TDA3MD: TIDL import tool failed

Hi, 

Could you please let me know how you are importing the model ? I mean steps/commands used for importing ?

Thanks,

Praveen

Hi, 

I run on command line:

tidl_model_import.out.exe cfg.txt

it says so in your UserGuide. Also i tried to Build TIDL reference executable using the command:

make TARGET_PLATFORM=PC TARGET_BUILD=release CORE=eve RUN_REF_FOR_STATS=1 all

and this is what i get:

C:\PROCESSOR_SDK_VISION_03_08_00_00\ti_components\algorithms\REL.TIDL.01.02.00.00\modules\ti_dl>make TARGET_PLATFORM=PC TARGET_BUILD=release CORE=eve RUN_REF_FOR_STATS=1 all
.
======== tidl ALGO =================
gmake -C ./algo -f makefile
process_begin: CreateProcess(NULL, gmake -C ./algo -f makefile, ...) failed.
make (e=2): The specified file cannot be found.

Hi,

It looks like "tempDir" is created while importing, could you please use "importTestCases.bat" to for importing and check if it works.

Regarding building TIDL reference executable, you no need to build it and you can use the executable from the release, and also you will not able to build as the TIDL source code build is not available publicly.  

Thanks,

Praveen

"importTestCases.bat" creates directories including "tempDir".

But "tempDir" is empty, and the rest is a text file, where the following is written

Couldn't open inputNetFile file: ..\..\test\testvecs\config\caffe_models\dense_examples\ti_dl_dense_1x1.prototxt  

Would it be possible for you to share details on your Project and requirements? Wanted to make sure, your target application can be supported on TDA3x or will it be better to try TDA4x devices

If it helps, can you consider trying your model with our TDA4x based tools available here

https://github.com/TexasInstruments/edgeai-tidl-tools

Hi Kymar! We were inspired by the «TDA3x SoC processor delivers cost effective ADAS solutions» and decided to create an ADAS system based on TDA3. The neural network component is supposed to be used for a small part of the image. In terms of technical characteristics, TDA3 completely satisfies us and we wouldn’t want to use TDA4, which has larger sizes and power consumption.

However, the TDA4 impressed us with its performance. Is it worth waiting for a smaller version of TDA4 in the near future, like TDA3 became a smaller version of TDA2 before that?

We do have parts from TI sitara processor family AM62 with Analytics capability ( recently launched) and we do have road-map to provide parts with higher analytics capability available this year with dedicated accelerators for ML and request to get in touch with Field for more details

www.ti.com/.../SK-AM62

Thanks for the info, we'll look into this later.

But at the moment we would like to embed a neural network on TDA3, it is ideal for our task. Based on the material "Deep Learning Inference For Embedded Applications Reference Design" one DSP and one EVE suits us completely.

However, when trying to embed a neural network, we ran into a problem. To check the conversion of the ONNX model to the TIDL format, a simple model was created with one convolutional layer and a non-linear ReLU activation function.

After running the conversion script, we got an error about missing files:

Processing config file .\tempDir\qunat_stats_config.txt !

Running TIDL simulation for calibration.

Could Not Open Files .\tempDir\temp_net.bin

Processing Frame Number : 0

Max values is -128 at index 0, width = 0, height = 0
Could Not Open Files .\tempDir\temp_net.bin
End of config list found !

The temp_net.bin file is simply missing and it is not clear how to get it.

Our actions are described in more detail in previous messages.

On your advice, we used "importTestCases.bat" to check the creation of the TempDir directory, in which the files necessary for the conversion should be generated. However, after running "importTestCases.bat", an empty TempDir directory is created.

Can you please tell me how to solve the problem with converting the model for TIDL?