| from __future__ import annotations |
|
|
| import torch |
| import torch.nn as nn |
| import timm |
|
|
|
|
| class SIFQBackbone(nn.Module): |
| """Backbone wrapper around timm models for grayscale fingerprints.""" |
|
|
| def __init__(self, model_name: str = "tiny_vit_5m_224.dist_in22k", pretrained: bool = True): |
| super().__init__() |
| self.encoder = timm.create_model( |
| model_name, |
| pretrained=pretrained, |
| num_classes=0, |
| in_chans=1, |
| ) |
| self.feature_dim = int(getattr(self.encoder, "num_features", 0)) |
| if self.feature_dim <= 0: |
| raise ValueError("Backbone num_features is missing or invalid") |
|
|
| def forward_spatial(self, x: torch.Tensor) -> torch.Tensor: |
| """Return spatial/token features where possible.""" |
|
|
| feats = self.encoder.forward_features(x) |
| if feats.ndim == 4: |
| b, c, h, w = feats.shape |
| return feats.permute(0, 2, 3, 1).reshape(b, h * w, c) |
| if feats.ndim == 3: |
| return feats |
| if feats.ndim == 2: |
| return feats.unsqueeze(1) |
| raise ValueError(f"Unexpected feature tensor shape: {tuple(feats.shape)}") |
|
|
| def forward(self, x: torch.Tensor) -> torch.Tensor: |
| spatial = self.forward_spatial(x) |
| return spatial.mean(dim=1) |
|
|
