Batch Normalization kernels are missing when running YOLOv8n

AI & Data Science Deep Learning (Training & Inference)
1

I am trying to profile YOLOv8n through NCU and do not find any kernel for batch normalization. Upon confirming I see in literature and even when printing the YOLO model that it contains a couple of BatchNorm2d layers.

2

The batch normalization layers get fused with convolution during export in Ultralytics.

You can get manually do that by running model.fuse() in Ultralytics.

3

Is there some way to preserve the batch normalization layers when exporting to ONNX format?