info8010-deep-learning: a PyTorch course delivered as markdown and notebooks

A university course you can read without enrolling

The repository is the course content for INFO8010, a deep learning course run at ULiège by Gilles Louppe, with Fanny Bodart, Elise Faulx, Julien Brandoit and Sacha Peters listed as teaching assistants. The agenda covers Spring 2026, with sessions on Friday at 8:30 AM in the B28 Mania Pavella amphitheater, and a Discord invite for questions. The course is graded through a project with a proposal deadline, a final deadline in mid-May and a poster session afterwards, plus two optional homeworks that each account for 5 percent of the final grade when submitted. None of that requires a university account to read. The lecture notes are written in markdown at the top level of the repository, the slides are mirrored as PDF files, and the code that accompanies each lecture is committed as a notebook. A self-learner can therefore follow the same sequence a student follows, minus the deadline and the grader.

Every lecture is a markdown file paired with a notebook

The top level of the repository is the course outline itself. Files named lecture0.md through lecture12.md hold the notes, and the agenda table links each one to a rendered page and to a PDF under pdf/. Alongside them sit course-syllabus.md, project.md, closing.md, and two files named archives-lecture-gan.md and archives-lecture-rnn.md. Directories cover the rest: code/ for notebooks, homeworks/ for the two assignments, plus assets/, figures/, poster/ and a template-report.tex used for writing the project report. The rendering layer lives at glouppe.github.io/info8010-deep-learning, where the markdown is served with a query parameter such as ?p=lecture3.md, so the same source produces the web page and the printable slide deck. One detail in the code/ layout stands out. Most lectures map to a single notebook, but the GPT lecture is a directory named code/gpt/ rather than one file, which suggests the from-scratch implementation is split into several pieces meant to be read in order.

Getting the code: clone the repository and open the notebooks

There is no pip install step to copy, and the project does not print one at the top of the page. Instead, the assignment section points readers to the homeworks folder for installation instructions, and it says plainly that the goal of the two assignments is to get you familiar with the PyTorch library. That folder is the place to start, and it is the only place where a working setup is described. A typical session is: clone the repository, follow the installation instructions found in the homeworks folder, then open the notebook that matches the lecture you are reading. The paths below are the actual entry points you will open, taken from the code directory:

code/lec1-polynomial-regression.ipynb
code/lec3-autodiff.ipynb
code/gpt/

The order of work follows the agenda: polynomial regression first, automatic differentiation in PyTorch later, and the GPT implementation last. Homework 2 assumes a working GPU setup and covers datasets, the Dataloader, running on GPU and training a convolutional neural network, so a functioning accelerator is expected by that point in the course.

The parked RNN and GAN lectures show the course expiring in place

Under an archived lectures heading, the project states that some lectures have become less relevant because of progress in the field, and it still hosts them for anyone interested. The two topics named in the repository file names are recurrent neural networks and generative adversarial networks. That admission is more useful than it first looks, because the same repository also holds a lecture on attention and transformers dated March 20 in the same agenda that still lists the older lecture notes. Each academic year also gets its own branch, and the tags follow the same pattern: v8-info8010-2025 for the 2024-2025 edition, v7-info8010-2024 for 2023-2024, and v6-info8010-2023 for 2022-2023, with earlier editions going back to v2 for 2018-2019. Older branches are not deleted, so a reader can compare a lecture between two years and see how the framing changed.

No autograder, no environment file, no solutions in the repository

This is where the project stops short of what a self-paced learning platform usually offers. The two homeworks are submitted on Gradescope, an external service tied to the course, and the grading workflow lives there rather than in the repository. There is no test suite, no continuous integration configuration and no visible dependency lock in the top level listing, so a reader who wants a reproducible environment must assemble one. There is also no solution set. The assignment notebooks ask for tensor operations, use of autograd and nn, and then a full convolutional training pipeline, which means the reader is expected to write the code rather than compare it with a reference. Project work follows the same pattern: instructions live in project.md and the report is expected to follow template-report.tex. The upside is that nothing in the repository is hidden behind a course registration. The downside is that nobody verifies your work for you.

How this differs from a packaged online course

The usual alternative is a hosted course platform that ships a prebuilt environment, a video per module and an automatic grader that marks exercises without a teacher ever looking at them. The approach here differs in three concrete ways. First, the course files are versioned by academic year in Git branches, so you can read exactly what a given cohort received rather than a moving web page. Second, the assessment is human and institutional: assignments are handed in on Gradescope and the project ends in a poster session, which means the notebooks are designed for a teacher to read, not for a script to mark. Third, the theory is written out as prose in markdown files rather than delivered as slides alone, so the explanation of why a technique works sits next to the code that implements it. If your priority is immediate feedback with zero setup, this route is slower. If your priority is a full semester of theory and code in one place, the notes and the notebooks are already written that way.

Releases, the BSD-3-Clause licence, and what the git history shows

The project carries a BSD-3-Clause licence, which in short terms permits reuse and redistribution with the notice conditions attached, and it is worth reading the LICENSE file yourself if you intend to republish the notebooks or slides. The repository is not archived, and the last push was on 2026-07-13. Published releases trail the teaching year rather than the code: the v8-info8010-2025 release is dated 2025-05-23, v7-info8010-2024 came on 2025-02-10, and v6-info8010-2023 on 2024-02-05. That gap is explained by the branching strategy, where each new edition starts from the previous one, and it means the master branch carries the current course while older editions stay reachable but frozen. Upgrading cost is low for readers, since you can pin a tag and follow one edition, and the format of the course means a new academic year largely rewrites notes and adds notebooks rather than changing any published interface.

Editorial conclusion

Adopt this repository if you want a structured, PyTorch-first path through the standard deep learning curriculum and are willing to supply your own environment and do your own grading. Skip it if you need a pinned dependency lock, automated assessment, or slides you can query from a single file. Before starting, open the homeworks folder and check the installation instructions yourself, then decide whether the lecture sequence fits the theory you already know.

Frequently asked questions

Official sources

1. glouppe/info8010-deep-learning on GitHub
2. Issues
3. License: BSD-3-Clause
4. README
5. Releases