testBNInMultiGPU

test Batch Normalization in multi GPU

syncBatchNorm in validSyncBN.py is most robust.

result

testBNInSingleGPU():

outputArray = [0. 0. 0.]

After normalization:
moving mean = [0.1 0.1 0.1]
moving variance = [0.9 0.9 0.9]

testBNInMultiGPU1():

outputArray = [0. 0. 0.]

After normalization:
moving mean = [0.2 0.2 0.2]
moving variance = [0.79999995 0.79999995 0.79999995]

testBNInMultiGPU2():

outputArray = [0. 0. 0.]

After normalization:
moving mean = [0.3 0.3 0.3]
moving variance = [0.79999995 0.79999995 0.79999995]

blog writen in Chinese

[TensorFlow]Batch Normalization在单机多卡上的一些实验