Open-source project
yandexdataschool/Practical_DL avatar
yandexdataschool/Practical_DL

Practical_DL's Dockerfile installs a Python 2.7 PyTorch wheel and a package called panda

DL course co-developed by YSDA, HSE and Skoltech

1,789 stars658 forksJupyter NotebookMIT

At a glance

What is it?
The course repository for the fall'26 deep learning term, with the default branch renamed every term and notebooks for five weeks while the syllabus describes three. The container image underneath it was written for Ubuntu Trusty, installs PyTorch 0.3.0 from a direct wheel URL, and pulls three frameworks from unpinned master archives.
Who is it for?
Practical_DL is course material, and it should be read that way. The notebooks are the product and they are genuinely good teaching artifacts with named authors per topic, but the surrounding scaffolding is a decade old and says so.
Can I use it commercially?
Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
Is it still maintained?
Yes. The repository last received commits 2 days ago.
What is it written in?
Mainly Jupyter Notebook, according to GitHub's language statistics.

Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.

From daily snapshots

Star history

  • yandexdataschool/Practical_DL1,789 ★ · +28
Star history
Repository2026-09-112026-11-15
yandexdataschool/Practical_DL1.8k1.8k

28 Stars in the last 65 days, measured from the daily GitHub snapshot. Compare with other projects

Editorial analysis

The default branch is named for the term and changes every September

The default branch of this repository is `fall26`, not main. The README opens by saying it supplements the deep learning course taught in fall'26 and that previous iterations live on their own branches, linking the 2025 branch as `fall25`. So the whole course is versioned by branch name and the branch you get from a plain clone is whichever term is current. There are no GitHub releases and no tags to pin, which means a notebook you cite in a paper or a blog post has to be cited by commit. It also means the two link styles in the page disagree with each other: the branch link uses `Practical_DL` while the issue link and the PR authorship link use lowercase `practical_dl`. Both resolve on GitHub, and it is a small reminder that this repository has been edited by several cohorts for several years without much tidying.

The syllabus stops at week three and the tree holds five weeks

The root of the repository contains `week01_backprop/`, `week02_autodiff/`, `week03_convnets/`, `week04_finetuning/` and `week05_interpretability/`, plus `resources/`. The syllabus in the README describes week one as an introduction to deep learning with backpropagation and adaptive optimization, paired with a seminar on neural networks in numpy. Week two is a catch-all lecture about deep learning tricks covering dropout, batch and layer normalization, frameworks and other tricks, paired with PyTorch basics. Week three is convolutional neural networks with a seminar on training your first ConvNet. Then the page says to be updated. So two of the five week folders, finetuning and interpretability, exist with no lecture line, no seminar line and no homework line, and the folder name for week two, autodiff, does not appear anywhere in that week's description even though week one's seminar is numpy-level and week three's is convolutional.

The image installs PyTorch 0.3.0 for Python 2.7 by direct URL

The Dockerfile starts `FROM andrewosh/binder-base` and declares a `MAINTAINER` line, which Docker has not supported for a long time in favour of labels. It appends the Ubuntu Trusty backports apt line, which points at a release long out of support, installs `gcc-4.9` and `g++-4.9`, and symlinks `swig3.0` to `swig`. Then the framework install is a URL rather than a version constraint:

code
RUN pip install --upgrade http://download.pytorch.org/whl/cu80/torch-0.3.0.post4-cp27-cp27mu-linux_x86_64.whl

Three things are encoded in that filename. The version is 0.3.0.post4, from before the modern module layout. `cu80` is the CUDA 8.0 build. And `cp27-cp27mu` is a CPython 2.7 manylinux wheel, so this line puts a Python 2 package into an image that also installs `python3-opengl` and later switches to a conda `python3` environment. The next lines install `torchvision` and `keras` with no version at all, on top of that Python 2 torch.

The scientific stack line installs panda instead of pandas

One line in the same file is a straightforward typo with a real consequence:

code
RUN pip install --upgrade matplotlib numpy scipy panda

`pandas` is spelled with an s; `panda` is a different package on PyPI, so this line installs a package nobody in this course intends to use and leaves the dataframe library absent. Everything downstream in the notebooks that expects `import pandas` fails at that import, and the failure looks like a missing dependency rather than a misspelling. Fixing it is one character, which is the useful part of the observation: this image has not been rebuilt for the current term. The image also carries a workaround comment about conda's libgcc blocking the system compiler and hiding `GLIBCXX_3.4.20`, fixed by uninstalling libgcc inside the conda python3 environment, which is the kind of fix that belongs in a git history rather than in a file people keep copying.

Three frameworks install from master archives, so the image is not reproducible

Theano, Lasagne and AgentNet are installed by URL with no commit and no tag:

code
RUN pip install --upgrade https://github.com/yandexdataschool/AgentNet/archive/master.zip

An archive of a branch tip means the same Dockerfile produces a different environment every time it runs, because whatever is at master today is what gets installed. Two of the three, Theano and Lasagne, are frameworks the course itself has moved away from, since week one asks for neural networks in numpy and week two asks for PyTorch basics, yet they remain in the image alongside a Python 2 torch. AgentNet is the same project's own library, pulled from its default branch. The image also installs gym with the atari and box2d extras, with a comment recording why mujoco is left out, since it fails on `import importlib.util`. That is a useful habit, documenting a known failure next to the workaround, and it is the one part of the file that reads as maintained on purpose.

Three ways to run it, and the only live channel is in Russian

The README says assignments can be completed locally or in Google Colab, with a readme inside each `week*` folder, and the Dockerfile adds a third path, a Binder-style base image. The primary language recorded for the repository is Jupyter Notebook, which is the honest description of what a student submits. Support is split: a Telegram chat room, explicitly described as the Russian-speaking HSE track, and GitHub issues for technical problems, bugs in the materials and contribution ideas. Deadlines and grading rules are not in the repository at all; they live on a separate wiki page. So the parts a student needs most, the dates and the rubric, are the parts most likely to drift out of date without a commit to show when they changed.

Week one tells you to worry about PyTorch through a six-year-old issue

The week one checklist ends with a line asking readers to begin worrying about installing PyTorch, with the link going to issue 6 in this repository. That is where the environment advice lives, in a discussion rather than in a file, which is fragile in a repository that renames its default branch every term. The rest of the checklist mixes three different kinds of item under one set of boxes: lecture topics, seminar topics, and announcements such as Homework 1 is out. Week two repeats the pattern with a catch-all lecture described as deep learning as a language alongside dropout, normalization and other tricks. Credits are attributed per contributor and per topic rather than generically, with named people on the lecture videos, the intro notebooks, generative models and autoencoders, the PyTorch and NLP notebooks, image captioning materials and variational autoencoder materials, plus a link to closed pull requests as the record of student fixes.

Editorial conclusion

Practical_DL is course material, and it should be read that way. The notebooks are the product and they are genuinely good teaching artifacts with named authors per topic, but the surrounding scaffolding is a decade old and says so. If you are taking the course, follow the current week's README and the wiki page for deadlines, use Colab or your own environment for the assignments, and treat the Dockerfile as an archaeology exhibit rather than a setup path. If you want to reuse the notebooks, the material to check first is week02_autodiff, since the Dockerfile has no modern autodiff library in it, and the second is week04_finetuning, which exists in the tree with no syllabus entry and no statement of what it expects. Either way, pin whatever you reuse to a commit, because the default branch is renamed each term and there are no releases to cite.

Frequently asked questions

What is yandexdataschool Practical_DL?

It is the MIT licensed repository of lecture and practice material for a deep learning course co-developed by YSDA, HSE and Skoltech, with the primary language recorded as Jupyter Notebook. The README says it supplements the course taught in fall'26, and the default branch is named `fall26`, with previous iterations kept on branches such as `fall25`.

Which weeks does the Practical_DL course cover?

The syllabus describes week one on deep learning introduction and backpropagation with a numpy seminar, week two on deep learning tricks including dropout and normalization with PyTorch basics, and week three on convolutional networks with a first ConvNet seminar, then says to be updated. The repository already holds week01_backprop, week02_autodiff, week03_convnets, week04_finetuning and week05_interpretability.

What does the Practical_DL Dockerfile install?

It builds on `andrewosh/binder-base`, adds the Ubuntu Trusty backports apt line and gcc-4.9, then installs PyTorch 0.3.0.post4 as a direct CPython 2.7 CUDA 8.0 wheel URL, torchvision and keras with no version, Theano, Lasagne and AgentNet from master archive URLs, gym with atari and box2d extras, and matplotlib, numpy, scipy and a package named panda.

Can I do the Practical_DL assignments in Google Colab?

Yes. The README says all assignments can be completed locally or in Google Colab, with readme files inside the week folders. Installing PyTorch is called out as a week one concern with a link to an issue in the repository, and support runs through a Telegram chat room for the Russian-speaking HSE track plus GitHub issues.

Where are the deadlines and grading rules for Practical_DL?

Not in the repository. The README links a separate wiki page for deadlines and grading rules, and it has no GitHub releases, so a course reader citing this material cites a commit on the term branch rather than a version.

Who wrote the Practical_DL course material?

Named contributors cover specific topics: main track lecture videos for the first eleven lectures, intro notebooks and admin, notebooks on generative models and autoencoders, PyTorch and NLP notebooks plus a bonus lecture, image captioning materials, variational autoencoder materials, notebooks and technical issue resolution, and admin work. Students and volunteers are credited through a link to closed pull requests.

Official sources

  1. Issues
  2. License: MIT
  3. README
  4. yandexdataschool/Practical_DL on GitHub
Add this badge to your README

If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.

Add this badge to your README

markdown
[![Hysen Labs](https://hysenlabs.com/badge/yandexdataschool-practical-dl.svg)](https://hysenlabs.com/projects/yandexdataschool-practical-dl)

Star history chart for your README

markdown
[![Star history](https://hysenlabs.com/chart/yandexdataschool-practical-dl.svg)](https://hysenlabs.com/projects/yandexdataschool-practical-dl)

Was this useful?