Tensorflow 2.0 GPU with JupyterLab docker environment

This repository is an helper for installing a docker environment that provides a tensorflow ecosystem through a JupyterLab interface.

Prerequisites

Linux distribution (Tested on Ubuntu and Arch Linux)
Cuda Drivers for your host GPU
docker
nvidia-docker

What is included

Tmux: Allows to manage multiple shells in the same docker container
Standard Tools: vim, wget
nodejs: jupyterlab extensions dependency
Rclone: Allows file transfer from several cloud services, as Google Drive, Dropbox, etc.
Python packages: requests, scikit-image, scikit-learn, pandas, seaborn, TF probability, SimpleITK (DICOM Imaging Library), jupyterlab, jupyter widgets
Fix for tensorflow docker not being able to hook to host cuda drivers

Setup and First Run

Clone this repository inside the project folder that you wish to mount inside the environment, e.g. ~/MyProject/ProjectEnvironment/
Run build-latest.sh to build the latest stable release of Tensorflow. Other specific versions are available in different scripts.
check that docker images returns the newly created image, e.g. edoardogiacomello/projectenv-latest if you ran build-latest.sh
choose a <name> for your new container and a <port> on your host to access jupyterlab (default is 8888)
run first_run_gpu.py <name> <tag> [<port>]. <tag> tag is the value shown by docker images (e.g. "latest")
Now a new container has been created on your machine and you are into a new shell inside the docker environment. Your project folder (e.g. ~/MyProject/) has been mounted on the folder /tf inside the container.
check if your cuda installation and nvidia-docker is running correctly using nvidia-smi
You can detach from the container (leaving it running) using ctrl + p, ctrl + q

Typical Usage (todo on every reboot)

Check if your container is still running, with docker ps
If it's not running, run it using docker start <name>, using the name you chose in the setup phase.
Attach to the container with docker attach <name>
(Optional) Open a TMUX interface using tmux. This allows to open multiple windows inside the docker shell, as jupyterlab would prevent you from using the shell your run it from.
Run ```jupyter lab --ip=0.0.0.0 --allow-root```` to run the jupyter-lab interface. Take note of the token and keep it secret.
You can now access your jupyter interface from http://<your_ip>:, e.g. http://localhost:8888, using the token provided.
You can detach safely leaving the container running using ctrl + p, ctrl + q. If you need another shell window, you can use tmux commands, such Ctrl+b, Ctrl+c for a new shell, etc.

Configuring File Cloud Access

You can configure Rclone to sync your Google Drive folders to the host machine, for example if you need to download models or datasets.

You can open a new terminal directly from JupyterLab or creating a new tmux window from the container shell.
use rclone config to configure the access to your account. Please refer to the RClone Documentation for your use case.

Launching Tensorboard

Tensorboard can be launched inside the python notebook using %load_ext tensorboard %tensorboard --logdir <path-to-your-logs>

Multiple Containers

If you need to run multiple containers you can launch the script first-run-gpu.py with another name and port. PLEASE NOTICE that the script is made for mounting the parent folder with respect from where it is launched (See above example). You can just edit the script replacing the mounted path $PWD/../ with one of your choice (must be absolute).

Note on GPU Usage: Tensorflow by default allocates ALL available GPU memory to the first user. To avoid this, include this part of code just after the first tensorflow import in your code:

import tensorflow as tf
# Setting allow_growth for gpu
physical_devices = tf.config.experimental.list_physical_devices('GPU')
if len(physical_devices) > 0:
    tf.config.experimental.set_memory_growth(physical_devices[0], True)
else:
    print("No GPU found, model running on CPU")

edoardogiacomello/ProjectsEnvironment