search

szagoruyko / imagine-nn

IMAGINE torch neural network routines
Other
109 stars 35 forks source link

SpatialCrossResponseNormalization bug #12

Closed qas32 closed 8 years ago

qas32 commented 9 years ago

When the input tensor is too big, number of blocks exceeds 65535, which causes crashes when compiling with arch=sm_20. That causes issues when trying to loadcaffe VGG_M network with 'cudnn', which uses inn.SCRN .

Any reason for not using sm_35 ?

szagoruyko commented 9 years ago

you can just use cudnn. I will update loadcaffe in a few days to support R3.

szagoruyko commented 8 years ago

fixed by https://github.com/szagoruyko/imagine-nn/pull/16