TDA4VE-Q1: yuv infer abnormal

? ?

Part Number: TDA4VE-Q1

We deployed 5 models in entry device , we find the inference results of 3 bin models are OK on device , but 2 bin models with 4 % AP decreased compared with simulation inference . So we think it as TI's API bug. How can wen solve the problem ?

9 months ago

0 ? ? 9 months ago

Prodigy 90 points

We use rtos to import the model . the sdk is ti-processor-sdk-rtos-j722s-evm-11_00_00_06 .

0 Chris Tsongas 9 months ago in reply to ? ?

TI__Genius 16500 points

Hi,

Could you please include your model, input data, and configuration files? There is very little we can determine with out them.

Regards,

Chris

0 ? ? 9 months ago in reply to Chris Tsongas

Prodigy 90 points

Hello , thanks for your reply . The models and evaluation result 3365.entry.tar.gz

0 ? ? 9 months ago in reply to ? ?

Prodigy 90 points

8738.img_in.zip

Fullscreen 4520.bev_seg2_inference_config_entry.txt Download

inFileFormat    = 2
postProcType = 3
inElementType = 0

netBinFile      = "/home/invo/2TB/evaluate/bev_seg2/output/pc_bin/entry/import_result/bev_seg2_int8.bin"
ioConfigFile   = "/home/invo/2TB/evaluate/bev_seg2/output/pc_bin/entry/import_result/bev_seg2_int8_1.bin"
outData =   "/home/invo/2TB/evaluate/bev_seg2/output/pc_bin/entry/infer_result/bev_seg2_int8_save.bin"
inData  =   "/home/invo/2TB/evaluate/bev_seg2/data/data_list_infer.txt"

#numFrames   = 14804
numFrames   = 500

#enableLayerPerfTraces = 2
#debugTraceLevel = 3
#writeTraceLevel = 3

Fullscreen 7343.bev_seg2_Import_config_entry.txt Download

#enableLayerPerfTraces = 1
#debugTraceLevel = 1
#writeTraceLevel = 3
modelType = 2
numParamBits = 8
numFeatureBits = 8
inputNetFile       = "/home/invo/2TB/evaluate/bev_seg2/data/onnx/model.onnx"

outputNetFile      = "/home/invo/2TB/evaluate/bev_seg2/output/pc_bin/entry/import_result/bev_seg2_int8.bin"
outputParamsFile   = "/home/invo/2TB/evaluate/bev_seg2/output/pc_bin/entry/import_result/bev_seg2_int8_"

#perfSimConfig = "/home/invo/2TB/evaluate/configs/device_config.cfg"
perfSimConfig = "/home/invo/2TB/rtos/ti-processor-sdk-rtos-j722s-evm-11_00_00_06/c7x-mma-tidl/ti_dl/test/testvecs/config/import/device_config.cfg"

inDataNorm  = 1
#inMean = 123.675 116.28 103.53
#inScale = 0.015625 0.015625 0.015625 
inMean = 0.0 0.0 0.0
inScale = 0.003921568627 0.003921568627 0.003921568627

#outElementType = 6 6
#addDataConvertToNet = 2
#outDataLayout = 1

#postProcType = 3
inDataFormat = 1
inFileFormat = 2
inWidth  = 448
inHeight = 448
#resizeWidth = 448
#resizeHeight = 448
inNumChannels = 3

numFrames = 3
#numFrames = 1
inData = "/home/invo/2TB/evaluate/bev_seg2/data/data_list_import.txt"

0 Chris Tsongas 9 months ago in reply to ? ?

TI__Genius 16500 points

Hi,

I noticed a problem with your entry/mos/model.onnx model that may be causing the issue.

[TIDL Import] UNSUPPORTED: Allowlisting : Layer name - /ArgMax : Only keepdims = 1 (default) is supported -- [tidlAllowlistingConstraints/tidl_constraint.cpp, 85]
[TIDL Import] UNSUPPORTED: Allowlisting : Layer name - /ArgMax_1 : Only keepdims = 1 (default) is supported -- [tidlAllowlistingConstraints/tidl_constraint.cpp, 85]

Can you please look at the ArgMax layers and ensure they are correct? The model in entru/3D/model.onnx imports and infers correctly.

./PC_dsp_test_dl_algo.out s:bev_seg2_inference_config_entry.txt

Processing config file #0 : bev_seg2_inference_config_entry.txt
----------------------- TIDL Process with REF_ONLY FLOW------------------------

# 0 . .. T 2257.09 .... ..... ... .... .....
# 1 . .. T 2259.54 .... ..... ... .... .....
# 2 . .. T 2253.60 .... ..... ... .... .....

As far as the 4% change it could be different set of inputs that could cause the variation.

Regards,

Chris

Processors

Processors forum

TDA4VE-Q1: yuv infer abnormal