TDA2EXEVM: different result between quantStatusTool and TDA2EXVEM

 there is my caffe model &prototxt & input images

2526.caffe_model.rarimages_9.rar

Hi,

Did you run your model on Vision SDK use case?
if so can you please try to run on standalone EVM by following steps in section 3.3.5 in the userguide.
And also, please refer to section 3.8 in the user guide to set qunat parameters for video sequence.

Thanks,
Praveen

The model on Vision SDK use case :file I/O,

There are two new problems:1.  change the input image size to 128X256.Run on eve the results were correct.but run on dsp ,the results were wong.

2.  Model files downloaded from the web, run on eve correctly, but run on DSP were wrong .

Did I miss something, please give me some advice?

Model file  &  images(32x32)

cifar-10_model.rarfour.rar

Hi,

1. For this image size of 128x256, could you please share the import config file and input image used to import the model bin files? Also, share the output results showing eve is correct and DSP is wrong?

2. This should work correctly on both EVE and DSP, can you please share EVE and DSP results along the config files used to run this.

Thanks,
Praveen

The import config for this image size of 128x256

Hi Praveen

The import config for this image size of 128x256 & input images used to import the model bin files.tidl_import_jseg21.rarimg_128_256.rarresults of image size of 128x256.rar

Mary

Hi Praveen

EVE and DSP results along the config files used to run this .Thank you very much.results of image size of 32x32.rar

Mary

Hi Mary,
First we will try to debug 32x32 image size.

I used the same model, proto text and input image (four.y) got the output of import , EVE and DSP are matching.
Final output is only 10 bytes as below..
A3 86 FF 8D 90 D1 6E E0 6F 80

Please use below import config and try to get same output as above for 32x32 image (four.y) in import, then EVE and DSP will match..
---------------------------------------------------------------------------------------------------------------------------------------------------------
# Default - 0
randParams = 0

# 0: Caffe, 1: TensorFlow, Default - 0
modelType = 0

# 0: Fixed quantization By tarininng Framework, 1: Dyanamic quantization by TIDL, Default - 1
quantizationStyle = 1

# quantRoundAdd/100 will be added while rounding to integer, Default - 50
quantRoundAdd = 25

numParamBits = 8
# 0 : 8bit Unsigned, 1 : 8bit Signed Default - 1
inElementType = 0

inputNetFile = "deploy.prototxt"
inputParamsFile = "cifar10_jacintonet11v2_iter_64000.caffemodel"
outputNetFile = "tidl_net_cifar.bin"
outputParamsFile = "tidl_param_cifar.bin"

rawSampleInData = 1
preProcType = 3
sampleInData = "..\..\test\testvecs\input\four.y"
tidlStatsTool = "..\quantStatsTool\eve_test_dl_algo.out.exe"
------------------------------------------------------------------------------------------------------------------------------------------------------------------------

We will get to 128x256 after we resolve this 32x32 image issue.

Thanks,
Praveen

Hi Praveen
I try to run on standalone DSP by following steps in section 3.3.5 in the userguide, Open the CCS and download the "dsp_test_dl_algo.out" through the emulator to the tda2xx development board. EVE and DSP are matching.
Final output is only 10 bytes similar to the...
A3 86 FF 8D 90 D1 6E E0 6F 80
-----------------------------------------------------------------------------------------------------------------------------
When I use the configuration file you provided.Generate these two files "Generate these two files
outputNetFile = "tidl_net_cifar.bin"
outputParamsFile = "tidl_param_cifar.bin"
Then copy the two files to the SD card, copy the four. y to the SD card, and pass the SD card boot program to run ALG_TIDL. Piece of selected the DSP result is wrong. Piece of selected the eve result is right.
--------------------------------------------------------------------------------------------------------------------------------
Thank you very much!
Mary

Hi Mary,

Could you please answer below questions to better understand the issue

1. Did you generate again the bin files ("tidl_net_cifar.bin" and "tidl_param_cifar.bin") after running on the standalone?
Ideally the bin files used on the standalone should work on VSDK.

2. Which usecase on VSDK you run to validate this ?

3. I assume you ran full network on EVE and its output is correct, but when you run full network on DSP, its output is wrong. Is this correct?
Because you mentioned " Piece of selected the DSP result is wrong ", what does this means?

Thanks,
Praveen

Hi Praveen
1. Did you generate again the bin files ("tidl_net_cifar.bin" and "tidl_param_cifar.bin") after running on the standalone?
Ideally the bin files used on the standalone should work on VSDK.
answer:Yes, I generate again the bin files ("tidl_net_cifar.bin" and "tidl_param_cifar.bin") after running on the standalone.

2. Which usecase on VSDK you run to validate this ?
answer: usecase file (I/O), The output section is modified to read the memory pointer directly and save to the SD card.This should not affect the ALG_TIDL section.

3. I assume you ran full network on EVE and its output is correct, but when you run full network on DSP, its output is wrong. Is this correct?
Because you mentioned " Piece of selected the DSP result is wrong ", what does this means?
answer: [IPU1-0] 1: DSP
[IPU1-0] 2: EVE
When I pick 1(DSP), the output is wrong, and when I pick 2(EVE), the output is correct.

I'm glad to see your reply again.
Mary

Hi Mary,

We will try to replicate this issue in VSDK usecase at our end and will come back to you.

Thanks,
Praveen

Hi Praveen,
Await your prompt reply!
thanks again!
Mary

Hi Mary,

Sorry for the late reply.
We tried to replicate the issue in VSDK use case and not able to match with the outputs that you shared.

You said in the previous post that, "output section is modified to read the memory pointer directly and save to the SD card". Could please try to run the use case without this change and check?

Thanks,
Praveen

Hi，Praveen
Thank you very much for remembering that . I tried not to modify the output, but the result was the same, the correct result was output on the eve, and the output was completely wrong on the DSP, I was strange, as if only I had encountered such a problem. I download the.dsp_test_dl_algo.out file through the emulator to DSP, the result is right, only when the SD card boot program, DSP can not output the correct results, I am very confused, if you later find the problem of the cloud, don't forget to tell me.
thanks
Mary