anko-intel commented on issue #14357:
URL:
https://github.com/apache/incubator-mxnet/issues/14357#issuecomment-649069970
Hi @adrianloy,
Using autograd.record() with default value of train_mode = True causes
BatchNorm to use currently given batch data to calculate mean and variance
anko-intel commented on issue #14357:
URL:
https://github.com/apache/incubator-mxnet/issues/14357#issuecomment-647960540
@adrianloy, @PatricZhao,
could you close the issue as the results on master (959143696) are the same
on GPU and CPU (see the comments above for details)?
anko-intel commented on issue #14357:
URL:
https://github.com/apache/incubator-mxnet/issues/14357#issuecomment-645038422
Hi @ThomasDelteil ,
According to the training script from to
https://github.com/apache/incubator-mxnet/issues/14357#issuecomment-497508722,
I didn’t manage to
anko-intel commented on issue #14357:
URL:
https://github.com/apache/incubator-mxnet/issues/14357#issuecomment-644887485
Hi @ThomasDelteil,
According to the training script from
https://github.com/apache/incubator-mxnet/issues/14357#issuecomment-497487102,
As I mentioned in
anko-intel commented on issue #14357:
URL:
https://github.com/apache/incubator-mxnet/issues/14357#issuecomment-644775609
Hi @adrianloy,
Your issue still exist on the 1.6 branch. On master branch „running_var” is
calculated only in backward pass on CPU and GPU backend as well, so your