While most models from torch.hub can be loaded via our existing timm.wrappers, I've observed that models instantiated this way can produce slightly different tensors in the forward pass than when instantiating the model using torch.hub.load, probably related to type castings. We should add a model wrapper that supports loading models using torch.hub.load. While this can be achieved already using the ModelFromFunction wrapper, this doesn't support out_indices which is required for segmentation tasks in eva.vision.
While most models from torch.hub can be loaded via our existing
timm.wrappers, I've observed that models instantiated this way can produce slightly different tensors in the forward pass than when instantiating the model usingtorch.hub.load, probably related to type castings. We should add a model wrapper that supports loading models usingtorch.hub.load. While this can be achieved already using theModelFromFunctionwrapper, this doesn't supportout_indiceswhich is required for segmentation tasks ineva.vision.