TDA4AH-Q1: When we test our model in SDK 9.2, we find that the results on the board does not match the simulation results.

Hi Guang;

We have downloaded your model, and will look into it.

Meanwhile, could you show us how you set the numItrPerf = 1 and 2.

Also, is your the TIDL version  10.01.04.00?

Thanks and regards

Wen Li

The SDK I used is  ti-processor-sdk-linux-adas-j784s4-evm-09_02_00_05-Linux-x86-Install.bin 
And the TIDL version is 10.01.03.02 which provided by your technical support to resolve the deconv issue.

I run the demo TI_DEVICE_armv8_test_dl_algo_host_rt.out with config files，and set numItrPerf =1 or 2 in the config file，like this

/cfs-file/__key/communityserver-discussions-components-files/791/2577.tidl_5F00_infer_5F00_smoke.txt

TI_DEVICE_armv8_test_dl_algo_host_rt.out is build at c7x-mma-tidl/arm-tidl/rt

Hi Guang,

Can you provide an onnx model? caffe is kind of old.

Regards,

Adam

Hi Guang,

Could you also provide the input files as you specified in your configs ?

inDataNamesList = blob0,

inData = ref_image_list.txt

inData  =   testvecs/config/test.txt

Regards,

Adam

/cfs-file/__key/communityserver-discussions-components-files/791/model_5F00_data.zip

Here are all the materials you wanted.The issue has no correlation with the data used. I randomly selected an image for you to reproduce the problem.

Hi Guang
We found the issue; it is because of back-to-back deconvolution. 

We are fixing it. We will update you by monday ( 24th march)

Thanks, and Regards
Vijay

Hi Guang
yesterday, we already shared the patch.
Hope you received it, and patch is working fine.
If you need any further assistance, kindly let us know.

Thanks, and Regards
Vijay

There are still issues in our full deployment model.

The results of the first frame in the model are consistent with the simulation, while the results of the second frame are not consistent with the simulation.

We are currently pruning the model to reproduce the issue.

When infer with the same imput data, only the result of the first frame is correct.

I have provided all the relevant materials. Could you please look into the issue as soon as possible?

/cfs-file/__key/communityserver-discussions-components-files/791/Infer_5F00_Consistency_5F00_Error.zip

Hi Guang
We need some more details. 
1) can you share the output, you are getting?
2) which outputs you compared, (host emulation and target), or (ONNX (expected output) and host emulation)?

Thanks, and Regards
Vijay

Hello GUANG

To clarify further, we tried to use this model on host emulation during the weekend and couldn't reproduce the issue. We couldn't check on EVM/Target due to EVM having SD card flashed with the exact configuration of 9.2 SDK with 10.1 patch got disconnected from power socket. This needs physical presence in office and my team is willing to go to office and check tomorrow even though it is holiday in India. So appreciate you providing the requested details from Vijay to help us debug the issue faster.

Below are specific questions

• Do you see the second and first frame difference in host emulation mode? [YES/NO/Didn't try]
• Do you see the second and first frame difference in Target? [YES/NO/Didn't try]
• Can you share the output obtained for first and second frame?

Thanks,

With Regards,

Pramod

Hello, Pramod 

Below are the answers:

• NO, second and first frame is same in host emulation mode
• YES
• we provide the output result obtained for first and second frame in the below file

output_result.zip

Hello, Pramod 

We have done a review, and found that there are a total of four models with this issue. And we systematically raised a ticket.

https://e2e.ti.com/support/processors-group/processors/f/processors-forum/1494305/tda4ah-q1-the-consistency-issues-between-simulation-and-on-board-inference-as-well-as-the-consistency-issues-of-multiple-inferences-on-the-board

The issues with these four models are consistent with the consistency issue mentioned here. Please help us resolve the consistency issues with these four models as soon as possible

Thanks we are able to see the issue on target (evm) and debugging it