OEAW Summer School on AI 2021

13.09 - 17.09. 2021 in Vienna, Austria, official website here.

Tuesday lecture session

Date: 14.09.2021, 09:00 - 17:30
Tutor: Pavol Harar (orcid)
Assistent: Leon Gerard (orcid)
Title: A tool a day keeps the bad review away
Alt-title: Planning, execution, and dissemination of a machine learning project
Slides: presentation.pdf

Tuesday exercise session

The main aim of the following exercises is to have a hands on experience with some of the tools presented during the lecture session. The following material builds on the jupyter notebook from a previous session in which you have built and trained some neural networks using the FashionMNIST dataet. All the commands bellow are suited for Ubuntu operating system but links to relevant resources are provided for users of other operating systems.

By the end of the day, you should be able to:

1. run a machine learning project within a virtual environment on your own computer
2. put your code into a git repository like this one
3. specify, build, and run a Docker container which runs your project
4. track your experiments in a nice and useful web interface
5. wrap your project into an interactive web application for common users
6. configure your project to run on Binder, e.g. for reviewers of your paper
7. deploy the web application such that it is available from the Internet

---

Exercise 1 - Jupyter in Virtualenv

• Make a volume folder.
• Inside the volume folder, make a project folder which will be the root of your project. Do the rest of the steps in the root of the project.
• Download the Solution_Exercise_6.ipynb from Monday session.
• Create requirements.txt.
• Install & create a virtualenv in .venv folder.
• Install dependencies from requirements.txt. (The torch+cpu package is not available for Mac users, so in case you have a Mac, remove the first line of the requirements.txt where the -f flag is specified and adjust the lines of torch and torchvision to not mention cpu. The torch+cpu package should then later be used within the Docker based on Ubuntu and also on Binder and Heroku. Torch which supports both cpu and gpu computation is quite large ~1GB and Heroku would complain about the size quota. You can alternatively keep two version of the file, e.g. 1. requirements.txt and 2. requirements-mac.txt)
• Run the jupyter notebook server.
• Run the cell 1 to see whether all imports work.
• If you run cells 1-4 the FashionMNIST data will be downloaded into volume > data folder which is needed only when you want to train the network. In the following exercises we will only need the pretrained models from Monday.
• Create a readme.md file with just a name of the project.

---

Exercise 2 - Git

Make sure you can authenticate with git using ssh key

• Create an account on github.com.
• Go to github.com > settings > SSH and GPG keys.
• Copy your public ssh key from the output of cat ~/.ssh/id_rsa.pub.
• Add it to the github keys and verify with ssh -T git@github.com.
• In case of different OS or some problems consult the github guide.

Create a new repository, initialize & push your project to origin

• Go to github.com and create a new repository oeawss21.
• Do not check the automatic creation of readme, license or other files.
• Go to your project folder.
• Create .gitignore file with .venv, .ipynb_checkpoints in it.
• Run git init.
• Run git remote add origin git@github.com:<your_username>/oeawss21.git.
• If you use git for the first time, you might be asked to configure your user name and commit email address with:
  • Run git config --global user.name "Your Name".
  • Run git config --global user.email "Your Name".
• Run git add ..
• Run git commit -m "Initial commit".
• Run git push origin master.
• Now your changes should be visible in your repository on github.com.
• In case github or you changed your HEAD branch from master to main, change the commands accordingly to avoid problems.

More in-depth but still absolutely simple guide on git:

• https://rogerdudler.github.io/git-guide/

---

Exercise 3 - Docker

Install Docker

• If you use different OS than Ubuntu, check Docker installation guide.
• On Ubuntu install with sudo apt install docker.io
• Check if it is installed correctly with sudo docker run hello-world

Pull Ubuntu minimal docker image

• Run sudo docker pull intelliseqngs/ubuntu-minimal-20.04:3.0.5.
• Add a file .dockerfile into your project. (Here we use a nonstandard name for a reason that we actually do not want Binder and Heroku to use our Dockerfile.)
• Base your .dockerfile on intelliseqngs/ubuntu-minimal-20.04:3.0.5.
• Fill the .dockerfile with commands to copy and install your project.
• Reference on writing the Dockerfile is here.
• In case you have problems, consult the solution in .dockerfile.

Build the image

• Run sudo docker build -t oeawss21:latest -f .dockerfile ..
• In case your Docker errors on "killed" Adjust Docker Preferences Resources RAM - make it bigger, i.e. 4 or 6GB in the settings of your Docker.
• Run sudo docker run --rm -v <path_on_host>/volume/data:/home/volume/data -p 8888:8888 oeawss21:latest jupyter notebook --allow-root --ip 0.0.0.0.
  • -p forwards port 8888 of the container to 8888 on the host
  • -v mounts specified host folder to container's folder (paths must be absolute)
  • --allow-root since all in the container runs as root
  • --ip 0.0.0.0 expose the jupyter server so host can see it
• Visit localhost:8888 in your browser and copy the token, the jupyter notebook should now run.

---

Exercise 4 - Weights and Biases

• Create an account at wandb.ai.
• Log in to your account and try the Example (wandb.me/intro) and run it until the "Run experiment" cell finishes.
• Check the results in the wandb.ai account.
• Check also these examples https://github.com/wandb/examples.
• If you feel motivated, change the Solution_Exercise_6.ipynb such that it tracks the training into wandb and view the results in the web interface.

---

Exercise 5 - Voila

• Go back to the notebook which is running in virtualenv.
• Create a new python3 notebook called webapp.ipynb.
• Copy trained models trained_cnn.pt, trained_model.pt from Monday to the project root.
• Build an interactive webapp using ipywidgets which allows user to specify a pretrained model to be used and image url to be specified. Then loads the selected pretrained model, computes and displays the prediction to the user.
• Click on Voila button in the jupyter notebook menu to test whether everything runs as a web app.
• Update the .dockerfile to include web app related files (webapp.ipynb, trained_model.pt, trained_cnn.pt) and rebuild your Docker image if you want to have a functional Docker container with a webapp inside.
• Push your changes to git.

---

Exercise 6 - Binder

• Make your git repo public if it is not already.
• Go to mybinder.org and fill in the form:
  • Repository URL: https://github.com/<your_username>/oeawss21
  • Git ref: master (or main depending on your repo)
  • Path to a URL (not a file): /voila/render/webapp.ipynb
• Copy the binder markup badge into your readme.md.
• Wait for app to run in Binder. It will take quite some time, but Binder is a free service, so...
• Check how to use Docker with binder if needed here.
• When the app runs, have fun... You can try this image for example.

*An example badge to run the webapp from this repository on Binder is bellow the title of this exercise. Try to click it.

---

Exercise 7 - Heroku [optional]

• Create a free account on Heroku. It might still ask you to fill in your credit card though.
• Add Procfile with web: voila webapp.ipynb --no-browser --port $PORT.
• Add runtime.txt into project folder with python-3.8.10.
• Push changes to the repo.
• Install Heroku cli by following the official guide.
• Deploy your app to heroku using git:
  • Run heroku update to make sure Heroku cli is up to date
  • Run heroku create to create a new Heroku app
  • Run git push heroku master to deploy.
  • Optionally set heroku ps:scale web=1.
  • Openy your app with heroku open.
• Or follow the deployment guide directly from Voila.

---

Some useful links

• How to use Docker with GPU
• Example on how to better structure the project
• Free Heroku slug must be at most 500MB