1.在进行

gdb python

r XX.py

where

调试时，报出以下错误：

1)每次运行都开38个线程，是否是线程超载[New Thread 0x7ffff2fd2700 (LWP 7415)]

[New Thread0x7ffff27d1700 (LWP 7416)]

[New Thread0x7fffeffd0700 (LWP 7417)]

[New Thread0x7fffeb7cf700 (LWP 7418)]

[New Thread0x7fffe8fce700 (LWP 7419)]

[New Thread0x7fffe67cd700 (LWP 7420)]

[New Thread0x7fffe3fcc700 (LWP 7421)]

[New Thread0x7fffe17cb700 (LWP 7422)]

[New Thread0x7fffdefca700 (LWP 7423)]

[New Thread0x7fffdc7c9700 (LWP 7424)]

[New Thread0x7fffd9fc8700 (LWP 7425)]

[New Thread0x7fffd77c7700 (LWP 7426)]

[New Thread0x7fffd4fc6700 (LWP 7427)]

[New Thread0x7fffd27c5700 (LWP 7428)]

[New Thread0x7fffcffc4700 (LWP 7429)]

[New Thread0x7fffcd7c3700 (LWP 7430)]

[New Thread0x7fffcafc2700 (LWP 7431)]

[New Thread0x7fffc87c1700 (LWP 7432)]

[New Thread0x7fffc5fc0700 (LWP 7433)]

[New Thread0x7fffc37bf700 (LWP 7434)]

[New Thread0x7fffc0fbe700 (LWP 7435)]

[New Thread0x7fffbe7bd700 (LWP 7436)]

[New Thread0x7fffbbfbc700 (LWP 7437)]

[New Thread0x7fffb97bb700 (LWP 7438)]

[New Thread0x7fffb6fba700 (LWP 7439)]

[New Thread0x7fffb47b9700 (LWP 7440)]

[New Thread0x7fffb1fb8700 (LWP 7441)]

[New Thread0x7fffaf7b7700 (LWP 7442)]

[New Thread0x7fffacfb6700 (LWP 7443)]

[New Thread0x7fffaa7b5700 (LWP 7444)]

[New Thread0x7fffa7fb4700 (LWP 7445)]

[New Thread0x7fffa57b3700 (LWP 7446)]

[New Thread0x7fffa2fb2700 (LWP 7447)]

[New Thread0x7fffa07b1700 (LWP 7448)]

[New Thread0x7fff9dfb0700 (LWP 7449)]

[New Thread0x7fff9b7af700 (LWP 7450)]

[New Thread0x7fff98fae700 (LWP 7451)]

[New Thread0x7fff967ad700 (LWP 7452)]

[New Thread0x7fff93fac700 (LWP 7453)]

2)现在报出：

ERROR (theano.gpuarray): Could notinitialize pygpu, support disabled

。。。

File”pygpu/gpuarray.pyx”, line 658, inpygpu.gpuarray.init

File”pygpu/gpuarray.pyx”, line 587, inpygpu.gpuarray.pygpu_init

GpuArrayException: cuDeviceGet: CUDA_ERROR_INVALID_DEVICE: invalid device ordinal

先不解决这个，先尝试测试一下：

发现，在import keras，也会报上述同样的错误！

conda install mkl

conda install mkl-service#使用以上两句均显示：#All requested packages already installed.

conda install blas

依旧不可以导入keras包。

3)将原有的conda环境删除，又新创建了环境，用conda安装了mkl之后，尝试import keras之后，仍然报错：

Using Theano backend.~/lib/python2.7/site-packages/theano/gpuarray/dnn.py:184: UserWarning: Your cuDNN version is more recent than Theano.

If you encounter problems, try updating Theano or downgrading cuDNN to a version >= v5 and <=v7.

warnings.warn(“Your cuDNN version is more recent than”ERROR (theano.gpuarray): Couldnotinitialize pygpu, support disabled

Traceback (most recent call last):

File”~/lib/python2.7/site-packages/theano/gpuarray/__init__.py”, line 227, in use(config.device)

File”~/lib/python2.7/site-packages/theano/gpuarray/__init__.py”, line 214, inuse

init_dev(device, preallocate=preallocate)

File”~/lib/python2.7/site-packages/theano/gpuarray/__init__.py”, line 99, ininit_dev**args)

File”pygpu/gpuarray.pyx”, line 658, inpygpu.gpuarray.init

File”pygpu/gpuarray.pyx”, line 587, inpygpu.gpuarray.pygpu_init

GpuArrayException: cuDeviceGet: CUDA_ERROR_INVALID_DEVICE: invalid device ordinal

在我的.theanorc配置文件中，是这么写的：

[global]

floatX=float32

device=cuda1

尝试去掉cuda编号？居然成功了！

Using Theano backend.~/.conda/envs/xhs/lib/python2.7/site-packages/theano/gpuarray/dnn.py:184: UserWarning: Your cuDNN version is more recent than Theano.

If you encounter problems, try updating Theano or downgrading cuDNN to a version >= v5 and <=v7.

warnings.warn(“Your cuDNN version is more recent than”Using cuDNN version7201on context None

Mapped name None to device cuda: GeForce GTX1080 Ti (0000:03:00.0)

接下来尝试解决 上述的用户警告。

由于theano已经是1.0.4最新版本，无法再进行更新，只能尝试将cuDNN版本降级。

但是使用conda list查看所有安装的包：

cudnn 6.0.21 cuda8.0_0 https://mirrors.tuna.tsinghua.edu.cn/a

#尝试此命令查看pygpu是否可用

DEVICE=”cuda” python -c “import pygpu; pygpu.test()”

此帮助里说，如果不是使用多个GPU可以忽略test_collectives error。

#尝试以下，

python test_gpu.py~/.conda/envs/xhs/lib/python2.7/site-packages/theano/gpuarray/dnn.py:184: UserWarning: Your cuDNN version is more recent than Theano. If you encounter problems, try updating Theano or downgrading cuDNN to a version >= v5 and <=v7.

warnings.warn(“Your cuDNN version is more recent than”Using cuDNN version7201on context None

Mapped name None to device cuda: GeForce GTX1080 Ti (0000:03:00.0)

[GpuElemwise{exp,no_inplace}((float32, vector)>), HostFromGpu(gpuarray)(GpuElemwise{exp,no_inplace}.0)]

Looping1000 times took 0.192847seconds

Resultis [1.2317803 1.6187935 1.5227807 … 2.2077181 2.2996776 1.623233]

Used the gpu

发现其使用的cudnn版本是7.2，明明是6.0但是却调用了7.2？

查看cuda的版本信息发现：

nvcc -V

nvcc: NVIDIA (R) Cuda compiler driver

Copyright (c)2005-2017NVIDIA Corporation

Built on Fri_Sep__1_21:08:03_CDT_2017

Cuda compilation tools, release9.0, V9.0.176

//发现安装cuda简直十分麻烦，所以下尝试一下运行程序。

Starting epoch 0…

段错误 (核心已转储)

#查看分配占空间的大小

ulimit -a#显示

stack size (kbytes, -s) 8192

也就仅仅8M大小，实在是太小了。

改为ulimit -s 102400，仍旧段错误。

试图将其调整为更大或者unlimit时，报错：

-bash: ulimit: stack size: 无法修改 limit 值: 不允许的操作

#使用sudo提示如下：

sudo: ulimit：找不到命令

在limit.conf下加了

#* soft stack unlimited

再使用ulimit -s unlimited就可以用了，但是运行程序发现仍是段错误，继续修改

#max locked memory (kbytes, -l) 64#尝试修改maxloc但是同样的方法不起作用

——————

终于解决了，在github上keras项目下发布的issue中找到了：

由于本机上的CUDA版本为9，所以又根据教程安装了CUDA8版本，以及cuDNN6.0版本，之后就可以了！！！

就是由于CUDA9不适合theano1.0!!!所以必须将版本，降版本之后就没有上述的warning了，就可以成功跑theano后端的keras代码了。