[FEAT] DEIMv2 implementation

[tetelias]

💡 Is your feature request related to a problem?

I would like to integrate DEIMv2 into lightly-train. DEIMv2 has 2 families of models,
DINOv3 and HGNetv2/LiteEncoder, but I would suggest only DINOv3-based for now.

🧰 Describe the solution you'd like

A lot of code is already in place, though some will need divergence resolution.
DEIMCriterion(looks close, but differs from DFINECriterion with mal_alpha=None,
use_uni_set=True and loss loss_mal added) and DEIMTransformer will need to be added.

🛠 Alternatives you've considered

Not sure, if applicable

📝 Additional context

Some of code with divergence:

• when class LQE creates MLP, lightly just uses default "relu", while DEIM allows to create MLP with a different one.
• box_cxcywh_to_xyxy function in DEIM adds clamp(min=0.0) to width and height
• layer name converter will be required when loading checkpoints
• HybridEncoder:
  • fusion of upsample_feat+feat_low and downsample_feat+feat_height in DEIMv2 can be via both sum and concatenation,
    while Lighly always go for concatenation. And DEIMv2 pretrained models with sum
  • DEIMv2 has larger encoder block variety covering RT-DETR, D-FINE and DEIM aggregation path.
  • DEIMv2 uses 1x1 conv + depthwise spatial stride-2 conv for downsampling in both D-FINE and DEIM
  • DEIMv2 does not always uses input projection and goes for nn.Identity when in_channels is equal to hidden_dim, which is the case for DINOv3
• DEIMv2 has additional parameters and imports in HungarianMatcher to accomodate late stage training, where it switches to class-score × IoU ranking objective.
• DEIMCriterion can inherit from DFINECriterion to add parameters and loss.

[liopeer]

Hey, that's a really cool suggestion!

We are actively working towards opening up our LT-DETR model to swappable backbones and the detector is already pretty similar to DEIMv2 anyway. I am not sure if matching DEIMv2 100% is a high priority for us at the moment, but over the next few weeks we will probably converge to something between what we currently have, DEIMv2 and also DEIMv2's successor EdgeCrafter.