Reproducible Deep Learning

PhD Course in Data Science, 2021, 3 CFU

This practical PhD course explores the design of a simple reproducible environment for a deep learning project, using free, open-source tools (Git, DVC, Docker, Hydra, ...). The choice of tools is opinionated, and was made as a trade-off between practicality and didactical concerns.

Local set-up

The use case of the course is an audio classification model trained on the ESC-50 dataset. To set-up your local machine (or a proper virtual / remote environment), configure Anaconda, and create a clean environment:

conda create -n reprodl; conda activate reprodl

⚠️ For an alternative setup without Anaconda, see issue #2.

Then, install a few generic prerequisites (notebook handling, Pandas, …):

conda install -y -c conda-forge notebook matplotlib pandas ipywidgets pathlib

Finally, install PyTorch and PyTorch Lightning. The instructions below can vary depending on whether you have a CUDA-enabled machine, Linux, etc. In general, follow the instructions from the websites.

conda install -y pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorch -c conda-forge
conda install -y pytorch-lightning -c conda-forge

This should be enough to let you run the initial notebook. More information on the use case can be found inside the notebook itself.

⚠️ For Windows only, install a backend for torchaudio:
pip install soundfile

Additional set-up steps

The following steps are not mandatory, but will considerably simplify the experience.

If you are on Windows, install the Windows Subsystem for Linux. This is useful in a number of contexts, including Docker installation.
We will use Git from the command line multiple times, so consider enabling GitHub access with an SSH key.
We will experiment with Docker reproducibility on the Sapienza DGX environment. If you have not done so already, set-up your access to the machine.

Organization of the course

The course is split into exercises (e.g., adding DVC support). The material for each exercise is provided as a Git branch. To follow an exercise, switch to the corresponding branch, and follow the README there. If you want to see the completed exercise, add _completed to the name of the branch. Additional material and information can be found on the main website of the course.

List of exercises:

Experimenting with Git, branches, and scripting (exercise1_git).
Adding Hydra configuration (exercise2_hydra).
Versioning data with DVC (exercise3_dvc).
Creating a Dockerfile (exercise4_docker).
Experiment management with Weight & Biases (exercise5_wandb).
Unit testing and formatting with continuous integration (exercise6_hooks).

An example

If you want to follow the first exercise, switch to the corresponding branch and follow the instructions from there:

git checkout exercise1_git

If you want to see the completed exercise:

git checkout exercise1_git_completed

You can inspect the commits to look at specific changes in the code:

git log --graph --abbrev-commit --decorate

If you want to inspect a specific change, you can checkout again using the ID of the commit.

Contributing

Thanks to Jeroen Van Goey for the error hunting. Feel free to open a pull request if you have suggestions on the current material or ideas for some extra exercises (see below).

⚠️ Because of the sequential nature of the repository, changing something in one of the initial branches might trigger necessary changes in all downstream branches.

Extra material (students & more)

Extra branches contain material that was not covered in the course (e.g., new libraries for hyper-parameter optimization), implemented by the students for the exam. They can be read independently from the main branches. Refer to the original authors for more information.

Author	Branch	Content
OfficiallyDAC	extra_optuna	Fine-tuning hyper-parameters with Optuna.

Advanced reading material

If you liked the exercises and are planning to explore more, the new edition of Full Stack Deep Learning (UC Berkeley CS194-080) covers a larger set of material than this course. Another good resource (divided in small exercises) is the MLOps repository by Goku Mohandas. lucmos/nn-template is a fully-functioning template implementing many of the tools described in this course.

sarvasvarora/reprodl2021