Error when running FpAnalyser with tiny yolov4

francisco sotofrancisco soto
in AI Model migration

Hello, I was trying to optimize my own onnx model (using onnx2onnx.py), which I couldn't make it, so I optimized the onnx model using another plataform, and then use the onnx2onnx.py.

This is the onnx model I used:

After that, I run the fpAnalyserCompilerIpevaluator_520.py script, giving the this error:

imagen.png

I would like to know how to solve this error,

Thanks!

Tagged:

Comments

  • kiddkidd

    which version of toolchain that you are using? And this is for which device? KL520 or KL720?

    I think there are two issues with the ONNX.

    1. your converted ONNX doesn't have shape. This might due to you didn't run the onnx2onnx.py
    2. There is a split layer in the model, which our devices don't support. See if you can modify with another op.
    image.png


  • francisco sotofrancisco soto

    Hi!

    I'm using the latest version of toolchain, and this is for a KL520 device.

    I couldn't optimized the onnx I sent you, so I used another plataform to optimized it.

    Then I using the onnx2onnx.py give me this onnx.


    So I'm using this onnx to continuing with the steps, but it gives me the error above.


    Thanks for the reply

  • kiddkidd

    it looks like your optimized doesn't convert the split correctly. It creates extra branch of slice. What was your intention to split? Get the first half of the channels?

    And BTW, KL520 also doesn't support slice as well.


    image.png


  • francisco sotofrancisco soto

    I understand, can you help about the error when I optimize the first ONNX I sent you?:

    imagen.png

    Thank you for the reply again.

  • Eric WuEric Wu
    edited April 2021

    Hi francisco soto,


    I saw there is "auto_pad" attribute in every convolution layer in the model you provided,

    image.png


    I have some experience about a bug comes from onnx builtin function due to "auto_pad" attribute, you can check my reply at the end in this post:

    https://www.kneron.com/forum/discussion/49/porting-tiny-yolov3-example#latest


    maybe that's why you got error message with "Inferred shape and existing shape differ",

    I think that would be better if you can modified your source code and set the explicit padding value for your convolution layer.

  • francisco sotofrancisco soto

    Thank you for the answer,

    I change the onnx model, so now I'm facing a new error:

    imagen.png

    Can you help me about this error?

  • Ethon LinEthon Lin

    Hi,

    The latest version of toolchain just update the opset from 9 to 11, if you want to run in the latest version, please use the python script "/workspace/libs/ONNX_Convertor/optimizer_scripts/onnx1_4to1_6.py" to upgrade the opset version.

    By the way, KL520 doesn't support the operators split and slice. If you want to inference model with the operators, please use KL720 instead of KL520.

  • francisco sotofrancisco soto

    Thank you about solving the error,

    Can the KL520 work in a aarch64 architecture?

  • Ethon LinEthon Lin

    No, KL520 is controlled through the cortex M4 which is 32 bit architecture.

  • francisco sotofrancisco soto

    Ok,  I understand.

    I have the requirement that the KL520 is connected to a NanoPi Neo 4, is that possible?

    Thanks.

  • Ethon LinEthon Lin

    KL520 could be used with RK3399, but the usb command of KL520 is transferred by libusb. https://github.com/libusb/libusb/releases/tag/v1.0.23

    User should make sure the libusb could work on their own environment.

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

© Kneron, Inc.