Why do you freeze batch norm parameters when training?

[leao1995]

would it be better to let batch norm parameters adapt to your current data?

[xichangzun]

it's a common practice.First, Because the pretrain network's bn layers have been trained. Second,Object Detection 's batchsize is small, hard to make bn parameter stable.

[prakashjayy]

use Group norm instead of batch norm . it is more stable.

[lxtGH]

Use synchronized batch normalization

[PhilipMay]

Use synchronized batch normalization

Using sync batch norm does not help with single GPU training and low batch sizes though.