Raspberry Pi - Submissions for EXPORT competition

[batrlatom]

Compete and Win

Use this discussion thread for YOLOv5 🚀 EXPORT Competition submissions in the Raspberry Pi category. Good luck!

[karanjakhar]

https://github.com/karanjakhar/yolov5-export-to-raspberry-pi

Please have a look.

[shivamswarnkar]

@karanjakhar did you benchmark your solution in raspi (or any other device)? If yes then what FPS did you get?

[bot-user-01]

Hi @karanjakhar for inferencing on video, the code is loading the tflite model for each frame? That's why it is giving slow results as well.

[karanjakhar]

@rajchem You are right. Thank you for pointing it out. I have updated it.

[bot-user-01]

@karanjakhar can you help how can I adjust the code for INT8 quantized tflite model?

[jaskiratsingh2000]

@karanjakhar
I just have a question that what is the core difference between the Pytorch Model and the Tflite model? As I could see that Pytorch model is not working good on the Raspberry Pi whereas the Tflite works better but I am not able to figure out at what are the reasons for that its working.

[AnhPC03]

Hello, I've create my repository to export YOLOv5 model from Pytorch framework to MNN framework (MNN is a lightweight deep neural network inference engine, embedded devices, such as IoT). Please take a look, feel free to comment or create an issue. Hope you enjoy it.
https://github.com/AnhPC03/yolov5-export-to-raspberry-pi-mnn

[batrlatom]

Thanks. We will look at it

[nocolour]

Good Job.
but slower than tflite. If can fine-tune it and increase the speed will be perfect.

[AnhPC03]

@nocolour I've not compared my model with tflite, I think you did. But my code default saves image/video/webcam into local. Maybe it makes my code slower than your tflite model? Or you had add --nodisplay --nosave when inference?

[shivamswarnkar]

@AnhPC03 good job man! btw, what FPS did you get (on the colab cpu and/or raspi)?

[jaskiratsingh2000]

Why there is a need to export Pytorch model to other for making it run on Raspberry Pi?

[jaskiratsingh2000]

Does anyone have any idea about this issue, please? - #4167

Would highly appreciate it if someone can help in know this.

[jaskiratsingh2000]

@batrlatom @glenn-jocher I just have a key question that what are we trying to do with exportation of current Yolov5 into the exported version? What is the need to export for Raspberry Pi like devices?

What is the difference between current yolov5 and the exported version of yolov5?

Can anyone of you please let me know? I would highly appreciate that.

[batrlatom]

@jaskiratsingh2000 Hi, the goal is to make a yolov5 inference as fast as possible on Raspi. In normal situation, yolov5 is not quite realtime on such system, so applying quantization or any other technique which makes it faster is what it needs to be done.

[jaskiratsingh2000]

@batrlatom Can you be more specific please because I am a researcher working with Yolo and Raspberry Pi and this is something very important for me to understand in terms of more detailed specifications.
Because I am kinda stuck here - #4167
Do you have any idea about this mentioned issue as well? That would really be of great help for me to understand.
Please let me know.

Thanks and Regards

[batrlatom]

@jaskiratsingh2000 I do not own raspi now, so I am unable to exactly try out the installation. But are you sure you have everything the same? The same checkout version, all libraries are on the same version? Computers are the same, but software layers could be different. If you do the pip list on both system, does it output the same version of all libraries? Did you try implementation of submissions from others? It looks like others were able to make it work. If you test the outputs visualy, are they really so different? The precission you mentioned would imply that raspi version does not output almost any rectangle correctly.

[jaskiratsingh2000]

@batrlatom Yes. Every Library on both the machines is the same. I tried the implementation of the others but none of them actually have implemented the Accuracy Performance and due to this reason only I was not able to understand the key question that what is the difference between the current Yolo version and the Exported version.
I would really appreciate it if you can help me with this issue since this is the core problem I see for Yolo on Rpi

[batrlatom]

in utils/general.py is a function def non_max_suppression . There is a variable time limit set to 10. Try to increase it, hopefully it helps with the mAP computation.

[jaskiratsingh2000]

@karanjakhar I am still looking forward to hearing your response on this - https://docs.google.com/document/d/1ej8hIktTFmsqTH0GCYVw-qD8sF9C9fa_Pe61EenNmkU/edit?usp=sharing

Can you please let me know?

If anyone else also has any idea about this then please let me know since I and @batrlatom have been discussing this but couldn't reach a conclusion.

Your response would be highly appreciable. Thanks!

[batrlatom]

@jaskiratsingh2000 Please solve the problem in the thread which is separate from the competition itself, thank you

[antran89]

Hi, I am a newbie, do we need to test it on Raspberry Pi device.

[blange48]

Hello, the new realease of yolo bring a amazing export model witch work perfectly on raspberry pi. With a 320 image, it is possible to reach 3 FPS and works perfectly with the raspberry camera.

[JNaranjo-Alcazar]

Okey, that is what I was thinking.... Thanks for sharing! 👍🏼

[JNaranjo-Alcazar]

One question regarding execution times. The inference time using the demo images is around 450 ms using int8 tflite model.

Can someone give some advice in how to optimize this times and further steps?

Thanks a lot for the help of this community! 👍🏼

[blange48]

That would be very nice @blange48 if you shared your running image. I would appreciate that 🥇

Hello, the image is available here : https://drive.google.com/file/d/1WrhaWJL2_eIJKJHgdf5KFnYqruFVlq2R/view?usp=sharing

[JNaranjo-Alcazar]

Nice! Thank you @blange48

[jaskiratsingh2000]

@blange48 I just have a quick question to understand regarding which I am having a hard time understanding it since long. What is actually inside the Pytorch model and Tflite model? I don't understand that what's the format inside the PyTorch model. When do we say that we are compiling the Pytorh model then what does happen intermediately?

Also, @blange48 Were you able to get the mAP value of any model on the RPi.

I am tagging other people if in case they have any idea about this then that would really lead to a healthy discussion.
cc @zldrobit @JNaranjo-Alcazar @batrlatom @glenn-jocher @shivamswarnkar

[jaskiratsingh2000]

Hey Everyone! @glenn-jocher @zldrobit @blange48 I just have a question that what is the core difference between the Pytorch Model and the Tflite model? As I could see that Pytorch model is not working good on the Raspberry Pi whereas the Tflite works better but I am not able to figure out at what are the reasons for that its working.

[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 [email protected] 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]

Hi everyone,

We were thrilled to have received your submissions to our YOLOv5 Export Competition for the Raspberry Pi category.

Unfortunately, none of the submissions passed the minimum set of requirements in each of the evaluation criteria. Therefore, no winner was selected for the category this time. However, there will be more chances for participants to compete again in the future.

Thank you all for your efforts and for being a valuable member of our open-source community. Contributions from you are what makes our community great. 🙌

Be sure to check out the winning repositories in the rest of the categories in our latest blog post: https://ultralytics.com/article/export-competition-winners

Stay amazing and keep creating! 🚀
Team Ultralytics