File size: 10,298 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 | """
Centralized configuration for the fingerprint graph model.
All model, backbone, training, data, loss, and inference settings are defined
here so that experiments can be reproduced by swapping a single config object.
"""
from dataclasses import dataclass, field
from typing import Literal
# βββββββββββββββββββββββββββββββββββββββββββββ Graph ββββββ
@dataclass
class GraphConfig:
"""Dynamic k-NN graph construction."""
k: int = 10
dynamic_graph: bool = True
distance_metric: Literal["euclidean", "cosine"] = "euclidean"
# βββββββββββββββββββββββββββββββββββββββββββββ RPE ββββββββ
@dataclass
class RelationalPEConfig:
"""Pairwise relational PE: (dx, dy, d, cos a, sin a, cos dtheta, sin dtheta)."""
input_dim: int = 7
hidden_dim: int = 64
output_dim: int = 64
num_layers: int = 2
activation: str = "gelu"
# βββββββββββββββββββββββββββββββββββββββββββββ Attention ββ
@dataclass
class AttentionConfig:
"""Local multi-head attention on k-NN graph."""
num_heads: int = 4
head_dim: int = 64
dropout: float = 0.1
# βββββββββββββββββββββββββββββββββββββββββββββ Pooling ββββ
@dataclass
class PoolingConfig:
"""Global pooling over variable-length minutiae sets."""
method: Literal["meanmax", "attentive", "multihead"] = "attentive"
num_heads: int = 4
hidden_dim: int = 256
# βββββββββββββββββββββββββββββββββββββββββββββ Backbone βββ
@dataclass
class BackboneConfig:
"""FLaRE CNN backbone settings."""
num_in: int = 1 # grayscale
extract_layer: str = "layer2" # 128D, H/4
image_size: tuple[int, int] = (256, 256) # (H, W)
# βββββββββββββββββββββββββββββββββββββββββββββ Sampler ββββ
@dataclass
class SamplerConfig:
"""Bilinear feature sampling at minutiae locations."""
append_geometry: bool = True # concat cos theta, sin theta
# βββββββββββββββββββββββββββββββββββββββββββββ Loss βββββββ
@dataclass
class ArcFaceConfig:
"""ArcFace (additive angular margin) loss."""
scale: float = 32.0
margin: float = 0.50
easy_margin: bool = False
@dataclass
class TripletConfig:
"""Triplet loss with hard mining."""
margin: float = 0.3
mining: Literal["hard", "semihard", "all"] = "semihard"
@dataclass
class LossConfig:
"""Combined loss wrapper."""
arcface: ArcFaceConfig = field(default_factory=ArcFaceConfig)
triplet: TripletConfig = field(default_factory=TripletConfig)
arcface_weight: float = 1.0
triplet_weight: float = 1.0
# βββββββββββββββββββββββββββββββββββββββββββββ Data βββββββ
@dataclass
class AugmentConfig:
"""Joint augmentation for image + minutiae."""
rotate: bool = True
rotate_range: float = 180.0
translate: bool = True
translate_range: float = 10.0
jitter_std: float = 2.0
minutia_dropout: float = 0.15
min_keep: int = 5
spurious_rate: float = 0.08
@dataclass
class DataConfig:
"""Dataset & dataloader parameters."""
image_dir: str = "data/images"
image_size: tuple[int, int] = (256, 256)
num_workers: int = 4
pin_memory: bool = True
augment: AugmentConfig = field(default_factory=AugmentConfig)
# βββββββββββββββββββββββββββββββββββββββββββββ Train ββββββ
@dataclass
class SchedulerConfig:
"""Warmup + cosine annealing schedule."""
warmup_epochs: int = 5
min_lr: float = 1e-6
@dataclass
class TrainConfig:
"""Training hyper-parameters."""
epochs: int = 100
batch_size: int = 32
lr: float = 3e-4
weight_decay: float = 1e-4
optimizer: Literal["adamw", "sgd"] = "adamw"
grad_clip: float = 1.0
scheduler: SchedulerConfig = field(default_factory=SchedulerConfig)
mixed_precision: bool = False
seed: int = 42
save_dir: str = "checkpoints"
log_every: int = 50
eval_every: int = 1
# βββββββββββββββββββββββββββββ Master Config βββββββββββββ
@dataclass
class Config:
"""Root configuration β single import gives access to everything."""
backbone: BackboneConfig = field(default_factory=BackboneConfig)
sampler: SamplerConfig = field(default_factory=SamplerConfig)
graph: GraphConfig = field(default_factory=GraphConfig)
rpe: RelationalPEConfig = field(default_factory=RelationalPEConfig)
attention: AttentionConfig = field(default_factory=AttentionConfig)
pooling: PoolingConfig = field(default_factory=PoolingConfig)
embed_dim: int = 256
num_layers: int = 6
output_dim: int = 192
train: TrainConfig = field(default_factory=TrainConfig)
loss: LossConfig = field(default_factory=LossConfig)
data: DataConfig = field(default_factory=DataConfig)
def get_default_config() -> Config:
"""Return a fresh default configuration."""
return Config()
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# ViT-Graph (Option C) configuration
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
@dataclass
class ViTConfig:
"""Pretrained ViT backbone settings.
Default: DINOv2 ViT-B/14 (768-D embeddings).
With image_size=224 and patch_size=14 β 16Γ16 = 256 patch tokens.
"""
model_name: str = "vit_base_patch14_dinov2.lvd142m"
pretrained: bool = True
freeze: bool = True
image_size: int = 224
@dataclass
class TRAMConfig:
"""TRAM multilayer centrality token selection.
K tokens are selected from the ViT patch grid based on attention
centrality β analogous to 30-80 minutiae on a fingerprint.
"""
num_tokens: int = 30
power_iterations: int = 10
layer_weights: Literal["uniform", "last_heavy", "exponential"] = "uniform"
@dataclass
class GridRPEConfig:
"""Grid-position relational PE (5-dim: Ξrow, Ξcol, dist, cos Ξ±, sin Ξ±)."""
input_dim: int = 5
hidden_dim: int = 64
output_dim: int = 64
num_layers: int = 2
activation: str = "gelu"
@dataclass
class ViTGraphConfig:
"""Root configuration for the ViT-Graph model.
Separate from Config (MDGT) β these two models can coexist.
Reuses GraphConfig, AttentionConfig, PoolingConfig, TrainConfig,
LossConfig, DataConfig from MDGT where applicable.
"""
vit: ViTConfig = field(default_factory=ViTConfig)
tram: TRAMConfig = field(default_factory=TRAMConfig)
grid_rpe: GridRPEConfig = field(default_factory=GridRPEConfig)
graph: GraphConfig = field(default_factory=lambda: GraphConfig(k=9))
attention: AttentionConfig = field(default_factory=AttentionConfig)
pooling: PoolingConfig = field(default_factory=PoolingConfig)
embed_dim: int = 256
num_layers: int = 3 # fewer than MDGT (ViT features are rich)
output_dim: int = 192
train: TrainConfig = field(default_factory=TrainConfig)
loss: LossConfig = field(default_factory=LossConfig)
data: DataConfig = field(default_factory=DataConfig)
def get_vit_graph_config() -> ViTGraphConfig:
"""Return a fresh ViT-Graph default configuration."""
return ViTGraphConfig()
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# MDGT v2 (ViT from scratch + TRAM + GNN) β docx plan
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
@dataclass
class V2TrainConfig:
"""Two-phase training configuration for MDGT v2.
Phase 1: Freeze ViT backbone, train GNN + pooling + heads.
Phase 2: Unfreeze ViT with smaller LR, train end-to-end.
"""
# Phase 1 β frozen ViT
phase1_epochs: int = 50
phase1_lr_gnn: float = 1e-3
phase1_batch_size: int = 64
# Phase 2 β end-to-end
phase2_epochs: int = 150
phase2_lr_vit: float = 1e-5
phase2_lr_gnn: float = 1e-4
phase2_batch_size: int = 32
# Shared
weight_decay: float = 0.05
warmup_epochs: int = 5
grad_clip: float = 1.0
seed: int = 42
save_dir: str = "checkpoints_v2"
log_every: int = 50
eval_every: int = 1
num_workers: int = 4
pin_memory: bool = True
@dataclass
class V2Config:
"""Root configuration for MDGT v2 pipeline.
Matches the docx plan: ViT-Tiny + TRAM + 2-layer GAT + multi-head pool.
"""
vit_variant: Literal["tiny", "small", "base"] = "tiny"
image_size: int = 224
tram_k: int = 30
gnn_layers: int = 2
gnn_dim: int = 0 # 0 = auto (192 for tiny, 256 for base)
gnn_heads: int = 4
gnn_k: int = 5
pool_heads: int = 4
output_dim: int = 256
drop_rate: float = 0.0
drop_path_rate: float = 0.1
# Loss
arcface_scale: float = 30.0
arcface_margin: float = 0.50
triplet_margin: float = 0.3
triplet_weight: float = 0.1
cls_weight: float = 0.1
# Training
train: V2TrainConfig = field(default_factory=V2TrainConfig)
def get_v2_config() -> V2Config:
"""Return a fresh MDGT v2 default configuration."""
return V2Config()
|