File size: 9,596 Bytes
dadf189 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 | from __future__ import annotations
"""Custom Vision Transformer trained from scratch.
Supports ViT-Tiny (192-D, 5.7M), ViT-Small (384-D), ViT-Base (768-D), and
a DINOv2-compatible ViT-B/14 profile (``dinov2_vitb14_reg``).
Returns patch tokens, CLS token, and all per-layer attention maps for TRAM.
Unlike the timm-based ``vit_backbone.py`` (pretrained models), this module
is designed to be trained from scratch on fingerprint data with 1-channel
(grayscale) input — no 3-channel replication needed.
"""
import torch
import torch.nn as nn
# ───────────────────────────────────────────── DropPath ────
class DropPath(nn.Module):
"""Stochastic depth (per-sample drop of residual branches)."""
def __init__(self, drop_prob: float = 0.0):
super().__init__()
self.drop_prob = drop_prob
def forward(self, x: torch.Tensor) -> torch.Tensor:
if self.drop_prob == 0.0 or not self.training:
return x
keep = 1.0 - self.drop_prob
shape = (x.shape[0],) + (1,) * (x.ndim - 1)
mask = torch.rand(shape, dtype=x.dtype, device=x.device).add_(keep).floor_()
return x / keep * mask
# ───────────────────────────────────────────── PatchEmbed ──
class PatchEmbed(nn.Module):
"""Image → non-overlapping patch tokens via Conv2d."""
def __init__(
self,
img_size: int = 224,
patch_size: int = 16,
in_chans: int = 1,
embed_dim: int = 192,
):
super().__init__()
self.in_chans = in_chans
self.grid_size = (img_size // patch_size, img_size // patch_size)
self.num_patches = self.grid_size[0] * self.grid_size[1]
self.proj = nn.Conv2d(
in_chans, embed_dim, kernel_size=patch_size, stride=patch_size,
)
def forward(self, x: torch.Tensor) -> torch.Tensor:
if x.shape[1] != self.in_chans:
# Allow transparent conversion between RGB and grayscale inputs.
if self.in_chans == 1 and x.shape[1] == 3:
x = x.mean(dim=1, keepdim=True)
elif self.in_chans == 3 and x.shape[1] == 1:
x = x.repeat(1, 3, 1, 1)
else:
raise RuntimeError(
f"PatchEmbed expected {self.in_chans} channels, got {x.shape[1]}"
)
return self.proj(x).flatten(2).transpose(1, 2) # (B, N, D)
# ───────────────────────────────────────────── Attention ───
class Attention(nn.Module):
"""Multi-head self-attention — returns output AND attention weights."""
def __init__(
self,
dim: int,
num_heads: int = 12,
attn_drop: float = 0.0,
proj_drop: float = 0.0,
):
super().__init__()
self.num_heads = num_heads
self.head_dim = dim // num_heads
self.scale = self.head_dim ** -0.5
self.qkv = nn.Linear(dim, dim * 3)
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop)
def forward(
self, x: torch.Tensor
) -> tuple[torch.Tensor, torch.Tensor]:
B, N, C = x.shape
H, d = self.num_heads, self.head_dim
qkv = self.qkv(x).reshape(B, N, 3, H, d).permute(2, 0, 3, 1, 4)
q, k, v = qkv.unbind(0) # each (B, H, N, d)
attn = (q @ k.transpose(-2, -1)) * self.scale
attn = attn.softmax(dim=-1) # (B, H, N, N)
attn_weights = attn.detach() # save for TRAM
out = (self.attn_drop(attn) @ v).transpose(1, 2).reshape(B, N, C)
out = self.proj_drop(self.proj(out))
return out, attn_weights
# ───────────────────────────────────────────── Block ───────
class Block(nn.Module):
"""Pre-norm Transformer block with stochastic depth."""
def __init__(
self,
dim: int,
num_heads: int,
mlp_ratio: float = 4.0,
drop: float = 0.0,
attn_drop: float = 0.0,
drop_path: float = 0.0,
):
super().__init__()
self.norm1 = nn.LayerNorm(dim)
self.attn = Attention(dim, num_heads, attn_drop, drop)
self.drop_path = DropPath(drop_path) if drop_path > 0 else nn.Identity()
self.norm2 = nn.LayerNorm(dim)
hidden = int(dim * mlp_ratio)
self.mlp = nn.Sequential(
nn.Linear(dim, hidden),
nn.GELU(),
nn.Dropout(drop),
nn.Linear(hidden, dim),
nn.Dropout(drop),
)
def forward(
self, x: torch.Tensor
) -> tuple[torch.Tensor, torch.Tensor]:
attn_out, attn_weights = self.attn(self.norm1(x))
x = x + self.drop_path(attn_out)
x = x + self.drop_path(self.mlp(self.norm2(x)))
return x, attn_weights
# ───────────────────────────────────────────── ViT ─────────
class ViT(nn.Module):
"""Vision Transformer with per-layer attention extraction.
Preset configurations::
tiny: 192-D, 12 layers, 12 heads (~5.7M params, 1ch input)
small: 384-D, 12 layers, 12 heads (~22M params)
base: 768-D, 12 layers, 12 heads (~86M params)
dinov2_vitb14_reg: 768-D, 12 layers, 12 heads (profile-compatible)
Parameters
----------
variant : str
``"tiny"`` | ``"small"`` | ``"base"`` | ``"dinov2_vitb14_reg"``.
img_size : int
Input image resolution (square).
patch_size : int
Patch side length in pixels.
in_chans : int
Input channels (1 for grayscale, 3 for RGB).
drop_rate : float
Dropout for embeddings and MLP.
attn_drop_rate : float
Dropout on attention weights.
drop_path_rate : float
Stochastic depth max rate (linearly increased per layer).
"""
PRESETS: dict[str, dict] = {
"tiny": dict(embed_dim=192, depth=12, num_heads=12),
"small": dict(embed_dim=384, depth=12, num_heads=12),
"base": dict(embed_dim=768, depth=12, num_heads=12),
"dinov2_vitb14_reg": dict(embed_dim=768, depth=12, num_heads=12),
}
def __init__(
self,
variant: str = "tiny",
img_size: int = 224,
patch_size: int = 16,
in_chans: int = 1,
drop_rate: float = 0.0,
attn_drop_rate: float = 0.0,
drop_path_rate: float = 0.1,
):
super().__init__()
preset = self.PRESETS[variant]
self.embed_dim: int = preset["embed_dim"]
depth: int = preset["depth"]
num_heads: int = preset["num_heads"]
self.patch_embed = PatchEmbed(img_size, patch_size, in_chans, self.embed_dim)
num_patches = self.patch_embed.num_patches
self.grid_size = self.patch_embed.grid_size
self.cls_token = nn.Parameter(torch.zeros(1, 1, self.embed_dim))
self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, self.embed_dim))
self.pos_drop = nn.Dropout(drop_rate)
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]
self.blocks = nn.ModuleList([
Block(self.embed_dim, num_heads, mlp_ratio=4.0,
drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i])
for i in range(depth)
])
self.norm = nn.LayerNorm(self.embed_dim)
self._init_weights()
# ------------------------------------------------------------------
def _init_weights(self):
nn.init.trunc_normal_(self.pos_embed, std=0.02)
nn.init.trunc_normal_(self.cls_token, std=0.02)
for m in self.modules():
if isinstance(m, nn.Linear):
nn.init.trunc_normal_(m.weight, std=0.02)
if m.bias is not None:
nn.init.zeros_(m.bias)
elif isinstance(m, nn.LayerNorm):
nn.init.ones_(m.weight)
nn.init.zeros_(m.bias)
elif isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode="fan_out")
if m.bias is not None:
nn.init.zeros_(m.bias)
# ------------------------------------------------------------------
def forward(
self, images: torch.Tensor
) -> tuple[torch.Tensor, torch.Tensor, list[torch.Tensor]]:
"""
Args:
images: ``(B, C, H, W)`` fingerprint images.
Returns:
patch_tokens: ``(B, N, D)`` — patch features (196 tokens for 224px).
cls_token: ``(B, D)`` — CLS token (for auxiliary classification).
attn_maps: list of L tensors ``(B, H, N+1, N+1)`` attention weights.
"""
x = self.patch_embed(images) # (B, N, D)
cls = self.cls_token.expand(x.shape[0], -1, -1) # (B, 1, D)
x = torch.cat([cls, x], dim=1) # (B, N+1, D)
x = self.pos_drop(x + self.pos_embed)
attn_maps: list[torch.Tensor] = []
for blk in self.blocks:
x, attn = blk(x)
attn_maps.append(attn)
x = self.norm(x)
return x[:, 1:], x[:, 0], attn_maps
|