Compiler: TDA4X - Tensorflow Model Import Layer-by-Layer Matching

Akshat Agrawal1

Tool/software: TI C/C++ Compiler

Hi,

In TDA4X Import tool (tidl_j7_00_09_00_00), we are importing a Tensorflow customized Object detection Model. We are able to generate .net and .params for the model. We are not achieving the expected results from the output layers.

As suggested by TI, we are now trying to match the layer-by-layer output of the layers.While matching, we are not able to generate the "OutQ" value for the layer which is required for verification. Please tell us how do we obtain this value so that we can verify the layer dumping.

Kindly do the needful.

Thanks and Regards,

Akshat Agrawal.

over 6 years ago

0 Anshu Jain over 6 years ago

TI__Guru 56820 points

Hi Akshat,

TIDL already supports dumping layer level traces of all layers in floating point. You can do this by setting writeTraceLevel to 3 in infer config file.

Regards,

Anshu

0 Anshu Jain over 6 years ago in reply to Anshu Jain

TI__Guru 56820 points

Hi Akshat,

Is this issue resolved? We have updated documentation in latest TIDL package. You can refer the following section :

"TIDL User Guide"->"Trouble shooting"->"Steps to Debug Functional Mismatch"

Regards,

Anshu

0 Vyom Mishra1 over 6 years ago in reply to Anshu Jain

Genius 4590 points

Hi Anshu,

This suggestion we have tried on our side and it worked for us.

Thanks for the support.

Regards,

Vyom Mishra

Processors

Processors forum

Compiler: TDA4X - Tensorflow Model Import Layer-by-Layer Matching