Preprocess for classification model

Hi, I try to infer the classification models on onnxruntime engin. however, I can`t get the consistent accuracy in the page except resnet mobilenet and shufflenet. I think the preprocess of these models is different and i use the preprocess in imagenet_preprocess.py. mobilenet and shufflenet converted from pytorch got normal accuracy while inceptionv1 and efficientnet-lite4 converted from caffe got low accuracy. I guess the preprocess of the models from caffe is different.

Start by reproducing the reported accuracy differences with the classification models in ONNX Runtime, then inspect imagenet_preprocess.py and compare its use for ResNet, MobileNet, ShuffleNet, InceptionV1, and EfficientNet-Lite4. Done means identifying whether the Caffe-converted models require different preprocessing and documenting or correcting the preprocessing needed for consistent accuracy.