测试vgg_16_cifar.py报错

[quietsmile]

ubuntu 14.04， cuda 7.5, cudnn 5.1.5 安装成功
但是跑demo/image_classification/train.sh时报错，错误信息如下：

[INFO 2016-08-31 17:20:21,497 layers.py:1430] channels=512 size=8192
[INFO 2016-08-31 17:20:21,497 layers.py:1430] output size for conv_8 is 4
[INFO 2016-08-31 17:20:21,498 layers.py:1430] channels=512 size=8192
[INFO 2016-08-31 17:20:21,499 layers.py:1430] output size for conv_9 is 4
[INFO 2016-08-31 17:20:21,501 layers.py:1490] output size for pool_3 is 2_2
[INFO 2016-08-31 17:20:21,502 layers.py:1490] output size for pool_4 is 1_1
[INFO 2016-08-31 17:20:21,507 networks.py:960] The input order is [image, label]
[INFO 2016-08-31 17:20:21,507 networks.py:963] The output order is [cost_0]
I0831 17:20:21.523936 13974 Trainer.cpp:169] trainer mode: Normal
I0831 17:20:21.546594 13974 PyDataProvider2.cpp:219] loading dataprovider image_provider::processData
[INFO 2016-08-31 17:20:21,682 image_provider.py:52] Image size: 32
[INFO 2016-08-31 17:20:21,682 image_provider.py:53] Meta path: data/cifar-out/batches/batches.meta
[INFO 2016-08-31 17:20:21,682 image_provider.py:58] DataProvider Initialization finished
I0831 17:20:21.682675 13974 PyDataProvider2.cpp:219] loading dataprovider image_provider::processData
[INFO 2016-08-31 17:20:21,682 image_provider.py:52] Image size: 32
[INFO 2016-08-31 17:20:21,682 image_provider.py:53] Meta path: data/cifar-out/batches/batches.meta
[INFO 2016-08-31 17:20:21,682 image_provider.py:58] DataProvider Initialization finished
I0831 17:20:21.683006 13974 GradientMachine.cpp:134] Initing parameters..
I0831 17:20:22.312453 13974 GradientMachine.cpp:141] Init parameters done.
.........
I0831 17:20:52.894659 13974 TrainerInternal.cpp:162] Batch=100 samples=12800 AvgCost=2.35864 CurrentCost=2.35864 Eval: classification_error_evaluator=0.833906 CurrentEval: classification_error_evaluator=0.833906
.........
I0831 17:21:00.884374 13974 TrainerInternal.cpp:162] Batch=200 samples=25600 AvgCost=2.15774 CurrentCost=1.95684 Eval: classification_error_evaluator=0.792148 CurrentEval: classification_error_evaluator=0.750391
.........
I0831 17:21:08.731333 13974 TrainerInternal.cpp:162] Batch=300 samples=38400 AvgCost=2.01417 CurrentCost=1.72705 Eval: classification_error_evaluator=0.753672 CurrentEval: classification_error_evaluator=0.676719
.........I0831 17:21:15.873359 13974 TrainerInternal.cpp:179] Pass=0 Batch=391 samples=50048 AvgCost=1.90795 Eval: classification_error_evaluator=0.71814
F0831 17:21:18.497601 13974 hl_cuda_cudnn.cc:779] Check failed: CUDNN_STATUS_SUCCESS == cudnnStat (0 vs. 5) Cudnn Error: CUDNN_STATUS_INVALID_VALUE
*** Check failure stack trace: ***
@ 0x7f609f255daa (unknown)
@ 0x7f609f255ce4 (unknown)
@ 0x7f609f2556e6 (unknown)
@ 0x7f609f258687 (unknown)
@ 0x8a98d4 hl_convolution_forward()
@ 0x5c66fc paddle::CudnnConvLayer::forward()
@ 0x62305c paddle::NeuralNetwork::forward()
@ 0x6b54af paddle::Tester::testOneBatch()
@ 0x6b5dc2 paddle::Tester::testOnePeriod()
@ 0x69a28c paddle::Trainer::trainOnePass()
@ 0x69d687 paddle::Trainer::train()
@ 0x53b0b3 main
@ 0x7f609e461ec5 (unknown)
@ 0x546695 (unknown)
@ (nil) (unknown)

更改cudnn版本，5.0.5， 4.0.4错误都一样~
求助！

[reyoung]

Please use command paddle version to print compile flag, and paste them here~~ Thanks.

[gangliao]

Hi, Can you post your GPU type name? For instance, K40?

[quietsmile]

PaddlePaddle 0.8.0b, compiled with
with_avx: ON
with_gpu: ON
with_double: OFF
with_python: ON
with_rdma: OFF
with_glog: ON
with_gflags: ON
with_metric_learning:
with_timer: OFF
with_predict_sdk:

gtx titanx, driver 352.39

[qingqing01]

@quietsmile Hi, there is no problem when we tested on Tesla K20/K40 with cuda 7.5 and cudnn 5.1, cudnn 4.0. But we don't have gtx titanx environment and wasn't able to to replicate this problem. We will solve it later.

[wangjiangb]

I have added a change list to fix it.

[qingqing01]

@quietsmile We have fixed this problem in GTX 980, see 341486d .

[hedaoyuan]

Fixed #107, and close issue.