TensorRT T4 fp16 speed misalignment

[cdy-for-grad]

Hello，i convert the official pretrained model to the onnx model by yolo export model=jameslahm/yolov10{n/s/m/b/l/x} format=onnx opset=13 simplify

Then, i convert the onnx to the tensorRT-fp16 in T4[TensorRT version 8.6.1.6],The result is as follow , latency is the mean value, the The last two lines are qps with different batch size.

wish you give the solution that i can get the result in your paper,Thank you

[JulioZhao97]

Also wondering how latency is measured? I think there is a standard procedure for measuring yolo models and can yolov10 author describe it to us? Gratefully appreciated!

[0margalal]

may you please state how you converted from Onnx to tensorRT ? plus, what are the requirements and could it be done on google colab ? Thanks in advance

[ShJacub]

I have the same problem. I had Yolov8s a bit better or equal fps than Yolov10 when benchmarking. on FP32 - approximately equal on FP16 - worse