What we wanted?

• Run TensorFlow easily
• Quickly reproducible environment
• Multiple TensorFlow versions and easy upgrades
• Project isolation

Use Docker

• Quickly reproducible environment
• Project isolation
• Hardware isolation

How?

• Just a wrapper around the docker command line tool
• Supports GPU isolation when you have multiple GPUs
• Specialized Docker images by NVIDIA:
  • Ubuntu
  • CentOS

Usage

nvidia-docker run -it tensorflow/tensorflow:latest-gpu

If you don’t want to use the nvidia-docker wrapper, you can add the command line arguments manually:

docker run \
	--device=/dev/nvidiactl \
	--device=/dev/nvidia-uvm \
	--device=/dev/nvidia0 \
	...

Ready-made images

Caffe

https://github.com/BVLC/caffe/

Caffe2

https://hub.docker.com/r/caffe2ai/caffe2/

TensorFlow

https://hub.docker.com/r/tensorflow/tensorflow/

MXNet

https://hub.docker.com/r/mxnet/

PyTorch

https://hub.docker.com/r/pytorch/pytorch/

We forked the TensorFlow image and added support for Python 3 https://hub.docker.com/u/quantlane/tensorflow

How to let a non-admin user use this

1. Add a non-privileged UNIX user…
2. …that is not a member of the docker group.
3. Have a long-running TensorFlow container
4. Allow the non-privileged user to run
   docker exec -it ...
   (E.g. using sudo.)

Future

• NVIDIA Docker 2
  • No wrapper command
  • GPU support for any Docker image
• Experimental support for scheduling GPUs in Kubernetes
  https://kubernetes.io/docs/tasks/manage-gpus/scheduling-gpus/