Deploy docker for DL project

Transcript

1. DevOps - Docker for DL environment docker + jupyter + screen + ssh
2. Scenarios
3. • DL environment relies on system hardware ◦ CUDA ◦ cuDNN • DL framework relies on system library ◦ tensorflow 1.4.1 > CUDA 8 ◦ tensorflow 1.5 > CUDA 9 • Isolate the project environment ◦ python2 vs. python3 ◦ python3.5 vs. python3.6 ◦ different package requirement across the different project Why should I use it? hardware CUDA, cuDNN keras, tensorflow ... python package isolation
4. Ubuntu Server Docker Image screen / tmux data code mount port binding :8888 :9888 local port forwarding :9888 browser preview
5. Introduction
6. Solution - Docker Traditional Virtual Machine Docker Container Reference: What is a Container?
7. Docker instruction $ docker images # list the images
8. Docker instruction $ docker rmi beetle-action:py36 # delete image by name $ docker rmi fe2255258306　# delete image by id
9. Docker instruction $ docker ps # list the container $ docker ps -a # list all containers $ docker rm c0aa8680536f # delete container
10. Script
11. Ubuntu Server Docker Image screen / tmux data code mount port binding :8888 :9888 local port forwarding :9888 browser preview
12. Activate Docker and Mount the project $ screen -R project-docker # Enter the screen $ nvidia-docker run \ -v ~/myProject:/root/project \ -v ~/myData:/root/project/data \ -v ~/myOutput:/root/project/output \ -p 9888:8888 \ -p 7006:6006 \ -it myDockerImage /bin/bash mount code / data/ output mount port • jupyter 8888 • tensorboard 6006 the command after the docker image activation e.g. nvidia/cuda:latest interactive mode Reference: docker run | Docker Documentation
13. Ubuntu Server Docker Image screen / tmux data code mount port binding :8888 :9888 local port forwarding :9888 browser preview
14. Activate jupyter in Docker image Because we are the root user in the docker image, we should activate jupyter with the paramter --allow-root $ jupyter notebook --allow-root
15. SSH - port forwarding $ ssh -NfL 9888:localhost:9888 USER@SERVER-IP If the connection still unreachable, please check the port usage with the following instruction in each phase $ netstat -altnp