TIDL import of ONNX model

Hi Kumar,

Thanks for your reply 

I tried to set  "outDataNamesList" as u told but still the error is not resolved, I have attached my onnx where the transpose is involved

Also the "concat " is  triggering an error  as below

Could not find const or initializer 648 !!!
Only float and INT64 tensor is suported

Any idea of why is it happening so.

Hi Lola,

  Can you share what values you have configuredin outDataNamesList?

Regards

Anshu

Hi Anshu

The  "outDataNamesList" parameter value is resolved anyhow but now it is resulting in below error.

Hi,

   Can you set numFrames = 1 in import config file and try?

Regards,

Anshu

Hi,

Thanks a lot for all your help the import is done without any errors.

But, I need clarification on how to implement the PostProcessing part, how to link my remaining model to find the confidences and locations.

My model is MobileNetV2+SSD, I have attached the .zip of my input ONNX model, Configuration file used for import and also the Graphviz svg file for visualise the output of Tidl-import.

4087.files.zip

Hi Lola,

   This model has 8 output tensors and dimensions for each of them are apart of the IOparams which is part of import tool.  Each output ID is associated with a name in the original model. This name is  available in the generated svg file during the import tool . You can use names to identify the tensors. You can refer the following document to understand how these output tensors are stored :

http://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/latest/exports/docs/tidl_j7_01_01_00_10/ti_dl/docs/user_guide_html/md_tidl_fsg_io_tensors_format.html

  Once you know which output maps to which output in your original network then you can implement the post processing.

Regards,

Anshu

Hi Anshu,

Can you help me with identifying the Output ID its name and how to map to the output of original network so as to implement the post processing.

Also mention which code or function i need to check so as to know how to perform post processing.

Regards,

Lola

Hi Lola,

   Is the documentation mentioned in above reply not helping : 

http://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/latest/exports/docs/tidl_j7_01_01_00_10/ti_dl/docs/user_guide_html/md_tidl_fsg_io_tensors_format.html 

   You can refer tidl_tb_postProc.c file in ti_dl/test/src/tidl_image_postproc.c file to understand how post processing is implemented in TIDL test bench.

Regards

Anshu

Hi Anshu 

The documentation you provided i am referring to it, it was helpful.

But the thing i struck is at  "Each output ID is associated with a name in the original model. This name is  available in the generated svg file during the import tool . You can use names to identify the tensors. "

When i check in svg file i cant identify any output ID, u can refer to the below attached image.

So i was not able to identify  which output maps to which output in my original network.

I hope i explained my issues you can ask me if you need any more details.

Hi Anshu,

Thanks for the links  provided.

Found out the required parameter values

Thanks