How to prevent tensorrt from fusing the final layers when trying to convert tiny yolov3 to int8?

[NIKHILDUGAR]

I used the example given in https://github.com/jkjung-avt/tensorrt_demos#demo-5-yolov4 to convert my tiny yolov3 to int8 but the issue is that when I use --verbose on onnx_to_tensorrt.py step I see that the final layers are being fused.
How do I stop this from happening?

Thanks

[jkjung-avt]

In general, you could do mark_output() on a specific layer to prevent TensorRT from fusing it with other layers. Reference: NVIDIA/TensorRT#252 (comment)

Otherwise, please be more specific about which layers in your tiny yolov3 model you're having the trouble.

[NIKHILDUGAR]

Hey, thanks for the quick response. I did look at the reference you provided and the problem I am having with it is that it doesn't really take layer names only tensors and the method I have seen mostly used by people is:-

last_layer = network.get_layer(network.num_layers -1) network.mark_output(last_layer.get_output(0))
Now since the output layers of tiny yolo ( in my case) or yolo are on different levels I can't correctly identify which layers to mark and how to do so correctly.