Possible bug in using Batch normalisation
mrtpk opened this issue · comments
Thomas Paul commented
-
training
parameter is not given while initialisingtf.layers.batch_normalization
. Thetraining
parameter defaults toFalse
. Hence while training the moving mean and moving variance won't be updated. -
Since this branch's tf api is <tf2,
tf.GraphKeys.UPDATE_OPS
andtrain_op
have to be executed before training [ref]. I could not find these updations in the training code.
Line 163 in becd222
lulupanPPY commented
boy,you saved this code