TDA4VM: Problem about TIDL simulation. The input is 10 channel float raw data

chen yizhao

Part Number: TDA4VM

Hi I have some problem about TIDL simulation.

The input of our model is 1x10x128x512 float raw data, not image. First, we test our model on TIDL in float32, and the output of TIDL is the same as pytorch. Then we test our model in int16, and we find that the output of TIDL is much different from pytorch. We think the problem is the input data. Because our input data is float32, so when testing float32 and int16 model in TIDL, the input will be different.

Can you give us some advise to deal with this problem? Do we need to set some extra commands when importing and testing TIDL model?

over 4 years ago

0 Jixiu Wu over 4 years ago

Prodigy 20 points

Hi, bro, have the problem resolved? I encounter the same problem with you, do you have any solutions to deal with the problem? Could you share it?

0 chen yizhao over 4 years ago in reply to Jixiu Wu

Prodigy 60 points

Nope, we haven't solved problem yet.

0 Jixiu Wu over 4 years ago in reply to chen yizhao

Prodigy 20 points

Hi, yizhao, I have solved the problem. There are two position you need to modify.

(1) in the function of tidl_ReadNetInput in tidl_tb_utils.c, add the following code

if(params->inFileFormat == 1)

{

elementSizeBytes = tidltb_getDatElementSize(ioPrms->rawDataInElementType[numBuffs]);

char ponitcloudFileName[500];

getFileNameAndIdxFromList((char*)params->inData, frameCount, numBuffs, ponitcloudFileName);

readRawFeatureData(ponitcloudFileName, (float32_tidl *)ins[numBuffs]->ptr, c, w, h, sizeof(float));

}

and the readRawFeatureData is as follows

int32_t readRawFeatureData(char *fileName, void *ptr, int16_t n, int16_t width, int16_t height, int32_t elemSize)

{

TI_FILE * fptr = FOPEN(fileName, "rb");

int32_t freadStatus = FREAD(ptr, elemSize, n * width * height, fptr);

if(freadStatus != (width * height * n * elemSize))

{

return 0;

}

return 1;

}

in test directory, recompile the code

make TARGET_PLATFORM=PC

(2) refer the following import configs, try different inQuantFactor, here in my case inQuantFactor = 64 is able to generate the correct results

modelType = 2
numParamBits = 8
numFeatureBits = 8
inputNetFile = "../../test/testvecs/models/public/onnx/polaris/model.onnx"
outputNetFile = "../../test/testvecs/config/tidl_models/onnx/tidl_net_model.bin"
outputParamsFile = "../../test/testvecs/config/tidl_models/onnx/tidl_io_model_"
inWidth = 1024
inHeight = 320
inNumChannels = 8

inQuantFactor = 64

rawDataInElementType = 6

inElementType = 1
outElementType = 1
inFileFormat = 1
inData = ../../test/testvecs/config/feature.txt
enableHighResOptimization = 1
enableLayerPerfTraces = 1
postProcType = 0
numFrames = 10
perfSimConfig = ../../test/testvecs/config/import/device_config.cfg

hope this can help you.

0 kumar.desappan over 3 years ago

TI__Mastermind 22145 points

The input data to TIDL during inference is expected to be in either fixed point 8-bit or 16-bit. Float32 bit mode is provided only for validation purposes.

Below Parameter can be used for the cases whether your input file different from model exeuction time

rawDataInElementType

This parameter is only applicable when inFileFormat is 1 and indicates the raw file's format. TIDL-RT test bench performs a conversion from rawDataInElementType to inElementType when these 2 are not same.
" Supported values are:
0 : unsigned char
1 : signed char
2 : unsigned short
3 : signed short
6 : float

software-dl.ti.com/.../md_tidl_sample_test.html

BTW, we recommend using Open Source Runtime APIs to avoid running into such issues.

https://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/08_00_00_12/exports/docs/tidl_j7_08_00_00_10/ti_dl/docs/user_guide_html/usergroup0.html

More details on the same and examples (Python/C++) can be found here

github.com/.../edgeai-tidl-tools

Processors

Processors forum

TDA4VM: Problem about TIDL simulation. The input is 10 channel float raw data