fastai/fastbook: the Deep Learning for Coders notebooks, and how to run them

What fastbook is, and what it is not

fastbook is a repository of Jupyter notebooks, not a Python package you install and import. The README describes the notebooks as covering "an introduction to deep learning, fastai, and PyTorch", and says they are used for a MOOC and form the basis of the book Deep Learning for Coders with fastai and PyTorch. The top-level entries are the chapters themselves, numbered 01_intro.ipynb through 20_conclusion.ipynb, plus two appendix notebooks, app_jupyter.ipynb and app_blog.ipynb.

The audience is someone who wants to learn by running code, not someone who needs a dependency. There is no library API to call. When you clone this repository you get teaching material: markdown explanations interleaved with executable cells, an images/ directory, a clean/ directory, a tools/ directory, a translations/ directory, and a utils.py helper module. The two releases listed for the project, v0.0.18 and v0.0.19, both landed on 2022-04-16, which tells you the versioned releases are old even though the repository itself has been pushed since.

If you arrived here expecting a software product, you are in the wrong place. The name is also shared with unrelated things: the search data around this project includes queries about a fastbook app, a fastbook login, a fastbook marketplace and a Treasury fastbook, none of which have anything to do with Jeremy Howard and Sylvain Gugger's notebooks.

How the notebooks and the fastai layer relate

The mechanism here is a stack of three layers, and the book walks down through them. At the top is fastai, described in the README as "a layered API for deep learning". In the middle is PyTorch. At the bottom, in the later chapters, is the mathematics and the raw implementation.

You can see that descent in the file names. Chapters 1 through 10 stay at the application level: image classification, production deployment, ethics, MNIST, pet breeds, multi-category targets, resizing and test-time augmentation, collaborative filtering, tabular data, and NLP. Chapter 11 is titled midlevel_data, which is where the mid-level data API is introduced. From there the notebooks move into internals: 12_nlp_dive.ipynb, 13_convolutions.ipynb, 14_resnet.ipynb, 15_arch_details.ipynb, 16_accel_sgd.ipynb on optimizers and callbacks, 17_foundations.ipynb, 18_CAM.ipynb on GradCAM, and 19_learner.ipynb.

That ordering is the actual design of the material. A learner starts by calling a high-level fastai function, and by chapter 19 has rebuilt enough of the learner to understand what that function did. The repository layout also shows the maintenance surface: utils.py, settings.ini and environment.yml sit alongside the notebooks, and the clean/ and tools/ directories support the build rather than the teaching.

Reading fastbook in Colab instead of installing it

The README is explicit that Colab is the recommended path: "Instead of cloning this repo and opening it on your machine, you can read and work with the notebooks using Google Colab", and it adds that this is the recommended approach for people just getting started because there is no need to set up a Python environment locally.

The README provides a direct Colab link per chapter. The pattern is the standard GitHub-to-Colab URL, and the appendix notebook for Jupyter is the gentlest entry point:

https://colab.research.google.com/github/fastai/fastbook/blob/master/app_jupyter.ipynb

Opening that link loads the notebook from the master branch. The chapter links follow the same shape, for example 01_intro.ipynb and 05_pet_breeds.ipynb in place of app_jupyter.ipynb. If you only want to read rather than run, the README points to a selection of chapters readable online at fastai.github.io/fastbook2e/.

Installing fastbook locally from requirements.txt

If you do want a local environment, the repository ships requirements.txt and environment.yml rather than an installer. The README gives no step-by-step local setup instructions, so the file contents are the only concrete guide. requirements.txt lists fastai>=2.0.0, graphviz, ipywidgets, matplotlib, nbdev>=0.2.12, pandas, scikit_learn, azure-cognitiveservices-search-imagesearch and sentencepiece.

A minimal local path is to clone the repository, create an environment, and install those requirements before launching Jupyter:

git clone https://github.com/fastai/fastbook.git
cd fastbook
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
jupyter notebook

After the install finishes, Jupyter opens in your browser and the chapter notebooks appear in the file listing. Two things in that list are worth noting before you run anything. The pin is fastai>=2.0.0, a floor rather than a ceiling, so pip will resolve to a current fastai release that may not behave exactly as the notebook text describes. nbdev>=0.2.12 is present because the book itself is built with nbdev, not because a learner needs it. And azure-cognitiveservices-search-imagesearch belongs to the image-downloading material in chapter 2, which means that notebook expects a Bing image search key; the README does not document how to supply one.

The licence is split, and that is the real constraint

This is the part most readers skim and should not. The README states that the code in the notebooks and the python .py files is covered by the GPL v3 license, while "the remainder (including all markdown cells in the notebooks and other prose) is not licensed for any redistribution or change of format or medium", apart from copying the notebooks or forking the repository for private use. Commercial and broadcast use are excluded.

The repository's licence field is reported as NOASSERTION, which matches that split rather than contradicting it: there is no single SPDX identifier that describes both halves. The README also notes that the published book "does not have the same GPL restrictions that are on this repository", so the purchased edition and the repository are not interchangeable from a rights standpoint.

Practically, this means you can read the prose and run the code, and you can fork the repository for your own private work. You cannot take the markdown explanations and republish them, reformat them, or build a commercial course from them. The README goes further and asks readers to tell anyone hosting a copy elsewhere that their actions are not allowed and may lead to legal action. Contributions carry their own term: the README states that pull requests assign copyright of that work to Jeremy Howard and Sylvain Gugger. This is a description of what the files say, not legal advice; if your use is commercial, read the LICENSE file and take your own advice.

Where fastbook is the wrong tool

Three situations make fastbook a poor fit. The first is treating it as a dependency. There is no fastbook package to pin in a requirements file, and the repository's own versioned releases stopped at v0.0.19 on 2022-04-16. If you need fastai in production, you install fastai, not this.

The second is expecting the notebooks to track the current fastai API. The README pins fastai>=2.0.0, and the notebook text was written against a specific point in fastai's history. The repository has been pushed since the last release, but the release history gives no evidence that the notebook content was re-verified against newer fastai versions. A cell that worked when the chapter was written may need adjusting now, and the README does not promise otherwise.

The third is licensing. If you want a permissively licensed deep learning curriculum you can adapt, remix and redistribute, this is not it. The GPL covers the code and the prose is explicitly restricted. An organisation that needs to embed teaching material in an internal platform with redistribution rights should look elsewhere before investing time in these notebooks.

Alternatives, and how they differ in approach

The most direct alternative for the same subject matter is the fastai documentation and its own course materials, referenced from the README at docs.fast.ai and course.fast.ai. The difference is scope and form: fastbook is a narrative sequence of notebooks that builds from high-level calls down to internals across twenty chapters, while the documentation is reference material organised around the API. If you want to look up what a function does, the documentation is the right place. If you want the guided path that ends with you understanding why the function exists, that is what the numbered notebooks provide.

A second comparison is the published book itself. The README says the book is available for purchase and explicitly does not carry the same GPL restrictions as the repository. For a reader who needs to quote passages, teach from the text, or work without a browser and a GPU session, the book removes the licensing constraint that the repository imposes. The trade-off is that the book is static while the notebooks are executable.

A third option, for anyone whose real goal is a stable library rather than a course, is to skip the teaching layer entirely and work with PyTorch directly. That is a different kind of decision: you lose the fastai abstractions and the guided progression, and you gain a dependency that is maintained as a library rather than as a book.

FAQ: what people actually ask about fastbook

The questions below come from search data about this project. Some of the searches attached to the name fastbook refer to unrelated products, and those are not answered here because the repository says nothing about them.

Editorial conclusion

Adopt fastbook if you are learning deep learning through fastai and PyTorch and you want the notebook code that matches the course and the book. Do not adopt it if you need a maintained library to import, a permissively licensed teaching corpus you can rebrand, or a reference for the current fastai API, since the last code release was v0.0.19 on 2022-04-16. Before you start, verify the licence boundary in the LICENSE file against your intended use, and check the notebook you plan to open against the fastai version installed by requirements.txt, since the repository pins fastai>=2.0.0 rather than a fixed release.

Frequently asked questions

Official sources

1. fastai/fastbook on GitHub
2. Issues
3. README
4. Releases