Spaces:
Paused
Paused
File size: 10,446 Bytes
8bcf79a 25bdf34 |
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 |
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Tuple, List, Optional
import numpy as np
class SemiHardTripletMiner:
"""Semi-hard negative mining for triplet loss training."""
def __init__(self, margin: float = 0.2):
self.margin = margin
def mine_triplets(
self,
embeddings: torch.Tensor,
labels: torch.Tensor
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""
Mine semi-hard triplets from embeddings.
Args:
embeddings: (N, D) tensor of normalized embeddings
labels: (N,) tensor of labels
Returns:
anchors, positives, negatives: (K, D) tensors where K is number of valid triplets
"""
# Compute pairwise distances
dist_matrix = self._compute_distance_matrix(embeddings)
# Find valid triplets
anchors, positives, negatives = self._find_semi_hard_triplets(
dist_matrix, labels
)
if len(anchors) == 0:
# Fallback to random triplets if no semi-hard ones found
return self._random_triplets(embeddings, labels)
return embeddings[anchors], embeddings[positives], embeddings[negatives]
def _compute_distance_matrix(self, embeddings: torch.Tensor) -> torch.Tensor:
"""Compute pairwise cosine distances between embeddings."""
# Normalize embeddings to unit length
embeddings = F.normalize(embeddings, p=2, dim=1)
# Compute cosine similarity matrix
similarity_matrix = torch.mm(embeddings, embeddings.t())
# Convert to distance matrix (1 - similarity)
distance_matrix = 1 - similarity_matrix
return distance_matrix
def _find_semi_hard_triplets(
self,
dist_matrix: torch.Tensor,
labels: torch.Tensor
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""Find semi-hard negative triplets."""
anchors = []
positives = []
negatives = []
n = len(labels)
for i in range(n):
anchor_label = labels[i]
# Find positive samples (same label)
positive_mask = (labels == anchor_label) & (torch.arange(n, device=labels.device) != i)
positive_indices = torch.where(positive_mask)[0]
if len(positive_indices) == 0:
continue
# Find negative samples (different label)
negative_mask = labels != anchor_label
negative_indices = torch.where(negative_mask)[0]
if len(negative_indices) == 0:
continue
# For each positive, find semi-hard negative
for pos_idx in positive_indices:
pos_dist = dist_matrix[i, pos_idx]
# Find negatives that are harder than positive but not too hard
# Semi-hard: pos_dist < neg_dist < pos_dist + margin
neg_dists = dist_matrix[i, negative_indices]
semi_hard_mask = (neg_dists > pos_dist) & (neg_dists < pos_dist + self.margin)
semi_hard_indices = torch.where(semi_hard_mask)[0]
if len(semi_hard_indices) > 0:
# Choose the hardest semi-hard negative
hardest_idx = semi_hard_indices[torch.argmax(neg_dists[semi_hard_indices])]
neg_idx = negative_indices[hardest_idx]
anchors.append(i)
positives.append(pos_idx)
negatives.append(neg_idx)
if len(anchors) == 0:
return torch.tensor([], dtype=torch.long), torch.tensor([], dtype=torch.long), torch.tensor([], dtype=torch.long)
return torch.tensor(anchors), torch.tensor(positives), torch.tensor(negatives)
def _random_triplets(
self,
embeddings: torch.Tensor,
labels: torch.Tensor
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""Generate random triplets as fallback."""
anchors = []
positives = []
negatives = []
n = len(labels)
max_triplets = min(1000, n // 3) # Limit number of random triplets
for _ in range(max_triplets):
# Random anchor
anchor_idx = torch.randint(0, n, (1,)).item()
anchor_label = labels[anchor_idx]
# Random positive (same label)
positive_mask = (labels == anchor_label) & (torch.arange(n, device=labels.device) != anchor_idx)
positive_indices = torch.where(positive_mask)[0]
if len(positive_indices) == 0:
continue
pos_idx = positive_indices[torch.randint(0, len(positive_indices), (1,))].item()
# Random negative (different label)
negative_mask = labels != anchor_label
negative_indices = torch.where(negative_mask)[0]
if len(negative_indices) == 0:
continue
neg_idx = negative_indices[torch.randint(0, len(negative_indices), (1,))].item()
anchors.append(anchor_idx)
positives.append(pos_idx)
negatives.append(neg_idx)
if len(anchors) == 0:
# Last resort: duplicate first sample
return embeddings[:1], embeddings[:1], embeddings[:1]
return torch.tensor(anchors), torch.tensor(positives), torch.tensor(negatives)
class OnlineTripletMiner:
"""Online triplet mining for batch training."""
def __init__(self, margin: float = 0.2, mining_strategy: str = "semi_hard"):
self.margin = margin
self.mining_strategy = mining_strategy
self.semi_hard_miner = SemiHardTripletMiner(margin)
def mine_batch_triplets(
self,
embeddings: torch.Tensor,
labels: torch.Tensor
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""
Mine triplets from a batch of embeddings.
Args:
embeddings: (B, D) tensor of normalized embeddings
labels: (B,) tensor of labels
Returns:
anchors, positives, negatives: (K, D) tensors
"""
if self.mining_strategy == "semi_hard":
return self.semi_hard_miner.mine_triplets(embeddings, labels)
elif self.mining_strategy == "hardest":
return self._hardest_triplets(embeddings, labels)
elif self.mining_strategy == "random":
return self._random_batch_triplets(embeddings, labels)
else:
raise ValueError(f"Unknown mining strategy: {self.mining_strategy}")
def _hardest_triplets(
self,
embeddings: torch.Tensor,
labels: torch.Tensor
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""Find hardest negative triplets."""
dist_matrix = self._compute_distance_matrix(embeddings)
anchors = []
positives = []
negatives = []
n = len(labels)
for i in range(n):
anchor_label = labels[i]
# Find positive samples
positive_mask = (labels == anchor_label) & (torch.arange(n, device=labels.device) != i)
positive_indices = torch.where(positive_mask)[0]
if len(positive_indices) == 0:
continue
# Find negative samples
negative_mask = labels != anchor_label
negative_indices = torch.where(negative_mask)[0]
if len(negative_indices) == 0:
continue
# For each positive, find hardest negative
for pos_idx in positive_indices:
pos_dist = dist_matrix[i, pos_idx]
# Find hardest negative (closest to anchor)
neg_dists = dist_matrix[i, negative_indices]
hardest_idx = torch.argmin(neg_dists)
neg_idx = negative_indices[hardest_idx]
# Only include if negative is closer than positive + margin
if neg_dists[hardest_idx] < pos_dist + self.margin:
anchors.append(i)
positives.append(pos_idx)
negatives.append(neg_idx)
if len(anchors) == 0:
return self._random_batch_triplets(embeddings, labels)
return torch.tensor(anchors), torch.tensor(positives), torch.tensor(negatives)
def _random_batch_triplets(
self,
embeddings: torch.Tensor,
labels: torch.Tensor
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""Generate random triplets from batch."""
return self.semi_hard_miner._random_triplets(embeddings, labels)
def _compute_distance_matrix(self, embeddings: torch.Tensor) -> torch.Tensor:
"""Compute pairwise cosine distances."""
embeddings = F.normalize(embeddings, p=2, dim=1)
similarity_matrix = torch.mm(embeddings, embeddings.t())
distance_matrix = 1 - similarity_matrix
return distance_matrix
def create_triplet_miner(
strategy: str = "semi_hard",
margin: float = 0.2
) -> OnlineTripletMiner:
"""Factory function to create a triplet miner."""
return OnlineTripletMiner(margin=margin, mining_strategy=strategy)
# Example usage
if __name__ == "__main__":
# Test with dummy data
batch_size = 32
embed_dim = 128
num_classes = 8
# Generate dummy embeddings and labels
embeddings = torch.randn(batch_size, embed_dim)
labels = torch.randint(0, num_classes, (batch_size,))
# Create miner
miner = create_triplet_miner(strategy="semi_hard", margin=0.2)
# Mine triplets
anchors, positives, negatives = miner.mine_batch_triplets(embeddings, labels)
print(f"Generated {len(anchors)} triplets from batch of {batch_size}")
print(f"Anchor indices: {anchors[:5]}")
print(f"Positive indices: {positives[:5]}")
print(f"Negative indices: {negatives[:5]}")
|