youngminpil/rollback

Backbone Can Not be Trained at Once: Rolling Back to Pre-trained Network for Person Re-Identification (AAAI2019)

rollback's Issues

Some doubts about training process

Hi, it is an interesting and enlightening work. But i have some doubts, first, why don't you just set the learning rate of block 4,3,2 to zero to intently train the block 1. It may be a substitute of rolling back block 2,3,4. second, why do not join triplet loss in the training process.
TAHNKS for your time to read this message.

where is the model.py for network?

from model import ft_net

where is ft_net ?

Thanks