Fail to run kl720_dme_serial_yolo with converted tiny-yolo-v3 from github

lc.wang

Platform: Raspberry Pi 4B + KL720 EVB USB dongle (PN: KP72B343A-M1).

Host lib version: host_lib_v0.9.1.

Model conversion and test:

Case 1.

1. Follow up http://doc.kneron.com/docs/#toolchain/yolo_example/, and change the model to yolov3-tiny in step 1, 520 to 720 in step 2 & 3.
2. Successfully convert to models_720.nef and transfer to Rbp4B.
3. Run original model successfully that verify connection is OK.
4. Replace with converted model in folder /home/pi/AIA_Kneron_social/host_lib/input_models/KL720/yolov3_tiny_416.
5. Run.

=> Fail.

Case 2.

1. Follow up https://www.kneron.com/forum/discussion/53/example-keras-kl520-how-to-convert-and-compile-tiny-yolo-v3-from-github-project, and change 520 to 720 in step 8.
2. Successfully convert to models_720.nef.
3. Others are the same as Case 1. It still fail to run.

Please help me.

Thanks

lc.wang

• Debug.

Case 1.

=> Change id from 1000 to 33, so as Case 2.

• Result.

=> Not the same as original model, why?

• Model size.

=> Therefore, neither yolo_example in docs nor example-keras-kl520-how-to-convert-and-compile-tiny-yolo-v3-from-github-project in forum are able to help the user convert to nef model file correctly.

Please post the correct procedure ASAP.

Jeffrey Chen

Hello,

which version SDK do you use? v1.1.1?

how about try to check the ISI mode?

It should be workable.

lc.wang

https://www.kneron.com/forum/discussion/comment/203#Comment_203

Hi 宇春,

Thanks for your reply.

1. I do not find your SDK for KL720 and http://doc.kneron.com/docs/#720_1.1.0/getting_start_720/ shows it is v1.1.0 so I use host_lib_v0.9.1. Please share SDK v1.1.1 for KL720.
2. I am in charge of building up an AI platform combined with our company's products.

I have some experience in Intel NCS2 and Google Coral Edge TPU.

To evaluate your chip with NCS2, TPU and Arm CPU, I need to build up benchmark flow.

I plan to use 100 still images for benchmark so that I use DME mode first.

ISI mode is the next step for streaming.

3. To shorten time, I plan to duplicate the inference result of kl720_dme_serial_yolo first.

a. I follow up 2 kind of your model converting procedures as case 1 and 2.

b. Replace models_720.nef with my converting and compare inference results.

c. It is fail. Please share your experience help me duplicate the inference result of kl720_dme_serial_yolo.

d. Comparing with converting Google TPU model, it is not necessary to feed 100 still images and is as the following.

Please share your "/data1/images" in https://www.kneron.com/forum/discussion/53/example-keras-kl520-how-to-convert-and-compile-tiny-yolo-v3-from-github-project#latest. It may help me duplicate the inference result of kl720_dme_serial_yolo.

4. The other question is how to convert png or bmp or jpg to bin file for DME_IMAGE_FILE input in kl720_dme_serial_yolo?

Otis Hsieh

Hi lc.wang,

Sorry for answering you after a long period of time.

We recommend you replace host lib with Kneron Plus because it is more user-friendly, efficient, and powerful.

For beginners, you can see the reference : http://doc.kneron.com/docs/#plus/getting_started/ , there are detail steps in the documentation.

If you still have any problem, we are willing to answer it and feel free to ask.

Thank you.

Otis Hsieh

Hi lc.wang,

If you want to use model yolov3-tiny, you can refer http://doc.kneron.com/docs/#plus/introduction/run_examples/

Our generic_inference_post_yolo is based on yolov3 tiny in both 520 / 720.

Thank you.