TDA4VM: discrepancy between PC emulation outputs and target execution outputs

Hi Kumar,

Sorry our project requirement is sdk7.1, can you please share document about it?

one more thing we have noticed Pc emulator results are same each run, but in EVM results vary each run. Can you please why this is happening also?

okay let me go throw it, is there any way where I can dump results for Model intermediate layers and compare on Pc and Evm? 

Hi Kumar,

This guide you shared is mostly covering to debug the issue between Pc emulator and Original Model.

Our issue is different, we already have the application, lets say app_tidl_od, it is giving different results when evaluated on PC and run on EVM.

Plus results are being varied in EVM each time, in PC that app results are same.

debugTraceLevel = 2 and writeTraceLevel = 3  are set in inference configuration file, but in app (app_tidl_od) we don't have that. We only have config file. so please help me to save values in that app.

Hi Kumar,

We followed the steps you have shared in the documentation. And we are able to dump the tensor outputs for each layer in TIDL test bench, ie by running the 

regards,

Gina

Is it possible to share the Layer configuration where you notice the first difference between the PC and EVM traces?

Also is it possible to share the model and sample image where you observe the difference ween the PC and EVM? This will help us in re-producing this issue.

Hi Kumar, 

We are working on how to dump it in QNX, can you share parameters?

For the model and images sorry our company is not allowing us to share with you guys. I have tried 16 bit segmentation and Object classification app but can't figure out if those issue remains. 

Do you have any 16bit models for object detection? like peelenet or any so we can test it with app_tidl_od and compare the results with our model.

Thank you

As of now build of "TI_DEVICE_a72_test_dl_algo_host_rt.out" is not supported in QNX SDK. We are working on adding this support in the next SDK release.

To make progress for now, could you please use Linux-based PSDKR to debug and isolate this issue. Once problem is fixed, you can continue to use QNX based flow

For the model and images sorry our company is not allowing us to share with you guys. I have tried 16 bit segmentation and Object classification app but can't figure out if those issue remains. 

Do you have any 16bit models for object detection? like peelenet or any so we can test it with app_tidl_od and compare the results with our model.

Thank you

Hi Mudasar,

We do have peele net part of our validation.

You can set numParamBits = 16 (Currently 8) in the release import config file and import the model and test it  

Hi Kumar, 

Thank you let me try.

Hi kumar, I followed your instructions and converted peele net, but here are issues during conversion

plus I downloaded the caffe model from this link https://github.com/GuoJaw/caffe-ssd-mobilenet/tree/master/models/Pelee

Please let me know, in case I am doing something wrong.

Could you please try the 8-bit mode and confirm that the model you are using is right model

Hi Kumar,

Still same issue as 16bit,

I am taking Model from this link https://github.com/GuoJaw/caffe-ssd-mobilenet/tree/master/models/Pelee 

In case you have link to model or models can you share?

We have tested the model from below

github.com/.../Pelee

Hi Kumar,

I tried  models from github.com/.../Pelee and results are as following can you confirm what is issue now?

okay, let me try Thank you