Feature Proposal: Integrate Hard Negative Mining into Partial FC

Hi authors,
First of all, thank you for your great work on Partial FC and InsightFace!
I’d like to propose a feature enhancement: integrating hard negative mining into Partial FC sampling:

- Maintain a lightweight confusion memory bank that records per-class "confused" predictions (e.g., from top-k logits).
- During sampling, combine:
- Positive class centers
- Confused (hard) negative class centers
- Random negatives (optional for diversity)

What do you think?