Normalization for pretrained feature extractor

[alamie]

For small image datasets (few hundreds of samples) one might want to use pretrained (on ImageNet) models to extract features from images and perform a classification task using a ML model (e.g. RandomForests, XGBoost,...). Pretrained models from torchvision.models can be used this way. The documentation of torchvision provides some recommendations regarding the use of pretrained models:

• input images should have shape (3, H, W) with H and W being at least 224.
• values have to be scaled to the [0, 1] range,
• image channels have to be normalized using ImageNet mean and std.

Pretrained models from timm could be used as well. Although I think that similar recommendations would apply for timm models, I was not able to find any documentation on this specific point. I have seen people do the following:

from timm import create_model

[...]

extractor = create_model(model_name, pretrained=True, in_chans=3, num_classes=0)
extractor.eval()

features = extractor(inputs)

However, inputs should be normalized before they are passed to the model, right?
The following seems more appropriate:

from timm import create_model
from timm.data.transforms_factory import transforms_imagenet_eval

extractor = create_model(model_name, pretrained=True, in_chans=3, num_classes=0)
extractor.eval()

cfg = model.default_cfg
img_size = cfg['test_input_size'][-1] if 'test_input_size' in cfg else cfg['input_size'][-1]
eval_transform = transforms_imagenet_eval(
    img_size=img_size, 
    interpolation=cfg['interpolation'], 
    mean=cfg['mean'],
    std=cfg['std'],
    crop_pct=cfg['crop_pct'])

inputs = scale(inputs)  # scale `inputs` to the [0, 1] range
inputs = eval_transform(inputs)
features = extractor(inputs)

Am I missing something?
Although extractor is fully convolutional, should the height and width of input images match default_cfg['input_size']?

[alamie]

If my question is unclear, I would be happy to reformulate :-)