Training with your config, you use norm_eval=True to freeze bn, why not settingnormalize=dict(type='BN', frozen=True)? Is there any difference？
ZhihuaGao
All 2 comments
frozen: stop gradient update in norm layers
norm_eval: stop moving average statistics update in norm layers
thangvubk
on 26 Feb 2019
👍6
Why need to stop moving average statistics update in norm layers when training model? As far as I know，BN layers usually need to record moving average statistics which is to be used when testing model.
blateyang
on 6 Jul 2019
Was this page helpful?
0 / 5 - 0 ratings
Related issues
Hemantr05
·
3Comments
tianxinhang
·
3Comments
happog
·
3Comments
fatLime
·
3Comments
michaelisc
·
3Comments
Most helpful comment
frozen: stop gradient update in norm layersnorm_eval: stop moving average statistics update in norm layers