COSIE.image_preprocessing.create_model

create_model(local_dir)[source]

Create and load a pre-trained Vision Transformer (ViT-L/16) model from a HuggingFace-compatible checkpoint.

Parameters

local_dirstr: Path to the folder containing the pre-trained model weights (e.g., pytorch_model.bin).

Returns

modeltorch.nn.Module: A ViT-Large (patch16, img224) model from the timm library, without classification head or global pooling.