Tool/software:
Hello TI Expert,
Our team is trying to use TIDL to compile our own model on the AM62A device. The input of our model is 16-bit Bayer data with a (.raw) suffix. During the model import process, we input the Bayer data in binary format as inData and perform 8-bit quantization. After the TIDL import and inference, the model accuracy drops significantly. Therefore, we would like to try calibrationOption = 7 for high-precision calibration. Could you please advise us on how to set inData? We would like to use more than 50 binary files as the calibration dataset to improve the accuracy of the quantized model. We tried saving the 50 binary files into a single binary file as input and set numFrames=50, but it was unsuccessful. Alternatively, we hope TI experts can provide other suggestions to improve the model accuracy.
Our import configuration file is as follows:
modelType = 2 numParamBits = 8 quantizationStyle = 3 calibrationOption = 7 biasCalibrationIterations = 1 inputNetFile = "../../test/testvecs/models/public/onnx/model.onnx" outputNetFile = "../../test/testvecs/config/tidl_models_denoise/onnx/tidl_net_denoise.bin" outputParamsFile = "../../test/testvecs/config/tidl_models_denoise/onnx/tidl_io_denoise_" inDataNorm = 0 inHeight = 972 inWidth = 1296 inNumChannels = 1 inFileFormat = 1 inData = ../../test/testvecs/input/denoise/bayer.bin numFrames = 1 postProcType = 0
