File size: 4,659 Bytes
edebaa5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
"""DINOv2 multi-head model for disaster building damage assessment.

Ported from hisaichi research code — simplified to inference-only,
original_only ROI mode (no two_stream, no CoVT, no cascade).
"""

import logging
from pathlib import Path

import timm
import torch
import torch.nn as nn
from peft import LoraConfig, get_peft_model

from .config import InferenceConfig

logger = logging.getLogger(__name__)


class DINOv2MultiHeadModel(nn.Module):
    """Multi-head model with DINOv2 backbone and optional auxiliary heads.

    Architecture:
        backbone (DINOv2 ViT-L/14) -> feature_transform -> head_full (6-class)
        Optional auxiliary heads: head_damage (2), head_disaster_type (2), head_severity (3)
    """

    def __init__(self, config: InferenceConfig) -> None:
        super().__init__()
        self.config = config

        # Backbone
        self.backbone = timm.create_model(
            config.model_name,
            pretrained=True,
            num_classes=0,
            img_size=config.image_size,
        )

        # Feature transform
        self.feature_transform = nn.Sequential(
            nn.Linear(config.hidden_dim, config.hidden_dim),
            nn.ReLU(),
            nn.Dropout(0.1),
        )

        # Main classification head
        self.head_full = nn.Linear(config.hidden_dim, config.num_classes)

        # Auxiliary heads (optional)
        if config.use_auxiliary_heads and not config.ce_only:
            self.head_damage = nn.Linear(config.hidden_dim, 2)
            self.head_disaster_type = nn.Linear(config.hidden_dim, 2)
            self.head_severity = nn.Linear(config.hidden_dim, 3)
        else:
            self.head_damage = None
            self.head_disaster_type = None
            self.head_severity = None

    def forward(self, x: torch.Tensor) -> dict[str, torch.Tensor]:
        """Forward pass through backbone and classification heads.

        Parameters
        ----------
        x : torch.Tensor
            Input images of shape ``(B, 3, H, W)``.

        Returns
        -------
        dict[str, torch.Tensor]
            Dictionary with key ``"full"`` (always present) and optionally
            ``"damage"``, ``"disaster_type"``, ``"severity"``.
        """
        feat = self.backbone(x)
        transformed = self.feature_transform(feat)

        outputs: dict[str, torch.Tensor] = {
            "full": self.head_full(transformed),
        }

        if self.head_damage is not None:
            outputs["damage"] = self.head_damage(transformed)
            outputs["disaster_type"] = self.head_disaster_type(transformed)
            outputs["severity"] = self.head_severity(transformed)

        return outputs


def build_model(config: InferenceConfig, device: str) -> nn.Module:
    """Create DINOv2 model with LoRA adapters applied.

    Parameters
    ----------
    config : InferenceConfig
        Model configuration.
    device : str
        Target device (e.g. ``"cuda"`` or ``"cpu"``).

    Returns
    -------
    nn.Module
        Model with LoRA adapters, moved to ``device``.
    """
    model = DINOv2MultiHeadModel(config).to(device)

    modules_to_save = [
        "feature_transform",
        "head_full",
        "head_damage",
        "head_disaster_type",
        "head_severity",
    ]

    lora_config = LoraConfig(
        r=config.lora_rank,
        lora_alpha=config.lora_alpha,
        target_modules=["qkv"],
        lora_dropout=config.lora_dropout,
        bias="none",
        modules_to_save=modules_to_save,
    )
    model = get_peft_model(model, lora_config)

    logger.info(
        "Built DINOv2 model with LoRA (r=%d, alpha=%d) on %s",
        config.lora_rank,
        config.lora_alpha,
        device,
    )
    return model


def load_checkpoint(model: nn.Module, checkpoint_path: Path, device: str) -> nn.Module:
    """Load trained weights from a checkpoint file.

    Parameters
    ----------
    model : nn.Module
        Model with LoRA adapters (from :func:`build_model`).
    checkpoint_path : Path
        Path to ``best_model.pth``.
    device : str
        Target device for weight mapping.

    Returns
    -------
    nn.Module
        Model in eval mode with loaded weights.

    Raises
    ------
    FileNotFoundError
        If ``checkpoint_path`` does not exist.
    """
    if not checkpoint_path.exists():
        raise FileNotFoundError(f"Checkpoint not found: {checkpoint_path}")

    state_dict = torch.load(checkpoint_path, map_location=device, weights_only=True)
    model.load_state_dict(state_dict)
    model.eval()

    logger.info("Loaded checkpoint from %s", checkpoint_path)
    return model