File size: 4,788 Bytes
e94400c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
"""
DINOv2 vision backbone wrapper.

Features:
  - Loads DINOv2 variants via torch.hub (with local fallback)
  - Exposes patch token features (x_norm_patchtokens)
  - Provides preprocessing (resize + normalization) for multi-view PIL images
  - Parallel per-view preprocessing using ThreadPoolExecutor
"""

from collections import OrderedDict
import os

from concurrent.futures import ThreadPoolExecutor
import torch

import torch
import torch.nn.functional as F
from torch import nn
from torchvision.models._utils import IntermediateLayerGetter
from typing import Dict, List
from torchvision import transforms


def apply_transform(view, transform):
    return transform(view)


# from llavavla.model.modules.dino_model.dino_transforms import make_classification_train_transform


class DINOv2BackBone(nn.Module):
    """
    Thin wrapper around a DINOv2 model.

    Args:
        backone_name: DINOv2 model id (e.g. dinov2_vits14, dinov2_vitb14).
        output_channels: (Unused placeholder; retained for future extension).

    Attributes:
        body: Loaded DINOv2 model.
        num_channels: Feature dimension of patch tokens.
        dino_transform: Preprocessing pipeline (resize + tensor + normalize).
    """

    def __init__(self, backone_name="dinov2_vits14", output_channels=1024) -> None:
        super().__init__()
        try:
            self.body = torch.hub.load("facebookresearch/dinov2", backone_name)
        except:
            import traceback

            traceback.print_exc()
            print(f"Failed to load dinov2 from torch hub, loading from local")
            TORCH_HOME = os.environ.get("TORCH_HOME", "~/.cache/torch/")
            weights_path = os.path.expanduser(f"{TORCH_HOME}/hub/checkpoints/{backone_name}_pretrain.pth")

            code_path = os.path.expanduser(f"{TORCH_HOME}/hub/facebookresearch_dinov2_main")

            self.body = torch.hub.load(code_path, backone_name, source="local", pretrained=False)

            state_dict = torch.load(weights_path)
            self.body.load_state_dict(state_dict)
        if backone_name == "dinov2_vits14":
            self.num_channels = 384
        elif backone_name == "dinov2_vitb14":
            self.num_channels = 768
        elif backone_name == "dinov2_vitl14":
            self.num_channels = 1024
        elif backone_name == "dinov2_vitg14":
            self.num_channels = 1408
        else:
            raise NotImplementedError(f"DINOv2 backbone {backone_name} not implemented")
        self.dino_transform = transforms.Compose(
            [
                transforms.Resize((224, 224)),
                transforms.ToTensor(),
                transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
            ]
        )
        # self.dino_transform = make_classification_train_transform()

    # @torch.no_grad()
    def forward(self, tensor):
        """
        Forward pass.

        Args:
            tensor: Image batch tensor [B*views, 3, H, W].

        Returns:
            torch.Tensor: Patch token features [B*views, N_tokens, C].
        """
        xs = self.body.forward_features(tensor)["x_norm_patchtokens"]

        return xs  # B*views, token, dim

    def prepare_dino_input(self, img_list):
        """
        Preprocess a batch of multi-view PIL image lists into a tensor suitable for DINO.

        Args:
            img_list: List of samples; each sample is List[PIL.Image] (multi-view).

        Returns:
            torch.Tensor: Flattened batch of shape [B * num_view, 3, H, W] on model device.
        """
        # img_list: is a list of [PIL], each representing multi views of the same example.
        # refer to https://github.com/facebookresearch/dinov2/blob/main/dinov2/data/transforms.py

        # use thread pool to parallel process each view
        with ThreadPoolExecutor() as executor:
            image_tensors = torch.stack(
                [
                    torch.stack(list(executor.map(lambda view: apply_transform(view, self.dino_transform), views)))
                    for views in img_list
                ]
            )

        # move the tensor to the device of DINO encoder
        B, num_view, C, H, W = image_tensors.shape
        image_tensors = image_tensors.view(B * num_view, C, H, W)
        device = next(self.parameters()).device
        image_tensors = image_tensors.to(device)

        return image_tensors


def get_dino_model(backone_name="dinov2_vits14") -> DINOv2BackBone:
    """
    Factory helper returning a configured DINOv2BackBone.

    Args:
        backone_name: DINOv2 variant name.

    Returns:
        DINOv2BackBone: Initialized backbone instance.
    """
    return DINOv2BackBone(backone_name)


if __name__ == "__main__":
    dino = DINOv2BackBone()
    pass