tensorflow

TensorFlow Docker requirements

1. Install Docker on your local host machine.
2. For GPU support on Linux, install nvidia-docker.

Docker is the easiest way to enable TensorFlow GPU support on Linux since only the NVIDIA® GPU driver is required on the host machine 

(the NVIDIA® CUDA® Toolkit does not need to be installed).

docker run [-it] [--rm] [-p hostPort:containerPort] tensorflow/tensorflow[:tag] [command]

$ docker run -it --rm tensorflow/tensorflow    python -c "import tensorflow as tf; print(tf.__version__)"

Note: nvidia-docker v1 uses the nvidia-docker alias, where v2 uses docker --runtime=nvidia.

CUDA 9.0 for TensorFlow < 1.13.0

nvidia-docker2 intsall

prerequisites:NVIDIA driver and Docker .

If you have a custom /etc/docker/daemon.json, the nvidia-docker2 package might override it.

Ubuntu 14.04/16.04/18.04, 

Ubuntu will install docker.io by default which isn't the latest version of Docker Engine. This implies that you will need to pin the version of nvidia-docker. 

# If you have nvidia-docker 1.0 installed: we need to remove it and all existing GPU containers
docker volume ls -q -f driver=nvidia-docker | xargs -r -I{} -n1 docker ps -q -a -f volume={} | xargs -r docker rm -f
sudo apt-get purge -y nvidia-docker

# Add the package repositories
curl -s -L https://nvidia.github.io/nvidia-docker/gpgkey | \
  sudo apt-key add -
distribution=$(. /etc/os-release;echo $ID$VERSION_ID)
curl -s -L https://nvidia.github.io/nvidia-docker/$distribution/nvidia-docker.list | \
  sudo tee /etc/apt/sources.list.d/nvidia-docker.list
sudo apt-get update

# Install nvidia-docker2 and reload the Docker daemon configuration
sudo apt-get install -y nvidia-docker2
sudo pkill -SIGHUP dockerd

# Test nvidia-smi with the latest official CUDA image
docker run --runtime=nvidia --rm nvidia/cuda:9.0-base nvidia-smi

ref:

https://github.com/NVIDIA/nvidia-docker

CUDA toolkit version	Driver version	GPU architecture
6.5	>= 340.29	>= 2.0 (Fermi)
7.0	>= 346.46	>= 2.0 (Fermi)
7.5	>= 352.39	>= 2.0 (Fermi)
8.0	== 361.93 or >= 375.51	== 6.0 (P100)
8.0	>= 367.48	>= 2.0 (Fermi)
9.0	>= 384.81	>= 3.0 (Kepler)
9.1	>= 387.26	>= 3.0 (Kepler)
9.2	>= 396.26	>= 3.0 (Kepler)
10.0	>= 384.130, < 385.00	Tesla GPUs
10.0	>= 410.48	>= 3.0 (Kepler)

CUDA images come in three flavors and are available through the NVIDIA public hub repository.

• base: starting from CUDA 9.0, contains the bare minimum (libcudart) to deploy a pre-built CUDA application.
  Use this image if you want to manually select which CUDA packages you want to install.
• runtime: extends the base image by adding all the shared libraries from the CUDA toolkit.
  Use this image if you have a pre-built application using multiple CUDA libraries.
• devel: extends the runtime image by adding the compiler toolchain, the debugging tools, the headers and the static libraries.
  Use this image to compile a CUDA application from sources.