VPL for large training datasets

@jiankangdeng
Does VPL training give improvements over Arcface on large training sets?

When we train with 3 Million identities, we observed that only 4% of identities are VPL trained (4% identities have lambda=0.15, rest 96% identities have lambda=0) at allowed_delta=200. In other words, only 4% identities use features saved from past to compute variational prototype , for the rest class centre is used as it is.

Is VPL effective when such low percentage of identities have valid features in queue?
How do we increase this percentage?

Observation:
We experimented with the effect of allowed_delta with on MS1MV3 with 128x4 batch size without momentum.
at allowed_delta=100, around 25% of identities are VPL trained, for the rest lambda=0.
at allowed_delta=200, around 50% of identities are VPL trained
We observed slightly better accuracies at allowed_delta=100.

vpl code from insightface repo uses a concept called VPL momentum. What is its effect? How to use VPL momentum effectively? Batch size reduces by 50% when VPL momentum is used.