Ensure that the number of filters in the last convolutional layers matches num_classes * (5 + num_output_channels), where num_classes in the official model might be greater than 1 (for multiple classes).

For example, in your model training configuration (likely a YAML file), ensure settings similar to the following:

# For a single class, assuming 3 output heads
nc: 1  # number of classes
nl: 3  # number of detection layers (if applicable)
anchors: [...]

Adjust your YAML file to align with that of the official settings if you have access to them, especially the anchor setups and the number of output channels and classes.

If this doesn't resolve it, consider revisiting the modifications you made for your single class adaptation and how they differ from the official version, particularly around the output layers in your model's architecture file. If you can share more detailed snippets of your configuration or any custom modifications to the model architecture, I could provide more targeted advice. 🚀