wenmin-wu/cv-lap-hard-negative-mining

LAP Hard Negative Mining

Overview

In metric learning, hard negatives (close but different-class examples) drive the most learning. Random negatives are too easy; per-anchor hardest negatives cause collapse. LAP (linear assignment problem) via the Jonker-Volgenant algorithm finds a globally optimal one-to-one assignment that maximizes overall difficulty across the entire batch, avoiding degenerate pairings.

Quick Start

import numpy as np
from lap import lapjv

def mine_hard_negatives(score_matrix, labels, t2i):
    cost = -score_matrix.copy()
    # Block same-class pairs with high cost
    for cls_indices in labels.values():
        idxs = [t2i[t] for t in cls_indices]
        for i in idxs:
            for j in idxs:
                cost[i, j] = 10000.0

    _, _, col_assignment = lapjv(cost)
    hard_pairs = []
    for j, i in enumerate(col_assignment):
        hard_pairs.append((i, j))
        cost[i, j] = 10000.0
        cost[j, i] = 10000.0
    return hard_pairs

Workflow

1. Compute pairwise similarity/score matrix from current embeddings
2. Mask same-class pairs with large cost to prevent them being selected
3. Run lapjv to find optimal one-to-one hard negative assignment
4. Use assigned pairs for contrastive/siamese training
5. Recompute assignments periodically as embeddings evolve

Key Decisions

• LAP vs random: LAP finds globally hard negatives; random wastes training on easy pairs
• LAP vs per-anchor hardest: per-anchor can cause model collapse; LAP distributes difficulty
• Recompute frequency: every epoch or every N batches — stale assignments degrade quality
• lap package: pip install lap for fast C++ Jonker-Volgenant solver

References

• Siamese (pretrained) 0.822