Open justusschock opened 6 years ago
Using fc5 as embedding is just ok. AngleLoss can reserve many many gaps in 512 high-dimensional space. Maybe unseen persons will be projected to the gap.
Hi @justusschock @clcarwin ! Can you please provide more details about how to use the AngleLoss with embedding for solving recognition problem?
Hi @clcarwin , your code seems to train a classifier. Is there any possibility only to train an embedding for later recognition of unseen persons with just a little modification of your code?