TDA4VM: Edgeai-torchvision QAT mAP issue

Vishwa Teja

Part Number: TDA4VM

HI,

I have followed the process provided in the official edgeai-torchvision document (edgeai-torchvision/Quantization.md at master · TexasInstruments/edgeai-torchvision · GitHub) for implementation of QAT on my OD model. I have trained my model for few epochs and generated ONNX successfully. After testing the model on TI Board (TDA4VM), it is found that the quantized int8 model and the floating-point fp32 model still have a large difference in mAP after QAT.

Used the import configurations mentioned below.

modelType = 2
numParamBits = 8
numFeatureBits = 8
quantizationStyle = 3
calibrationOption = 64
inputNetFile = "/home/prakash/TI_SDK_8/ti-processor-sdk-rtos-j721e-evm-08_05_00_11/tidl_j721e_08_05_00_16/ti_dl/test/testvecs/models/public/onnx/QAT/ResNet_CenterNet_FR_TI_QAT_15bestepoch_W512_H320_C3_05_05_2023_simplified.onnx"
outputNetFile = "../../test/testvecs/config/tidl_models/onnx/CenterNet_ResNet18_FR_W736_H512_C3_23_01_2023/QAT/05_05_23/M1/8/tidl_net_centernetresnet_qat_34_8.bin"
outputParamsFile = "../../test/testvecs/config/tidl_models/onnx/CenterNet_ResNet18_FR_W736_H512_C3_23_01_2023/QAT/05_05_23/M1/8/tidl_io_centernetresnet_qat_34_8"
inDataNorm = 1
inMean = 97.155 96.39 94.605
inScale = 0.029935638 0.031372549 0.031882672
inWidth = 512
inHeight = 320
inNumChannels = 3
inDataFormat = 0
inFileFormat = 2
inElementType = 1
#writeTraceLevel = 3
#debugTraceLevel = 1
writeOutput = 2
inData = "../../test/testvecs/config/import_resnet_ABC_50_vis.txt"
outDataNamesList = "heat_map,wh_map,reg_map"