RT-DETRV2 and normalization

🤗Transformers
1

i’m working on fine tuning a RT DETRV2 model and looking at the official fine tuning script here
they dont set do_normalize=True


image_processor = AutoImageProcessor.from_pretrained(
        args.model_name_or_path,
        do_resize=True,
        size={"max_height": args.image_square_size, "max_width": args.image_square_size},
        do_pad=True,
        pad_size={"height": args.image_square_size, "width": args.image_square_size},
        use_fast=args.use_fast,
        **common_pretrained_args,
    )

I looked at the notebook for fine tuning the model as well here
and they also don’t set do_normalize = True

image_processor = AutoImageProcessor.from_pretrained(
    checkpoint,
    do_resize=True,
    size={"width": image_size, "height": image_size},
    use_fast=True,
)

but the pre processor config for the RT-DETRV2 Image Processor here has do_normalize set to False by default even though it also provides image_mean and image_stdso the model is fine tuned without normalizing the inputs as far as I understand.
I’m confused, do RT-DETR models not need normalization when fine tuning?

1 Like