目前提供的镜像ubuntu-gpu-mxnet1.3、ubuntu-gpu-tf1.12，似乎有一些问题，尝试修复了一下

[icemoon1987]

#116 为了解决这个issue，做了一些尝试。

采用这个Dockerfile构建镜像，可以正常运行nvidia-smi，并可以正常进入GPU模型训练阶段。

主要修改：

1. 依赖于提供的GPU镜像；（避免重新构建所有环境）
2. 删除GPU镜像中的cuda和nvidia相关库；（解决nvidia-driver冲突）
3. 采用Nvidia官方的方式，重新安装用到的cuda计算库；（解决安装cuda大包联带安装nvidia-driver，解决cuda 9库小版本问题）
4. 补充必要的环境变量，是的nvidia-docker加载宿主机nvidia-driver相关库和工具。（解决nvidia-docker启动镜像问题）
5. 稍微修改pip脚本，使其可用。（解决pip问题）

[CLAassistant]

Thank you for your submission, we really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

EC2 Default User seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}