Nvidia Jetson Nano - Submissions for EXPORT competition

[batrlatom]

Compete and Win

Use this discussion thread for YOLOv5 🚀 EXPORT Competition submissions in the Nvidia Jetson Nano category. Good luck!

[juantoca]

Hi. We are preparing a submission for the competition. Even though it currently works as expected and it will be finished soon, we would like to know which Jetpack version are we expected to support. Are we supposed to support the last stable release(4.5.*) or shall we support also the older versions of Jetpack(4.4.*)?

[batrlatom]

4.5.* is fine to us

[sean-wade]

Where can I see the code submitted by the contributers? Must we wait until the end of the competition to see it?

[batrlatom]

Participants should place their submissions here. There are no entries to this category yet, therefore, no links are visible yet.

[sean-wade]

Ok, thanks a lot !!!

[SaiNikhileshReddy]

Is Jetson Nano 2GB Developer Kit eligible for competition?

[batrlatom]

Yes, we would love to see YOLOv5 on 2GB kit!

[deepconsc]

Coded & wrote tutorial for Jetson Nano 🙌

Needs a little review and refactor, which you'll be able to see in the end of July 12-16 week.

✅ Bonus l - Includes INT8 conversion & calibration tutorial.
✅ Bonus ll - as I've personally integrated distributed inferences with Xavier mini-nodes in couple of organizations, I guess some of our colleagues would find it helpful to have clear walkthrough and documentation on Jetson Xavier. Plus, INT8 acceleration on it is INSANE! I'll add its documentation as well.

https://github.com/deepconsc/tensorrt_yolov5

[batrlatom]

Thanks for your entry, we will test it as soon as possible. We changed the date of competition ending, so you can still work on the codebase.

[deepconsc]

Glad to hear! As I read evaluation is planned between 1-16 September. I'll update the repository to the final version much earlier than that 🙌🏻

[shivamswarnkar]

@deepconsc did you benchmark your solution on Nanor, Xavier, or any other devices? It'd be great if you could share that (I'm trying to optimize overall inference pipeline)

[MAli-Farooq]

Hi

Can I take part in this competition using Nvidia Xavier development kit ? I have implemented it and it's working good .

[batrlatom]

Hi, xavier and nano has almost the same software stack. It is fine to develop on xavier, it will work quite well. But we will test it on jetson nano, since it is much more affordable piece of hardware. We are looking fornyour submission

[shivamswarnkar]

Hi,

@aditya-dl and I have worked on this submission for Nvidia Jetson Nano.

https://github.com/BlueMirrors/Yolov5-TensorRT

We are getting identical outputs as PyTorch Inference(FP16). This uses letterboxing (so no need to forcefully resize) and you can directly test it on Colab as well (we got up to 165FPS on Colab T4 GPU in benchmarking).

P.S. We are working on int8 and batching support.

[batrlatom]

We are going to start the evaluation process next week. Stay tuned!

[alxmamaev]

I made experiments with yolov5 on mu Jetson Nano, and made docker container to build tensorrt models. Also I made python wrapper with NMS for it. Wrapper allows you process photos, batch of photos, videos and camera streaming. Demos included.

https://github.com/alxmamaev/jetson_yolov5_tensorrt

[batrlatom]

thank you for your submission. We are going to test it next week

[alxmamaev]

Update: I'm update NMS algorithm for more comparability with original implementation. Also I pushed docker image to docker hub, now you does not need build all by your own.

[stefani-kovachevska]

Hey @alxmamaev, I've tried to reach you at hi@alxmamaev.me. Could you get in touch by replying or emailing me at stefani@ultralytics.com? Thanks!

[alxmamaev]

Hey, yes!

[jario-jin]

Submission for YOLOv5 EXPORT Competition (Jetson platform)

I tested the detection accuracy and speed on COCO dataset on nano (472 gflops), TX2 (1.33 tflops), Xavier NX (21 tops) and AgX Xavier (32 tops). Two different languages (Python and C++) are provided, and it is found that C++ is much more efficient than python. Finally, a detailed tutorial is provided.

https://github.com/jario-jin/yolov5-on-nvidia-jetson

[batrlatom]

Good job, looks promissing. We are going to start the evaluation next week!

[deepsworld]

YOLOv5 EXPORT Competition (Jetson platform)

• All the dependencies including ones to train and validate in a compact docker container.
• On the fly TensorRT engine build with FP32, FP16, INT8 with calibration.
• No need for any custom plugins.
• At least 4X inference speed up...could be more with different devices and configs.
• Dynamic shape input is also supported

https://github.com/deepsworld/Yolo5Optimized

[glenn-jocher]

@deepsworld thank you for your submission!

[stefani-kovachevska]

Hi participants,

Thank you all for your submissions, we have already reviewed them and are about to announce the winners! Please send me an email at stefani@ultralytics.com including your full name and GitHub username, so I can share more information with you.

Stay tuned - the results are coming soon!

Stefani Kovachevska
Project Manager at Ultralytics

[stefani-kovachevska]

Congratulations to @alxmamaev, our winner in the Nvidia Jetson Nano category! 🎉
Check out the rest of the winners in the official announcement: https://ultralytics.com/article/export-competition-winners

Thank you all for participating and for making our community great!

Stay amazing and keep creating! 🚀
Team Ultralytics