Kneron toolchain not converting the YOLOv5s to KL520

Franklin OliveiraFranklin Oliveira
edited February 7 in KL520 related discussion

Hi,

I'm using the kneron/toolchain:v0.28.0 docker image to convert the YOLOv5s noupsample model to .nef for KL520. The steps that i'm following are:

1 Container run and model train and export to onnx

franklinls@MLOPS-Machine:~$ sudo docker run -it --rm --name=kneron kneron/toolchain

(onnx1.13) root@dc4f5de59e3d:/workspace# conda deactivate

(base) root@dc4f5de59e3d:/workspace# cd /workspace/ai_training/detection/yolov5/yolov5/

(base) root@dc4f5de59e3d:/workspace/ai_training/detection/yolov5/yolov5# wget https://raw.githubusercontent.com/kneron/Model_Zoo/main/detection/yolov5/yolov5s-noupsample/best.pt

(base) root@dc4f5de59e3d:/workspace/ai_training/detection/yolov5/yolov5# CUDA_VISIBLE_DEVICES='0' python3 train.py --data coco128.yaml --cfg yolov5s-noupsample.yaml --weights 'best.pt' --batch-size 4 --epoch 2 --workers 2

...

Optimizer stripped from runs/train/exp/weights/best.pt, 13.8MB

2 epochs completed in 0.032 hours.

(base) root@dc4f5de59e3d:/workspace/ai_training/detection/yolov5/yolov5# python3 ../exporting/yolov5_export.py --data ../yolov5/data/model_paths_520_coco128.yaml

self.vanish_point 0.0

Starting ONNX export with onnx 1.7.0...

****onnx file**** ./yolov5s-noupsample-coco128.onnx

...

ONNX export success, saved as ./yolov5s-noupsample-coco128.onnx

2 Model conversion from ONNX to NEF (KL520)

(base) root@dc4f5de59e3d:/workspace/ai_training/detection/yolov5/yolov5# conda activate onnx1.13

(onnx1.13) root@dc4f5de59e3d:/workspace/ai_training/detection/yolov5/yolov5# (onnx1.13) root@dc4f5de59e3d:/workspace/ai_training/detection/yolov5/yolov5# python onnx_to_nef.py 

2.1 km.evaluate() output

docker_version: kneron/toolchain:v0.28.0

comments:

kdp520/input bitwidth: int8

kdp520/output bitwidth: int8

kdp520/cpu bitwidth: int8

kdp520/datapath bitwidth: int8

kdp520/weight bitwidth: int8

kdp520/ip_eval/fps: 4.64802

kdp520/ip_eval/ITC(ms): 215.145 ms

kdp520/ip_eval/RDMA bandwidth GB/s: 0.8

kdp520/ip_eval/WDMA bandwidth GB/s: 0.8

kdp520/ip_eval/GETW bandwidth GB/s: 0.8

kdp520/ip_eval/cpu_node: Sigmoid: Sigmoid_129, Sigmoid_131, Sigmoid_133

gen fx model report: model_fx_report.html

gen fx model json: model_fx_report.json

2.2 km.analysis() output

 Failure for model "input/input" when running "kdp520/unimplemented feature"

Fixed-point analysis done. Saved bie model to '/data1/kneron_flow/input.kdp520.scaled.bie'

2.3 km.compile() output

[common][error][exceptions.cc:41][KneronException] UnimplementedFeature: undefined CPU op [Sigmoid] of node [Sigmoid_129]

[tool][error][batch_compile.cc:528][BatchCompile] Failed to compile input.kdp520.scaled.bie


As explained, the conversion process is not working due to the Sigmoid operation in Sigmoid_129 node mapped to CPU. So, how can i fix this?

Tagged:

Comments

  • Franklin OliveiraFranklin Oliveira

    Code


  • Maria ChenMaria Chen

    Hi,

    The nodes labeled as "cpu_nodes" are nodes that are not supported by KL520 NPU:

    kdp520/ip_eval/cpu_node: Sigmoid: Sigmoid_129, Sigmoid_131, Sigmoid_133

    Since KL520 doesn't support Sigmoid nodes, you will need to cut them in your onnx model, then add them back to your postprocess.

    You can find a list of supported nodes here: Hardware Supported Operators - Document Center

  • Franklin OliveiraFranklin Oliveira

    Hi Maria,

    Thanks for your reply.

    I've found the sigmoid in postprocess of kl520_kn-model-zoo_generic_image_inference_post_yolov5 example.

  • Franklin OliveiraFranklin Oliveira

    Hi again Maria.

    In my project, i'm using the Banana Pi P2 Zero with Armbian to access the KL520.

    So, i'm interested in use the Kneron DFUT on this board to flash the firmware but only the Raspberry Pi binary is available. How can i proceed in this case?

  • Maria ChenMaria Chen

    Hi Franklin,

    These are the platforms you can build and run Kneron PLUS on:

    image.png

    If you are not using the above platforms, to flash the KL520 firmware, you could download Kneron PLUS and try building the DFUT console:

    image.png

    Build with DFUT console: Build Kneron PLUS - Document Center

    Then, you could use the console to update the KL520 to USB Boot mode or Flash Boot mode, then upload the firmware.

    DFUT console: Run Examples - Document Center

  • Franklin OliveiraFranklin Oliveira

    Hi Maria,

    Thanks for your reply. I'll follow this build process.

  • Franklin OliveiraFranklin Oliveira
    edited February 12

    Hi again Maria,

    Is there any HW library available for KL520 TFBGA 8x8 159io (Symbol, footprint)? As reference, we're using Kicad

  • Maria ChenMaria Chen

    Hi again Franklin,

    For KL520 hardware information, you could refer to these documents available on our Developer Center: Developers | Kneron – Full Stack Edge AI

    image.png


  • Franklin OliveiraFranklin Oliveira

    Hi Maria,

    The 96 board/Main board files are not available for me. Can you send me, please?

    Captura de tela 2025-02-13 080105.png


  • Ethon LinEthon Lin

    Hi Franklin,

    The files are only provided to limited users. Please contact your purchase window to get access rights

  • Franklin OliveiraFranklin Oliveira

    Hi Ethon,

    I've tried to contact info@kneron.us, but without success until now. Do you know other contact (email or skype)?

  • Ethon LinEthon Lin

    Hi Franklin,

    Did you purchase any Kneron devices, and where did you get them? Please contact your purchase window instead of 'info@kneron.us.' We will provide download access for the product you bought.

    If you intend to purchase but have not bought the products yet, please contact our sales representative, Brian (brian.lin@kneron.us).

  • Franklin OliveiraFranklin Oliveira

    Hi Ethon,

    At my company, we're testing the KL520 USB dongle with the intention of integrating it into our product. So, we purchased it from Mouser for testing, and now we're interested in designing our own board based on the KL520 chip, for continuous purchase. Because of this, we need the footprint and symbol for Kicad.

    Would it be better to contact them through the purchase window, or should I reach out to Brian for assistance?

    Is that the contact through purchase window?

    Captura de tela 2025-02-18 083031.png


  • Maria ChenMaria Chen

    Hi Franklin,

    Yes, you should reach out to Brian (brian.lin@kneron.us), our sales representative. Thank you for your interest!

    You could also fill out the form on our website: Contact us | Kneron – Full Stack Edge AI

  • Franklin OliveiraFranklin Oliveira

    Hi Maria,

    I've tried reaching out through Brian and the Contact us | Kneron – Full Stack Edge AI, but unfortunately, I haven’t received a response. Do you have another contact I could try? My issue is simple to resolve, but the lack of communication is delaying our production.

    Thanks,

  • Maria ChenMaria Chen

    Hi Franklin,

    I'm sorry for their lack of communication. Could you contact Jeffrey (jeffrey-yc.chen@kneron.us) instead and cc Brian just in case? Thank you!

The discussion has been closed due to inactivity. To continue with the topic, please feel free to post a new discussion.

© Kneron, Inc.