run vgg_16 error about paddle HOT 13 CLOSED

@fbiswt please give me the version of your cuda and cudnn. Thanks.

from paddle.

@qingqing01 cuda 7.5 ,cudnn v4 . Many Thanks

from paddle.

@qingqing01 Should I change the version of cudnn to V5?

from paddle.

@fbiswt It is ok when I tested with cuda 7.5, cudnn v4 or cudnn v5. But my GPU is Tesla K40, what is your GPU type?

from paddle.

@qingqing01 My GPU is also K40, but I install paddle on centos, is this problem?

from paddle.

@qingqing01 When I make paddle ,I got a problem is
CMake Error at cmake/cudnn.cmake:33 (get_filename_component):
get_filename_component unknown component DIRECTORY
Call Stack (most recent call first):
CMakeLists.txt:51 (include)
so I delete the line "get_filename_component(CUDNN_LIB_PATH ${CUDNN_LIBRARY} DIRECTORY)
" in cudnn.make

from paddle.

@qingqing01 I upgrade cmake to cmake3 and solved "get_filename_component unknown component DIRECTORY" error ,and reinstall paddle, and also got that problem

from paddle.

Anyone answer this question ?

from paddle.

@fbiswt I'm sorry for this question and can't give you a solution now. In fact, There are two implementations for BatchNorm. Maybe you can try another implementation. A simple way, you can set batch_norm_type="batch_norm" in batch_norm_layer in file of python/paddle/trainer_config_helpers/layers.py and reinstall paddle.

from paddle.

@qingqing01 Thank you so much for helping me solve this problem ! Now the paddle runing without error!

from paddle.

@qingqing01 Thank you!
I can confirm the same problem happens on Fedora with CUDA 7.5, cuDNN 4, Tesla K80 GPU, and batch_norm_type="batch_norm" works.

from paddle.

@aaronzs & @fbiswt
We check the code, and the bug is paddle use the CUDNN V4 RC not CUDNN V4.
The reason for this bug is that the cudnnBatchNormalizationBackward API is different between the two versions.
We will fix this bug soon.
Before this use cudnn 5.0 is better.

from paddle.

We have fixed this problem, the code is #71

from paddle.