因为机器权限限制，已安装了NDVI的驱动、CUDA8.0、cuDNN7.1。
但后续安装tensorflow-gpu都是坑。
如果用pip安装，不同tensorflow-gpu版本对应不同cuDNN,CUDA版本，如果强制安装，会报错各种类似libcuda.so.1包找不到
image.png
即使pip install tensorflow-gpu==1.1.0 降版本安装，也报错/lib64/libc.so.6: version `GLIBC_2.16' not found，需要升级系统。
to make life easy，推荐 一键构建全套的虚拟环境（需要提前安装anaconda3）
conda create --name tf_gpu tensorflow-gpu
然后通过以下命令激活和关闭环境
source activate tf_gpu
source deactivate
但是最坑爹的是该集群上activate都没有授权执行功能。
各种拜托同学后换了一个新的实验环境，CUDA10.0，cuDNN7.4。pip 可直接安装tensorflow-gpu 13.0，太方便了有没有，感慨下大清校园的资源方便。
测试
import tensorflow
import tensorflow as tf
matrix1 = tf.constant([[3., 3.]])
matrix2 = tf.constant([[2.],[2.]])
product = tf.matmul(matrix1, matrix2)
sess = tf.Session()
如果输出GPU信息，则代表安装成功，如下图
2019-03-25 14:19:52.462303: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
2019-03-25 14:19:53.132123: I tensorflow/compiler/xla/service/service.cc:150] XLA service 0x55a65b582100 executing computations on platform CUDA. Devices:
2019-03-25 14:19:53.132194: I tensorflow/compiler/xla/service/service.cc:158] StreamExecutor device (0): Tesla P100-PCIE-16GB, Compute Capability 6.0
2019-03-25 14:19:53.136793: I tensorflow/core/platform/profile_utils/cpu_utils.cc:94] CPU Frequency: 2300190000 Hz
2019-03-25 14:19:53.140707: I tensorflow/compiler/xla/service/service.cc:150] XLA service 0x55a65b6503f0 executing computations on platform Host. Devices:
2019-03-25 14:19:53.140747: I tensorflow/compiler/xla/service/service.cc:158] StreamExecutor device (0): <undefined>, <undefined>
2019-03-25 14:19:53.141484: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1433] Found device 0 with properties:
name: Tesla P100-PCIE-16GB major: 6 minor: 0 memoryClockRate(GHz): 1.3285
pciBusID: 0000:04:00.0
totalMemory: 15.90GiB freeMemory: 15.61GiB
2019-03-25 14:19:53.141528: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1512] Adding visible gpu devices: 0
2019-03-25 14:19:53.152529: I tensorflow/core/common_runtime/gpu/gpu_device.cc:984] Device interconnect StreamExecutor with strength 1 edge matrix:
2019-03-25 14:19:53.152579: I tensorflow/core/common_runtime/gpu/gpu_device.cc:990] 0
2019-03-25 14:19:53.152596: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1003] 0: N
2019-03-25 14:19:53.153330: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1115] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 15190 MB memory) -> physical GPU (device: 0, name: Tesla P100-PCIE-16GB, pci bus id: 0000:04:00.0, compute capability: 6.0)
题外话：
pip安装太慢？换下镜像
pip install -i https://pypi.tuna.tsinghua.edu.cn/simple tensorflow-gpu keras
conda下载也很慢？
conda config --add channels 'https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free/'
conda config --set show_channel_urls yes
后记
良好的习惯很重要。需要构建多个虚拟环境对python开发环境进行管理。
比如virtualenv（利用pip安装软件）和anaconda自带的conda工具都可以方便构建虚拟环境。
