Use size instead of scale factor to specify the output size of nn.interpolate(): this avoids any rounding issue leading to mismatching output size and consistently generate the same output size as with the previous kludge (from facebookresearch/dino#8).

Account for register tokens in get_intermediate_layers

* Update param_groups.py

Update lr decay rates for reg tokens

* Update param_groups.py

Add new backbones (and matching linear classification heads) trained with 4 registers following [Vision Transformers Need Registers](https://arxiv.org/abs/2309.16588).

Pass scale factor as a tuple of floats to F.interpolate() to allow tracing.

Add streamlined model versions w/o the mmcv dependency to directly load them via torch.hub.load().

Add streamlined model versions w/o the mmcv dependency to directly load them via torch.hub.load().

Ignore import warnings in hubconf and trigger lint workflow on PR.

Rework support code for torch.hub.load() to allow reusing shared functions and eventually expose more models.

Update the top-level README to make it clearer what's currently available.

Add semantic segmentation (Mask2Former based on ViT-Adapter) code + update demo notebook for segmentation with a dedicated section.

Add semantic segmentation (linear) code + demo notebook

Add depth estimation code + demo notebook

Add (optional) extras dependencies for dense tasks (mmcv and mmsegmentation) to conda and pip requirements.

Update code and models license from CC-BY-NC to Apache 2.0 in headers and other files.

We put default number of nodes to 1 in particular so that the linear evaluation with 8 GPUs has a global batch size of 1024, to reproduce the results.

Allow disabling the use of xFormers (for inference) by simply setting the XFORMERS_DISABLED environment variable

Add missing explicit layers argument

Passing a list as default argument is not recommended.

Issue: flake8 processes the venv dir.
Change: added `exclude = venv` to flake8 conf

Signed-off-by: Alexander Seiler <seileralex@gmail.com>

Remove hardcoded selection of operator implementation and use xFormers fMHA dispatcher instead.

Fix linear classifier wrapper in PyTorch Hub configuration module to support multi-sample batches.

Document and fix implementation of extra metadata generation for ImageNet-1k.

Remove unused sample decoding interface in datasets.

Installation, dependencies and Python module search path as well as a few other things.

Fix author name.