Philipp Wirth

Hi @aymuos15 that's great! I will assign the issue to you. To start off, I would recommend you fork the repo and read the [guide on how to contribute](https://github.com/lightly-ai/lightly/blob/master/CONTRIBUTING.md#pull-requests). Let...

Hi @aymuos15, how is it going? Can we help you in any way?

Hey @masc-it this should work almost out of the box. We have the necessary tools for momentum updates here: - https://github.com/lightly-ai/lightly/blob/c1c4ee2b2de8e3fa2e831cf142d4864e80c0b448/lightly/models/utils.py#L159 - https://github.com/lightly-ai/lightly/blob/c1c4ee2b2de8e3fa2e831cf142d4864e80c0b448/lightly/models/utils.py#L186 For masking and ViTs we can use...

I quickly put this example together: ```python # Note: The model and training settings do not follow the reference settings # from the paper. The settings are chosen such that...

Thanks for the feedback, @guarin. I went through the paper again and added dummies for focal masks, projection heads, and prototypes. The only thing missing in the proposal above is...

Quick update, we can't simply use the `SimCLRCollateFunction` as shown above because they use different augmentations for each of the anchor images (hence we need a `MultiViewCollateFunction` instead).

Also, correct me if I'm wrong, but I think the `focal_mask` is simply implemented as a crop in the official repo: https://github.com/facebookresearch/msn/blob/4388dc1eadbe3042b85d3296d41b9b207656e043/src/deit.py#L213

Looks like all we need is the adapted loss as the architecture is the same as for MoCo.

Doesn't the `lightly.data.collate.BaseCollate` class offer an interface for such pipelines? The pipeline need only be implemented as a `torchvision.transforms.Compose`.

Yes, this issue is still open and nobody is currently working on it.