Add DinoV3 Feature extractor

[TNodeCode]

I want to add a DinoV3 Feature extractor to the repository and build a Detektion Model that uses DinoV3 as a Feature extractor. I've already ritten some Code for the Feature extractor. Integrate thet Code into the framework.
ˋˋˋpy
class DINOv3ViT(nn.Module):

def __init__(self, finetuning=False, output_patches=False, layers=[2,5,8,11], layer_norm=True, *args, **kwargs):
    super(DINOv3ViT, self).__init__()
    self.finetuning = finetuning
    self.layers = layers
    self.output_patches = output_patches
    self.config = DINOv3ViTConfig()
    self.model = DINOv3ViTModel(self.config)
    self.layer_norm = layer_norm and (len(layers) + int(output_patches)) > 0
    if self.layer_norm:
        self.norms = nn.ModuleList([nn.LayerNorm(384, eps=1e-5, elementwise_affine=True) for _ in range(len(layers) + int(output_patches))])
    if not self.finetuning:
        self._freeze()


def _freeze(self):
    for param in self.model.parameters():
        param.requires_grad = False


def forward(self, x):  # should return a tuple
    """Forward image through model

    Args:
        x (torch.Tensor): image of shape (B, C, H, W)

    Returns:
        torch.tensor: Model output
    """
    outputs = {}
    if self.layers:
        def get_hook(name):
            def hook(module, input, output):
                outputs[name] = output
            return hook

        # register hooks
        if self.output_patches:
            self.model.embeddings.register_forward_hook(get_hook("embeddings"))
        for layer in self.layers:
            self.model.layer[layer].register_forward_hook(get_hook(layer))

    z = self.model(x)
    if len(outputs.keys()) == 0:
        outputs[11] = z.last_hidden_state

    for i, k in enumerate(outputs):
        if self.layer_norm:
            z = self.norms[i](outputs[k])
        else:
            z = outputs[k]
        # remove the [CLS] token
        z = z[:, 1:, :]
        # batch_size, num_patches, hidden_size
        B, P, D = z.shape
        h = w = int(P ** 0.5)
        z = z.permute(0, 2, 1)
        z = z.reshape(B, D, h, w)
        outputs[k] = z

    if len(outputs.keys()) > 1:
        return tuple(outputs[k] for k in outputs.keys())
    return outputs.get(next(iter(outputs.keys())))

ˋˋˋ

ˋˋˋpy
import torch.nn as nn
from mmdet.registry import MODELS
from transformers import DINOv3ConvNextConfig, DINOv3ConvNextModel

@MODELS.register_module()
class DINOv3ConvNext(nn.Module):

def __init__(self, finetuning=False, layers=[0,1,2,3], *args, **kwargs):
    super(DINOv3ConvNext, self).__init__()
    self.finetuning = finetuning
    self.layers = layers
    self.config = DINOv3ConvNextConfig()
    self.model = DINOv3ConvNextModel(self.config)
    if not self.finetuning:
        self._freeze()


def _freeze(self):
    for param in self.model.parameters():
        param.requires_grad = False


def forward(self, x):  # should return a tuple
    """Forward image through model

    Args:
        x (torch.Tensor): image of shape (B, C, H, W)

    Returns:
        torch.tensor: Model output
    """
    outputs = {}
    if self.layers:
        def get_hook(name):
            def hook(module, input, output):
                outputs[name] = output
            return hook

        # register hooks
        for i in self.layers:
            self.model.stages[i].register_forward_hook(get_hook(i))

    z = self.model(x)
    if len(outputs.keys()) == 0:
        outputs[0] = z.last_hidden_state

    if len(outputs.keys()) > 1:
        return tuple(outputs[k] for k in outputs.keys())
    return outputs.get(next(iter(outputs.keys())))

ˋˋˋ

Also write a custom Detektion head that uses

[TNodeCode]

Can you also create a custom Detektion head using hungarian matching?