This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

WEBENCH® Tools/PROCESSOR-SDK-DRA8X-TDA4X: TIDL Import tool for not supported networks

Part Number: PROCESSOR-SDK-DRA8X-TDA4X

Tool/software: WEBENCH® Design Tools

I would like to ask you about Yolo models the following questions:

Could you please elaborate more about how to implement models importing properly. 

 

1. How the config file is constructed for a custom made network? Any example on YOLOv3 model would help

2. How are the not supported layers (e.g., Leaky ReLU in YOLO) replaced with supported layers?

3. Is the size of the model a burden for the board? How?

  • Hi Alex,

      Please find my answers :

    1. How the config file is constructed for a custom made network? Any example on YOLOv3 model would help

    We have not validated YOLOv3 model but we have validated other objection detection models ( SSD based). You can refer import confile files of peelenet (import\public\caffe\tidl_import_peeleNet.txt) or mobilenetv1 ssd (import\public\tensorflow\tidl_import_mobileNetv1_0.75_ssd.txt) as an example

    2. How are the not supported layers (e.g., Leaky ReLU in YOLO) replaced with supported layers?

    Leaky Relu should get mapped to PRelu in TIDL

    3. Is the size of the model a burden for the board? How?

    Size should not be a problem as models are stored in external memory ( DDR).


    Regards,

    Anshu