For my custom semantic segmentation model, I am following the conversion pipeline from PyTorch to ONNX, HAR, HAR optimized, and finally HEF.
I checked that my torch model and one model produce the same outputs.
My torch model takes a normalized image as input in the dataloader (mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) and produces an output.
In this case, when I optimize har with calib data with the command:
hailo optimize --hw-arch hailo8 --calib-set-path …/data_npy/calib_512_1500.npy PIDNet.har
should I use the normalized numpy array (created with the same normalization used with Pytorch) for the calibration data npy file rather than using the real image values in the [0, 255] range?
Note that my custom model doesn’t have an input normalization layer.
I proceeded this way, but the segmentation results from using my custom HEF are performing very poorly.