Do you speed up by TensorRT ?

[wu-ruijie]

v5 is so fast! I even dare to imagine how fast it is speeded up by TensorRT, Do you has any job about it ?

[github-actions]

Hello @wu-ruijie, thank you for your interest in our work! Please visit our Custom Training Tutorial to get started, and see our Jupyter Notebook , Docker Image, and Google Cloud Quickstart Guide for example environments.

If this is a bug report, please provide screenshots and minimum viable code to reproduce your issue, otherwise we can not help you.

If this is a custom model or data training question, please note that Ultralytics does not provide free personal support. As a leader in vision ML and AI, we do offer professional consulting, from simple expert advice up to delivery of fully customized, end-to-end production solutions for our clients, such as:

• Cloud-based AI systems operating on hundreds of HD video streams in realtime.
• Edge AI integrated into custom iOS and Android apps for realtime 30 FPS video inference.
• Custom data training, hyperparameter evolution, and model exportation to any destination.

For more information please visit https://www.ultralytics.com.

[glenn-jocher]

@wang-xinyu did a great TensorRT implementation of our https://github.com/ultralytics/yolov3 repo here (which supports both YOLOv3 and YOLOv4), he might best answer this question.
https://github.com/wang-xinyu/tensorrtx/tree/master/yolov3-spp

[HaxThePlanet]

Tensor core support would be amazing!

[sljlp]

Hi! I tested yolo5-s on cpu by directly running detect.py and the inference speed is only 3 fps.Could you please give me some advice?I want to make it 30 fps at least.

[glenn-jocher]

@sljlp you might want to see 'Running yolov5 on CPU' #37

The default --img-size for detect.py is 640, which you can reduce significantly to get the FPS you are looking for.

[glenn-jocher]

@sljlp one caveat is --img-size must be a multiple of the largest stride, 32. So acceptable sizes are 320, 288, 256, etc.

[glenn-jocher]

Update: I've pushed more robust error-checking on --img-size now in 099e6f5, so if a user accidentally requests an invalid size (which is not divisible by 32), the code will warn and automatically correct the value to the nearest valid --img-size.

[thancaocuong]

@glenn-jocher Can you provide yolov5.weights file. I've found that to convert yolo to tensorrt, we need the weights file to use with (https://github.com/wang-xinyu/tensorrtx/)

[glenn-jocher]

@thancaocuong there is no such file.

[TrojanXu]

I have a python implementation here, with NMS, https://github.com/TrojanXu/yolov5-tensorrt

[wang-xinyu]

Hi @glenn-jocher

I just implemented yolov5-s in my repo https://github.com/wang-xinyu/tensorrtx/tree/master/yolov5
, and test on my machine. yolov5-m, yolov5-l, etc, will come out soon.

Models	Device	BatchSize	Mode	Input Shape(HxW)	FPS
YOLOv3-spp(darknet53)	Xeon E5-2620/GTX1080	1	FP16	608x608	38.5
YOLOv4(CSPDarknet53)	Xeon E5-2620/GTX1080	1	FP16	608x608	35.7
YOLOv5-s	Xeon E5-2620/GTX1080	1	FP16	608x608	167
YOLOv5-s	Xeon E5-2620/GTX1080	4	FP16	608x608	182
YOLOv5-s	Xeon E5-2620/GTX1080	8	FP16	608x608	186

[wang-xinyu]

Update! My tensorrt implementation already updated according to this commit 364fcfd

The PANet updated.

Please find my repo https://github.com/wang-xinyu/tensorrtx

[alexandrebvd]

Update! My tensorrt implementation already updated according to this commit 364fcfd

The PANet updated.

Please find my repo https://github.com/wang-xinyu/tensorrtx

Thanks for sharing! Do you have plans to implement other yolov5 versions as well?

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[wang-xinyu]

We have updated the yolov5 tensorrt according to the v2.0 release of this repo.

And made speed test on my machine.

Models	Device	BatchSize	Mode	Input Shape(HxW)	FPS
YOLOv5-s	Xeon E5-2620/GTX1080	1	FP16	608x608	142
YOLOv5-s	Xeon E5-2620/GTX1080	4	FP16	608x608	173
YOLOv5-s	Xeon E5-2620/GTX1080	8	FP16	608x608	190
YOLOv5-m	Xeon E5-2620/GTX1080	1	FP16	608x608	71
YOLOv5-l	Xeon E5-2620/GTX1080	1	FP16	608x608	40
YOLOv5-x	Xeon E5-2620/GTX1080	1	FP16	608x608	27

please find https://github.com/wang-xinyu/tensorrtx.

@glenn-jocher could you also add a link to https://github.com/wang-xinyu/tensorrtx in your Tutorials section?

[glenn-jocher]

@wang-xinyu thanks, yes this is a good idea. Can you submit a PR for the README please?

EDIT: I'll add a link to the export tutorial also.

[ttanzhiqiang]

https://github.com/ttanzhiqiang/onnx_tensorrt_project

[fire717]

We have updated the yolov5 tensorrt according to the v2.0 release of this repo.

And made speed test on my machine.

Models Device BatchSize Mode Input Shape(HxW) FPS
YOLOv5-s Xeon E5-2620/GTX1080 1 FP16 608x608 142
YOLOv5-s Xeon E5-2620/GTX1080 4 FP16 608x608 173
YOLOv5-s Xeon E5-2620/GTX1080 8 FP16 608x608 190
YOLOv5-m Xeon E5-2620/GTX1080 1 FP16 608x608 71
YOLOv5-l Xeon E5-2620/GTX1080 1 FP16 608x608 40
YOLOv5-x Xeon E5-2620/GTX1080 1 FP16 608x608 27
please find https://github.com/wang-xinyu/tensorrtx.

@glenn-jocher could you also add a link to https://github.com/wang-xinyu/tensorrtx in your Tutorials section?

Thx for your work, I just wonder how do u test FPS with batchsize.
Cause our video is just one img flow, every img is in a serial line, so why could u use batchsize more than 1?