Computer Vision

Tutorial 2: Reproducible Deep Learning

This lesson introduces essential tools for developing deep learning projects to promote code reproducibility and reusability. We will cover three open-source tools: (1) Git, (2) Docker, and (3) Hydra . The choice of tools is questionable but offers a good compromise between practicality and educational needs.

Local set-up

You can run this project in any local environment. Just as an example, you may choose an Anaconda virtual enviroment. If so, you can create a virtual environment with the following command:

conda create -n reprodl python=3.12 pip; conda activate reprodl

and install all the requirements.

pip install -r requirements.txt

Steps

You can move through existing branches using Git.

git checkout existing_branch

We will end up with the following branches.

Branch	Content
main	It's our starting point, which contains just a notebook.
step1_going_modular	In this step, we will rewrite our notebook to make our code modular. We will end up moving all code in separate files.
step2_docker	We introduce a Dockerfile and isolate our code from the OS.
step3_hydra	In this final step, we want to make our code more easily configurable. We will remove all variables and move them to a dedicated file.

Optional steps

In this lesson, we saw some very simple tools with which you can improve the reproducibility, portability, and reuse of your Deep Learning projects. Remember that there are dozens of other tools for every step of your project. Here are some ones:

• DVC: it allows to track your data changes. You can think about it as a Git for data.
• Docker Hub: you can publish your Docker images in public repositories the same way you publish code on Github.
• Weights&Biases: it is a very simple and powerful tool to log all your experiments and parameters on a dashboard in the cloud.
• MLFlow: it allows to keep track of your model versions as well as logging your training parameters.