Explanation¶
Explanation answers why rather than how. These articles cover design decisions, trade-offs, and the non-obvious behavior behind the library.
Articles¶
- Why Imbalanced Losses — why standard cross-entropy fails under class imbalance, and what each loss in this library actually optimizes
- Memory Queue Design — why a circular buffer is necessary for ranking losses at low positive rates, and how it interacts with DDP
- Temperature and Soft Ranking — how the sigmoid approximation replaces the discrete rank function and what temperature controls
- DDP All-Gather and Gradients — why rank-based losses are biased under standard DDP sharding and how all-gather fixes this without breaking autograd