SpatialBatchNormalization Issue

[apage224]

SpatialBatchNormalization Issue

[soumith]

@apage2 that sounds strange, happy to investigate. have you tried this code: https://github.com/szagoruyko/cifar.torch and see the same behavior?

[szagoruyko]

I know what's the issue, I forgot to add reset function, preparing a PR

[fmassa]

There is already a PR for that #114

[szagoruyko]

@fmassa it wont work, no check for weight, bias and no running mean and std reset.

[fmassa]

There is no option to remove the weight and bias from cudnn.SpatialBatchNormalization, plus the running mean and std are already set to zeros and ones.

[soumith]

ugh guys we cant have this kind of regressions. can you add tests for a case where we catch this issue?

[szagoruyko]

@fmassa ah right. reset function is not just for ctor, can be called elsewhere.

[soumith]

so should I merge #116?

[fmassa]

I think #116 is fine.
BTW, this regression was introduced in the original BatchNorm 4.0 PR #101

[soumith]

ok there's #116 and there's #114
Sergey, merge whichever is correct, i'll update distro in 30 mins

[szagoruyko]

@soumith let me add a test. we should it to nn too btw.

[soumith]

@apage2 updating your cudnn.torch package via the command: (luarocks install cudnn) should fix things. sorry about that.

[apage224]

Nice, that was fast. I reran and it is working as expected.

Thanks.