diifferent when inference

when use pytorch model, img is normalised,
img = torch.from_numpy(img).unsqueeze(0).float()
img.div_(255).sub_(0.5).div_(0.5)

but use onnx model :
batch = (batch - model.input_mean) / model.input_std

it didn't normalise to 0-1

Start by locating the PyTorch and ONNX inference preprocessing entry points and comparing how each applies input mean and standard deviation. Reproduce the reported normalization difference with the issue's snippets, then verify that both paths produce the intended input range and add or update a focused regression test if the relevant test area is found.

Written by the indexing model from the issue text.