OS X R2 Tanh and SoftMax tests fail about cudnn.torch

[soumith]

aaaah, it is very likely they have a bug on OSX version. Cant think of any other explanation, as it works cleanly on Linux. you can report bugs to them via the nvidia developer tool.

[szagoruyko]

Yes, on Linux only MaxPooling fails sometimes, as they mention in docs. On OS X actually all ReLU, Tanh, Sigmoid and SofMax fail a lot. Will report a bug.

[szagoruyko]

Just caught the same problem in Linux!

--------------------------------------------------------------------------------
ReLU_single
error on state (backward)
 LT(<) violation   val=nan, condition=0.01
    /usr/local/share/lua/5.1/torch/Tester.lua:26: in function 'assertlt'
    test/test.lua:350: in function <test/test.lua:321>

--------------------------------------------------------------------------------

[soumith]

ok, i am going to run the unit tests a few thousand times and see how that goes.
Also, are you making sure to use Cuda 6.5 on Linux?

[soumith]

can you give me other details about your linux, for a possible reproduction

[szagoruyko]

yes, cuda-6.5, 4 Titan Blacks, 340.29 driver, torch, cutorch, nn and cunn updated to the last version, and also got another machine (mostly equal) on which it fails too, like here Sigmoid:

--------------------------------------------------------------------------------
Sigmoid_single
error on state (backward)
 LT(<) violation   val=nan, condition=0.01
    ...ocks/torch-distro/install/share/lua/5.1/torch/Tester.lua:26: in function 'assertlt'
    test/test.lua:484: in function <test/test.lua:455>

--------------------------------------------------------------------------------

[soumith]

thanks, having a look

[szagoruyko]

it's Ubuntu 14.04 btw, and it fails on all 4 cards in one test, not just one.

[soumith]

ok that's an interesting detail.

[soumith]

ok over several hundred runs, reproduced this once on my Tesla K40. Now trying to print out the specific input shape etc, and reproduce this consistently.

[soumith]

i'm not able to reproduce this even over thousands of runs. i got the nan once, but cant seem to get it again.
can you run this https://github.com/soumith/cudnn.torch/blob/burepro/test/test.lua#L415
And then send me the files badTanh.t7 badSoftmax.t7, badSigmoid.t7 etc.

[szagoruyko]

hm, no luck after 100, will leave it to run overnight

[soumith]

ok i reproduced the nans. it is very likely that cudnn guys are using the fast approximations, so in very extreme precision cases, generating nans. we went down that path in the past and reverted, we have a long history of these things. I will report this to them.

Fast approximations docs:
http://docs.nvidia.com/cuda/cuda-math-api/group__CUDA__MATH__INTRINSIC__SINGLE.html

[philvdm]

I believe it is because you use Beta=0 and it is not handled properly in cuDNN R2 RC for the activation functions.
When beta=0, we are supposed to write directly (without reading the input) because 0 x Nan = Nan
It will be fixed in R2 Final release

[soumith]

thanks @philvdm will wait for the final release.