File size: 8,899 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 | from __future__ import annotations
"""Image-only dataset for the V2 (ViT + TRAM + GNN) pipeline.
Unlike ``FingerprintDataset`` which requires paired minutiae files,
this dataset loads only images + identity labels. No minutiae extractor
is needed — the ViT backbone learns features end-to-end.
Supports directory layouts:
1. ImageFolder: ``root/identity_name/sample.{ext}``
2. PolyU: ``root/{first,second}_session/finger_sample.{ext}``
(identity parsed from filename prefix before last underscore)
"""
import os
import random
from collections import defaultdict
from pathlib import Path
import torch
from torch.utils.data import Dataset, Sampler
from PIL import Image
from .augmentation import build_train_transform, build_val_transform
IMAGE_EXTS = {".bmp", ".png", ".tif", ".tiff", ".jpg", ".jpeg"}
def infer_device_from_name(path_or_name: str) -> str | None:
stem = Path(path_or_name).stem
parts = stem.split("_")
if len(parts) < 4:
return None
token = parts[1]
token_lower = token.lower()
if token_lower in {"roll", "plain"}:
return token_lower
if token.isalpha() and len(token) <= 3:
return token
return None
class ImageDataset(Dataset):
"""Load fingerprint images with identity labels for metric learning.
Returns:
image: ``(1, H, W)`` normalised [0, 1]
label: int
"""
def __init__(
self,
image_dir: str,
image_size: int = 224,
augment: bool = True,
repeat_factor: int = 1,
augment_profile: str = "standard",
):
super().__init__()
self.transform = (
build_train_transform(image_size, profile=augment_profile) if augment
else build_val_transform(image_size)
)
self.repeat_factor = max(1, int(repeat_factor))
self.samples: list[tuple[str, int]] = [] # (path, label)
self.labels: list[int] = []
self.devices: list[str | None] = []
self.label_map: dict[str, int] = {}
self._discover(image_dir)
# ------------------------------------------------------------------
def _discover(self, image_dir: str):
root = Path(image_dir)
if not root.exists():
return
subdirs = sorted([d for d in root.iterdir() if d.is_dir()])
session_like = subdirs and all("session" in d.name.lower() for d in subdirs)
if subdirs and not session_like:
has_images = any(
any(f.suffix.lower() in IMAGE_EXTS for f in d.iterdir() if f.is_file())
for d in subdirs[:5]
)
if has_images:
self._discover_imagefolder(root, subdirs)
return
self._discover_flat(root)
def _append_sample(self, img_path: Path, label: int):
self.samples.append((str(img_path), label))
self.labels.append(label)
self.devices.append(infer_device_from_name(img_path.name))
def _discover_imagefolder(self, root: Path, subdirs: list[Path]):
"""``root/identity/sample.ext`` layout."""
for idx, identity_dir in enumerate(subdirs):
identity = identity_dir.name
self.label_map[identity] = idx
for img_file in sorted(identity_dir.iterdir()):
if img_file.suffix.lower() in IMAGE_EXTS:
self._append_sample(img_file, idx)
def _discover_flat(self, root: Path):
"""Flat/PolyU layout — parse identity from filename."""
all_images: list[Path] = []
for ext in IMAGE_EXTS:
all_images.extend(root.rglob(f"*{ext}"))
all_images = sorted(all_images)
identity_of: dict[str, str] = {}
for img in all_images:
stem = img.stem
parts = stem.rsplit("_", 1)
identity = parts[0] if len(parts) > 1 else stem
identity_of[str(img)] = identity
unique_ids = sorted(set(identity_of.values()))
id_to_label = {name: idx for idx, name in enumerate(unique_ids)}
self.label_map = id_to_label
for img in all_images:
identity = identity_of[str(img)]
label = id_to_label[identity]
self._append_sample(img, label)
# ------------------------------------------------------------------
@property
def num_classes(self) -> int:
return len(self.label_map)
def __len__(self) -> int:
return len(self.samples) * self.repeat_factor
def __getitem__(self, idx: int) -> dict[str, object]:
idx = idx % len(self.samples)
path, label = self.samples[idx]
pil_img = Image.open(path).convert("L")
image = self.transform(pil_img)
return {"image": image, "label": label}
class ImageListDataset(Dataset):
"""Image dataset backed by an explicit list of ``(path, label)`` samples.
Useful for continual learning where the effective training set is assembled
dynamically from the current stage plus replay exemplars from previous stages.
"""
def __init__(
self,
samples: list[tuple[str, int]],
image_size: int = 224,
augment: bool = True,
augment_profile: str = "standard",
):
super().__init__()
self.transform = (
build_train_transform(image_size, profile=augment_profile) if augment
else build_val_transform(image_size)
)
self.samples = [(str(path), int(label)) for path, label in samples]
self.labels = [label for _path, label in self.samples]
self.devices = [infer_device_from_name(path) for path, _label in self.samples]
unique_labels = sorted(set(self.labels))
self.label_map = {str(label): label for label in unique_labels}
@property
def num_classes(self) -> int:
return len(self.label_map)
def __len__(self) -> int:
return len(self.samples)
def __getitem__(self, idx: int) -> dict[str, object]:
path, label = self.samples[idx]
pil_img = Image.open(path).convert("L")
image = self.transform(pil_img)
return {"image": image, "label": label}
class PKSamplerV2(Sampler):
"""P identities × K samples per batch for metric learning."""
def __init__(self, dataset: ImageDataset, p: int = 8, k: int = 4, device_aware: bool = False):
self.p = p
self.k = k
self.device_aware = device_aware
self._len = len(dataset) // (p * k)
self.label_to_indices: dict[int, list[int]] = defaultdict(list)
self.label_to_device_indices: dict[int, dict[str | None, list[int]]] = defaultdict(lambda: defaultdict(list))
for idx, label in enumerate(dataset.labels):
self.label_to_indices[label].append(idx)
self.label_to_device_indices[label][dataset.devices[idx]].append(idx)
self.labels = sorted(self.label_to_indices.keys())
def _sample_indices(self, label: int) -> list[int]:
indices = self.label_to_indices[label]
if not self.device_aware:
return (
random.sample(indices, self.k)
if len(indices) >= self.k
else random.choices(indices, k=self.k)
)
by_device = self.label_to_device_indices[label]
distinct_devices = [dev for dev in by_device if dev is not None]
if len(distinct_devices) <= 1:
return (
random.sample(indices, self.k)
if len(indices) >= self.k
else random.choices(indices, k=self.k)
)
chosen: list[int] = []
device_keys = distinct_devices.copy()
random.shuffle(device_keys)
for dev in device_keys:
if len(chosen) >= self.k:
break
chosen.append(random.choice(by_device[dev]))
remaining_pool = [idx for idx in indices if idx not in chosen]
needed = self.k - len(chosen)
if needed > 0:
if len(remaining_pool) >= needed:
chosen.extend(random.sample(remaining_pool, needed))
else:
chosen.extend(remaining_pool)
if len(chosen) < self.k:
chosen.extend(random.choices(indices, k=self.k - len(chosen)))
return chosen
def __iter__(self):
pool = self.labels.copy()
random.shuffle(pool)
ptr = 0
for _ in range(self._len):
if ptr + self.p > len(pool):
pool = self.labels.copy()
random.shuffle(pool)
ptr = 0
batch_labels = pool[ptr:ptr + self.p]
ptr += self.p
batch: list[int] = []
for lbl in batch_labels:
batch.extend(self._sample_indices(lbl))
yield batch
def __len__(self) -> int:
return max(1, self._len)
|