UFR-Fing / src /data /enhanced_sampler.py
anbinh39's picture
Add files using upload-large-folder tool
dadf189 verified
Raw
History Blame Contribute Delete
16.2 kB
from __future__ import annotations
"""Enhanced PK Sampler with guaranteed cross-device sampling and hard negative mining.
PKSamplerV3 addresses limitations of the original PKSampler:
- Guarantees cross-device positives (when available)
- Filters identities with insufficient samples
- Robust fallback strategies for edge cases
- Better hard negative mining through label diversity
"""
import random
from collections import defaultdict
from torch.utils.data import Sampler
from .image_dataset import ImageDataset, ImageListDataset
class PKSamplerV3(Sampler):
"""Enhanced PK sampler for metric learning with cross-device guarantees.
Samples P identities × K samples per batch, with special handling for:
- Cross-device sampling: Ensures samples from different devices within each identity
- Hard negatives: Maximizes label diversity across batches
- Quality filtering: Only uses identities with sufficient samples
Parameters
----------
dataset : ImageDataset
The fingerprint dataset.
p : int
Number of identities per batch (P).
k : int
Number of samples per identity (K).
ensure_cross_device : bool
If True, prioritize sampling from different devices within each identity.
min_devices_per_identity : int
Minimum number of distinct devices required for an identity to be viable
when ensure_cross_device=True.
hard_negative_ratio : float
Currently unused. Reserved for future hard negative mining based on
pre-computed similarity matrix.
Notes
-----
Batch size = P × K (e.g., 8 identities × 4 samples = 32).
"""
def __init__(
self,
dataset: ImageDataset,
p: int = 8,
k: int = 4,
ensure_cross_device: bool = True,
min_devices_per_identity: int = 2,
hard_negative_ratio: float = 0.5,
):
self.p = p
self.k = k
self.ensure_cross_device = ensure_cross_device
self.min_devices_per_identity = min_devices_per_identity
self.hard_negative_ratio = hard_negative_ratio
# Build index structures
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)
device = dataset.devices[idx]
self.label_to_device_indices[label][device].append(idx)
# Filter viable labels based on cross-device requirements
if ensure_cross_device:
self.viable_labels = self._filter_viable_labels_cross_device()
else:
self.viable_labels = self._filter_viable_labels_basic()
if len(self.viable_labels) < p:
print(
f"⚠️ Warning: Only {len(self.viable_labels)} viable labels with "
f"cross-device requirement (need {p}). Relaxing constraints..."
)
# Fallback: Use all labels with at least k samples
self.viable_labels = [
lbl
for lbl, indices in self.label_to_indices.items()
if len(indices) >= k
]
if len(self.viable_labels) < p:
print(
f"⚠️ Critical: Only {len(self.viable_labels)} labels with ≥{k} samples. "
f"Using all available labels."
)
self.viable_labels = list(self.label_to_indices.keys())
# Compute number of batches
total_samples = sum(
len(self.label_to_indices[lbl]) for lbl in self.viable_labels
)
self._len = max(1, total_samples // (p * k))
print(
f"PKSamplerV3: {len(self.viable_labels)} viable labels, "
f"{self._len} batches per epoch (P={p}, K={k})"
)
# ------------------------------------------------------------------
def _filter_viable_labels_cross_device(self) -> list[int]:
"""Filter labels that have sufficient cross-device samples."""
viable: list[int] = []
for label, dev_dict in self.label_to_device_indices.items():
# Count distinct non-None devices
distinct_devices = [dev for dev in dev_dict if dev is not None]
num_devices = len(distinct_devices)
# Check total samples
total_samples = sum(len(indices) for indices in dev_dict.values())
if num_devices >= self.min_devices_per_identity and total_samples >= self.k:
viable.append(label)
return viable
# ------------------------------------------------------------------
def _filter_viable_labels_basic(self) -> list[int]:
"""Filter labels with at least K samples."""
return [
lbl
for lbl, indices in self.label_to_indices.items()
if len(indices) >= self.k
]
# ------------------------------------------------------------------
def _sample_cross_device_indices(self, label: int) -> list[int]:
"""Sample K indices ensuring cross-device diversity.
Strategy:
1. Identify available devices for this identity
2. Distribute K samples across devices as evenly as possible
3. Fill remaining with random samples if needed
Args:
label: Identity label to sample from.
Returns:
List of K sample indices.
"""
by_device = self.label_to_device_indices[label]
devices = [
dev for dev in by_device if dev is not None and len(by_device[dev]) > 0
]
# Fallback: Not enough devices, use random sampling
if len(devices) < 2:
indices = self.label_to_indices[label]
if len(indices) >= self.k:
return random.sample(indices, self.k)
else:
return random.choices(indices, k=self.k)
# Strategy: Distribute K samples across devices
chosen: list[int] = []
devices_shuffled = devices.copy()
random.shuffle(devices_shuffled)
# Compute samples per device
k_per_device = max(1, self.k // len(devices))
for dev in devices_shuffled:
if len(chosen) >= self.k:
break
available = by_device[dev]
n_sample = min(k_per_device, len(available), self.k - len(chosen))
if n_sample > 0:
sampled = (
random.sample(available, n_sample)
if len(available) >= n_sample
else available
)
chosen.extend(sampled)
# Fill remaining slots
if len(chosen) < self.k:
remaining_pool = [
idx for idx in self.label_to_indices[label] if idx not in chosen
]
needed = self.k - len(chosen)
if len(remaining_pool) >= needed:
chosen.extend(random.sample(remaining_pool, needed))
else:
# Last resort: Add all remaining + duplicate from existing
chosen.extend(remaining_pool)
while len(chosen) < self.k:
chosen.append(random.choice(self.label_to_indices[label]))
return chosen[: self.k]
# ------------------------------------------------------------------
def _sample_basic_indices(self, label: int) -> list[int]:
"""Sample K indices without device constraints."""
indices = self.label_to_indices[label]
if len(indices) >= self.k:
return random.sample(indices, self.k)
else:
return random.choices(indices, k=self.k)
# ------------------------------------------------------------------
def _create_batch(self) -> list[int]:
"""Create one batch with P identities × K samples.
Returns:
List of sample indices for this batch.
"""
# Randomly select P identities
if len(self.viable_labels) >= self.p:
batch_labels = random.sample(self.viable_labels, self.p)
else:
# Sample with replacement if not enough labels
batch_labels = random.choices(self.viable_labels, k=self.p)
# Sample K indices per identity
batch: list[int] = []
for label in batch_labels:
if self.ensure_cross_device:
indices = self._sample_cross_device_indices(label)
else:
indices = self._sample_basic_indices(label)
batch.extend(indices)
return batch
# ------------------------------------------------------------------
def __iter__(self):
"""Yield batches for one epoch."""
for _ in range(self._len):
yield self._create_batch()
def __len__(self) -> int:
return self._len
class ContinualReplayPKSampler(Sampler):
"""PK sampler that mixes current-stage labels with replay labels per batch.
The combined dataset is expected to contain:
1. Current-stage samples in the range ``[0, current_size)``
2. Replay exemplar samples in the range ``[current_size, len(dataset))``
Each batch draws ``current_p`` identities from the new stage and
``replay_p`` identities from replay memory, which reduces abrupt domain
shift between stages and gives old identities direct rehearsal batches.
"""
def __init__(
self,
dataset: ImageDataset | ImageListDataset,
current_size: int,
p: int = 8,
k: int = 4,
replay_p: int = 2,
ensure_cross_device: bool = True,
min_devices_per_identity: int = 2,
):
self.dataset = dataset
self.current_size = max(0, int(current_size))
self.p = int(p)
self.k = int(k)
self.ensure_cross_device = ensure_cross_device
self.min_devices_per_identity = min_devices_per_identity
replay_size = max(0, len(dataset.samples) - self.current_size)
if replay_size <= 0:
replay_p = 0
replay_p = max(0, min(int(replay_p), self.p - 1))
self.replay_p = replay_p
self.current_p = self.p - self.replay_p
self.current_label_to_indices: dict[int, list[int]] = defaultdict(list)
self.current_label_to_device_indices: dict[int, dict[str | None, list[int]]] = (
defaultdict(lambda: defaultdict(list))
)
self.replay_label_to_indices: dict[int, list[int]] = defaultdict(list)
self.replay_label_to_device_indices: dict[int, dict[str | None, list[int]]] = (
defaultdict(lambda: defaultdict(list))
)
for idx, label in enumerate(dataset.labels[: self.current_size]):
self.current_label_to_indices[label].append(idx)
device = dataset.devices[idx]
self.current_label_to_device_indices[label][device].append(idx)
for idx, label in enumerate(
dataset.labels[self.current_size :], start=self.current_size
):
self.replay_label_to_indices[label].append(idx)
device = dataset.devices[idx]
self.replay_label_to_device_indices[label][device].append(idx)
self.current_labels = self._filter_viable_labels(
self.current_label_to_indices,
self.current_label_to_device_indices,
)
self.replay_labels = self._filter_viable_labels(
self.replay_label_to_indices,
self.replay_label_to_device_indices,
)
if len(self.current_labels) == 0:
raise ValueError(
"ContinualReplayPKSampler requires at least one current-stage label"
)
if len(self.replay_labels) == 0:
self.replay_p = 0
self.current_p = self.p
total_current_samples = sum(
len(self.current_label_to_indices[label]) for label in self.current_labels
)
self._len = max(1, total_current_samples // max(1, self.current_p * self.k))
print(
f"ContinualReplayPKSampler: current_labels={len(self.current_labels)} "
f"replay_labels={len(self.replay_labels)} current_p={self.current_p} "
f"replay_p={self.replay_p} batches={self._len}"
)
def _filter_viable_labels(
self,
label_to_indices: dict[int, list[int]],
label_to_device_indices: dict[int, dict[str | None, list[int]]],
) -> list[int]:
viable: list[int] = []
for label, indices in label_to_indices.items():
if len(indices) < self.k:
continue
if not self.ensure_cross_device:
viable.append(label)
continue
distinct_devices = [
device
for device, device_indices in label_to_device_indices[label].items()
if device is not None and len(device_indices) > 0
]
if len(distinct_devices) >= self.min_devices_per_identity:
viable.append(label)
if viable:
return viable
return [
label
for label, indices in label_to_indices.items()
if len(indices) >= self.k
] or list(label_to_indices.keys())
def _sample_indices(
self,
label: int,
label_to_indices: dict[int, list[int]],
label_to_device_indices: dict[int, dict[str | None, list[int]]],
) -> list[int]:
indices = label_to_indices[label]
if not self.ensure_cross_device:
return (
random.sample(indices, self.k)
if len(indices) >= self.k
else random.choices(indices, k=self.k)
)
by_device = label_to_device_indices[label]
devices = [
device
for device in by_device
if device is not None and len(by_device[device]) > 0
]
if len(devices) < self.min_devices_per_identity:
return (
random.sample(indices, self.k)
if len(indices) >= self.k
else random.choices(indices, k=self.k)
)
chosen: list[int] = []
devices_shuffled = devices.copy()
random.shuffle(devices_shuffled)
for device in devices_shuffled:
if len(chosen) >= self.k:
break
chosen.append(random.choice(by_device[device]))
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)
while len(chosen) < self.k:
chosen.append(random.choice(indices))
return chosen[: self.k]
def _sample_labels(self, labels: list[int], n_labels: int) -> list[int]:
if n_labels <= 0:
return []
if len(labels) >= n_labels:
return random.sample(labels, n_labels)
return random.choices(labels, k=n_labels)
def __iter__(self):
for _ in range(self._len):
batch: list[int] = []
for label in self._sample_labels(self.current_labels, self.current_p):
batch.extend(
self._sample_indices(
label,
self.current_label_to_indices,
self.current_label_to_device_indices,
)
)
for label in self._sample_labels(self.replay_labels, self.replay_p):
batch.extend(
self._sample_indices(
label,
self.replay_label_to_indices,
self.replay_label_to_device_indices,
)
)
yield batch
def __len__(self) -> int:
return self._len