TDA2SX: TIDL SSD model setup

Sasa Mihajlica

Part Number: TDA2SX

Hi everyone,
I have SSD object detection model which I want to run on SoC. During conversion eve_test_dl_algo.exe is run and I get correct results in stats_out_tool.bin file. I am running full net with eve_test_dl_algo.exe application. However, by instructions, I should run it partially on EVE and partially on DSP core. Since I don't have DSP test app, a made usecase similar to TIDL OD one where I am running 1 EVE core and 1 DSP core with TIDL algorithm link with this model. But, I am getting empty list of bounding boxes as a result. Also, I tried this model with TIDL usecase and results are the same as for the previous usecase.

Does anyone know why I am getting correct results only during the conversion but only when it is not run by isntructions and on SoC not at all?

Regards,
Sasa

over 7 years ago

0 kumar.desappan over 7 years ago

TI__Mastermind 22145 points

Did you get the model working with the original TIDL OD use case without any modification to the usecase?
Most cases the difference result are observed if the input used for import and use cases are not the same.
Check the input tensor passed to use case and make sure, it matches with the first tensor of import (trace_0*.y)

0 Sasa Mihajlica over 7 years ago in reply to kumar.desappan

Intellectual 860 points

Input to import and usecases are same and identical to trace_dump_0*.y . I didn't check the model with original TIDL OD usecase yet because originali it takes video input instead of singe image so modifications are necessarry. Can I somehow get output from all layers on SoC like during the import process having trace_dump_* file for all layers?

0 kumar.desappan over 7 years ago in reply to Sasa Mihajlica

TI__Mastermind 22145 points

No, generation trace on the SoC is not supported by the library.

Processors

Processors forum

TDA2SX: TIDL SSD model setup