1. 程式人生 > >tesnorflow Batch Normalization

tesnorflow Batch Normalization

graph point ilo eps net store scope psi bat

1.train或者從checkpoint restore後發現moving_mean和moving_variance都是0和1

bn1_mean = graph.get_tensor_by_name("bn1/moving_mean/read:0")

"bn1/moving_variance:0"

將updates_collections=None即可

net = slim.batch_norm(net,epsilon=0.001,updates_collections=None,scale=True,activation_fn=nn.relu,scope = ‘bn1‘)

tesnorflow Batch Normalization