Reduce weight size

[hochschulfreunde]

Is there a way to reduce the size of the weight size?

We would like to use a self trained version in a JavaScript web application on user side only for privacy reasons. So they should not have to download something around 6mb. gzip compression does not help very much.

Everything till 1mb would be okay.

The model should just classify cards of two different types or detect cards of two different types in a picture.

To transfer the picture to the server and back is no oppertunity.

[glenn-jocher]

@hochschulfreunde hello! To reduce the weight size of your YOLOv8 model, you can consider the following strategies:

1. Model Pruning: This involves removing less important weights from the model. Pruning can lead to a smaller model size with a potential trade-off in accuracy.

2. Quantization: You can quantize your model weights to lower precision, which can significantly reduce the model size.

3. Knowledge Distillation: Train a smaller "student" model to mimic the behavior of your larger "teacher" model. The student model will be smaller in size.

4. Smaller Architecture: Start with a smaller base model or backbone that is designed for efficiency, such as YOLOv8-Nano, if available.

5. Fewer Classes: Since you're only detecting two types of cards, ensure your model is trained only on those classes to avoid unnecessary complexity.

6. Optimized Export Formats: When exporting your model, choose formats that are optimized for size and performance, like ONNX or TensorFlow Lite, and apply post-training optimizations.

Remember, each of these methods may affect the performance of your model, so you'll need to balance size reduction with maintaining adequate accuracy for your task. For more detailed guidance on these methods, please refer to our documentation at https://docs.ultralytics.com. Good luck with your web application! 🚀🃏