Skip to content

Latest commit

 

History

History
373 lines (288 loc) · 23.2 KB

CONTRIBUTING.md

File metadata and controls

373 lines (288 loc) · 23.2 KB
Raw

Contributing to OpenVINO Notebooks

Thank you for being interested in contributing to the OpenVINO Notebooks repository! This guide explains the design decisions, requirements, and coding guidelines for the OpenVINO Notebooks repository. Please read the Design Decisions and Validation sections before jumping to the Getting Started section.

The goal of this document is to make it as easy as possible to contribute to the OpenVINO Notebooks repository, while maintaining the quality and consistency of the notebooks in the repository.

If you have a question, about the notebooks or about contributing to the repository, please create a discussion!

Design Decisions

The goals of the OpenVINO Notebooks are:

  • to make it easy to get started with OpenVINO.
  • to teach how to use OpenVINO tools to do inference, convert and quantize models.
  • to make it easy to use models from OpenVINO's Open Model Zoo and other public models.

To do this, there are a few requirements that all notebooks need to pass.

General design considerations

  1. The notebooks work on Windows, macOS and Linux (see supported operating systems) with Python 3.7, 3.8, 3.9 and 3.10.
  2. As a rule, the notebooks do not require installation of additional software that is not installable by pip. We do not assume that users have installed XCode Developer Tools, Visual C++ redistributable, cmake, etc. Please discuss if your notebook does need C++ - there are exceptions to this rule.
  3. The notebooks should work on all computers, and in container images. We cannot assume that a user will have an iGPU or a webcam, so using these should be optional. For example, In the case of webcam inference, provide the option to use a video.
  4. The notebooks should work in Jupyter Lab and Jupyter Notebook. If a dependency does not work in either of these, the notebooks should still be usable: use graceful degradation. For example, if a visualization library only works in Jupyter Lab, but offers substantial advantages, make sure that a user who runs Jupyter Notebook still sees the output, even if it is not interactive/3D/annotated/etc.
  5. With the exception of notebooks that demonstrate training of neural networks, all notebooks should by default run in less than five minutes with "Run All Cells" (excluding time required to download files). If this means using a smaller model or smaller dataset that gives less than optimal results, or having a less amazing visualization, provide the better option that takes longer as an option.
  6. Not everyone who uses the notebooks will have a fast computer and/or fast internet. It is not always possible to use a smaller model or a smaller dataset, but if it is, please do that, and provide an option for the larger model or dataset.
  7. The target audience for the notebooks includes both experienced and new developers. The goal is not just to show the output of a model, but to teach how OpenVINO works, by interacting with it. Not all notebooks need to be full-fledged tutorials, but it is always good to explain steps and add comments.
  8. Respect for human rights is rooted in our values at Intel. We will not accept contributions that perform facial recognition or analyze demographics like age and gender.

Implementation choices

  1. Notebooks in this repository typically rely on a shared requirements.txt file. However, contributors are encouraged to install the required packages at the top of their notebook using %pip install -q ... commands. This allows the notebooks to be run independently as standalone examples. To maintain package compatibility, contributors are expected to install the same versions of packages as specified in the shared requirements.txt file. This helps ensure consistency in our testing pipelines and prevents dependency conflicts.
  2. The notebooks are located in the "notebooks" subdirectory. There is a subdirectory for every notebook, with generally the same base name as the notebook. For example, the 001-hello-world.ipynb notebook can be found in the 001-hello-world directory.
    • See the Notebook naming section below, for the numbering of the notebooks.
    • Add a README.md to the notebook subdirectory. Add a screenshot that gives an indication of what the notebook does if applicable.
    • Avoid adding any other files to the notebook's subdirectory. Instead, rely on models and data samples available online and fetch them within the notebook. Please refer to the Notebook utils section.
  3. In case you want to utilize one of the Open Model Zoo models, refer to the 104 Model Tools notebook.
  4. The notebooks should provide an easy way to clean up the downloaded data, for example with a commented-out cell at the end of the notebook.

Coding Guidelines

  1. See PEP 20
  2. Format notebook code with Black, with a line width of 100. See Tools.
  3. Use f-strings for string formatting: PEP 498
  4. Use keyword/named arguments when calling a function with more than one parameter: function(a=1, b=2) instead of function(1, 2)
  5. Use from pathlib import Path for path manipulation instead of os.path
  6. Add type hints to functions: PEP 484
  7. Add REST style docstring (see 403 for an example). It is not necessary to specify the parameter type in the docstring, since type hints are already added to the function definition.
  8. Do not use global variables in functions: a function should not depend on values that are defined outside it.
  9. Use ALL_CAPS for constants.
  10. Prefer consistency. Example: if other notebooks use import numpy as np do not use import numpy in yours.

Other things to keep in mind

  1. Always provide links to sources. If your notebook implements a model, link to the research paper and the source GitHub (if available).
  2. Use only data and models with permissive licenses that allow for commercial use, and make sure to adhere to the terms of the license.
  3. If you include code from external sources in your notebook add the name, URL and license of the third party code to the licensing/third-party-programs.txt file.
  4. Don't use HTML for text cells, use Markdown markups instead.
  5. Add Table of content to top of the Notebook, it helps to get quick fist understanding of content and ease of navigation in the dev environment. There is no need to think about it during development, it can be built or updated after changes with .ci\table_of_content.py. Just run the script with the parameter -s/--source, specifying a Notebook or a folder with several notebooks as value, the changes will be applied to all of them.

Notebook naming

Names should be descriptive but not too long. We use the following numbering scheme:

  • 000- hello world like notebooks: very small tutorials that help to quickly show how OpenVINO works.
  • 100- OpenVINO tool tutorials: explain how to optimize and quantize notebooks.
  • 200- OpenVINO model demos: demonstrate inference on a particular model.
  • 300- Training notebooks: notebooks that include code to train neural networks.
  • 400- Live demo notebooks: demonstrate inference on a live webcam.

Please use the first available number in the branch, trying to fill the holes e.g. choose 206, when there are 205 and 207 and 206 is missing.

README.md files

Every notebook must have a README.md file that briefly describes the content of the notebook. A simple structure for the README.md file is described below:

# Title of Tutorial
[brief intro, basic information about what will be described]

## Notebook Contents
[more details, possibly information about research papers, the model(s) used and/or data]
Additional subsections, e.g license information.


## Installation Instructions
[link to installation guide, other important information for install process]

Notebooks that work in Binder have a Launch Binder badge in the README.md files. In the same way, notebooks that work in Google Colab have a Launch Colab badge in the README.md files.

File Structure

To maintain consistency between notebooks, please follow the directory structure outlined below.

notebooks/
└──<three-digit-number>-<title>/
   ├── README.md
   ├── <three-digit-number>-<title>.ipynb
   ├── utils/
   ├── model/
   ├── data/
   └── output/

In case the example requires saving additional files to disk (e.g. models, data samples, utility modules, or outputs), please create corresponding folders on the same level as README.md file.

Recommendations for File Structure

  • Model

We recommend to load your models using URL or other ways of distribution of pre-trained models, like PyTorch Hub or the Diffusers package.

  • Data

We recommend to use embedded URL for image/video data. Follow the below instructions to create embedded URL in GitHub:

  • Go to any issue on GitHub.

  • In the comment section, you can attach files. Just drag/drop, select or paste your image.

  • Copy the code/link displayed in the text area

  • License

If you download or include a model, it must be licensed under an open source license like Apache 2.0 which allows for redistribution, modification and commercial use.

Any datasets, images or videos used for fine-tuning, quantization or inference inside a notebook must be licensed under Creative Commons 4.0 (CC 4.0) with permission for commercial use. If commercial use is not allowed, but the data is under CC 4.0, special approval will be required. Please let us know in your pull request if your data has any restrictions on use.

Notebook utils

The notebook_utils.py file in the notebooks/utils directory contains utility functions and classes that can be reused across notebooks. It contains a download_file() function that optionally shows a progress bar, and a standard way to convert segmentation maps to images and display them. The Python file is generated from notebook_utils.ipynb notebook in the same directory. If you want to add a function or class to notebook_utils.py, please add it to the notebook, and generate the Python file with jupyter nbconvert notebook_utils.ipynb --TagRemovePreprocessor.remove_cell_tags=hide --to script Add a "hide" tag to any demo cells (from the right side gear sidebar) to prevent these cells from being added to the script.

Interactive inference with Gradio

To enhance the functionality of your notebook, it is recommended to include an interactive model inference interface at the end. We recommend using Gradio for this purpose.

Here are some guidelines to follow:

  • Install the latest version of Gradio by running pip install -q -U gradio.
  • If you're using a gradio.Interface object in your demo, disable flagging by setting the allow_flagging keyword argument to 'never'.
  • Launch the interface with debug=True. This mode blocks the main thread, identifies any possible inference errors and stops the running Gradio server when execution is interrupted. It is important to disable it when executing on CI to prevent main thread hanging. It is done by adding test_replace metadata key to the cell containing the line with launch method, and replacing this line with the same one but excluding debug argument. More detailed instructions can be found here.
  • Avoid setting share=True for the launch method of gradio.Interface. Enabling this option generates an unnecessary public link to your device, which can pose a security risk. Instead, if the interface window is not rendering in your case, consider temporarily setting the server_name and server_port parameters to the address where the server is located. This workaround is particularly useful when you are using remote Jupyter server. To assist other users, please leave a comment in your notebook explaining this solution. It will help them quickly resolve the issue if they encounter the same problem. The comment we recommend to use for this:
# if you are launching remotely, specify server_name and server_port
# demo.launch(server_name='your server name', server_port='server port in int')
# Read more in the docs: https://gradio.app/docs/
  • We use Gradio Blocks only when we need to create a complex interface. The Gradio Interface class provides an easy-to-use interface and saves development time, so we use it whenever possible. However, for more complex interfaces, Gradio Blocks gives us more flexibility and control.

Notebooks Metadata

Each notebook file has metadata that includes additional information about the notebook. Some metadata fields (e.g. title, creation date, links to GitHub, Colab, Binder etc.) are generated automatically from notebook content or related README.md file. However other fields (e.g. tags, image URL) should be defined by notebook contributor. As each notebook file has JSON format, manually defined metadata fields are stored in corresponding .ipynb file in global notebook metadata object (metadata.openvino_notebooks field in the end of notebook JSON structure).

Example of such manually defined notebook metadata:

"openvino_notebooks": {
 "imageUrl": "...",
 "tags": {
  "categories": [
   "First Steps"
  ],
  "libraries": [],
  "other": [],
  "tasks": [
   "Image Classification"
  ]
 }
}

Notebook tags in metadata can have several values and should be a subset of defined tags that can be found in ./selector/src/shared/notebook-tags.js.

  • tags.categories tags relate to notebook groups like "AI Trends", "First Steps", "Model Demos" etc.
  • tags.tasks tags relate to particular AI tasks that are demonstrated in notebook.
  • tags.other tags are free-form tags and can be any string (please follow capitalization naming convention).

Requirements

Contributors are encouraged to install the required packages at the top of their notebook using %pip install ... commands. This allows the notebooks to be run independently as standalone examples. To maintain package compatibility, contributors are expected to install the same versions of packages as specified in the shared requirements.txt file located in the repository root folder. Additional guidelines:

  1. Specify the widest compatible package version range. If your notebook has only a lower bound on some package version, consider specifying it with ">=" sign instead of "==". Specifying the exact version of package might lead to dependency conflict between notebooks.
  2. Do not use spaces between package, version and comparison operator when specifying the package installed. Use "package==version" instead of "package == version".

Validation

Automated tests

We use GitHub Actions to automatically validate that all notebooks work. The following tests run automatically on a new notebook PR:

  • treon: tests that the notebooks execute without problems on all supported platforms.

  • Code check:

    • Uses flake8 to check for unnecessary imports and variables and some style issues
    • Verifies that the notebook is included in the main README.md and the README.md in the notebooks directory.
    • Runs the check_install.py script to test for installation issues

  • Spell check: spell checking is performed by PySpelling module which requires Aspell spell checker in conjunction with our custom word list dictionary (.ci/spellcheck/.pyspelling.wordlist.txt). For information about dealing with found spelling problems please refer to the PySpelling section below.

  • docker_treon: tests that the docker image builds, and that the notebooks execute without errors in the Docker image. To manually run this test, build the Docker image with docker build -t openvino_notebooks . and run the tests with docker run -it --entrypoint /tmp/scripts/test openvino_notebooks. It is recommended to build the image on a clean repository because the full notebooks folder will be copied to the image.

  • CodeQL

  • Notebooks Metadata Validation: verifies that all added or modified notebooks in PR have valid metadata and visualizes them in workflow summary.

    • In the rest of this guide, the automated tests in GitHub Actions will be referred to as CI (for Continuous Integration).

If your notebook takes longer than a few minutes to execute, it may be possible to patch it in the CI, to make it execute faster. As an example, if your notebook trains for 20 epochs, you can set it to train for 1 epoch in the CI. If you do inference on 100 frames of a video, you can set it to do inference on only 1. See this Wiki page for more information.

Manual test and code quality tools

See Getting started about installing the tools mentioned in this section.

treon

Tests are run in the CI with treon, a test framework for Jupyter Notebooks.

To run treon locally, run treon to run the tests for all notebooks, or treon notebook.ipynb for just one notebook. treon fails if the notebook environment is not openvino_env.

nbqa

nbqa allows using a variety of code quality tools on Jupyter Notebooks. For example nbqa flake8 notebook.ipynb will warn about unused imports.

nbdime

nbdime has several useful tools, among which nbdiff-web to show the difference between two notebooks in a web browser. nbdiff can also be used as the standard diff tool for git, with much more useful output than the regular git diff output.

JupyterLab Code Formatter

JupyterLab Code Formatter adds a button to Jupyter Lab to automatically format the code in notebooks with black and isort. Please use either this extension or a different way to automatically format your notebook.

PySpelling

PySpelling is a module to help with automating spell checking and it is essentially a wrapper around the Aspell command line utility. Additional custom (project and domain specific) word list dictionary that extends standard Aspell dictionary is located in .ci/spellcheck/.pyspelling.wordlist.txt file. PySpelling configuration file can be found in .ci/spellcheck/.pyspelling.yml file.

To run spell checking locally, execute the following command:

python .ci/spellcheck/run_spellcheck.py

If spell check is failing, there are any typos or new words, you have two possible options how to fix it:

  1. Add new word (abbreviation, name, term etc.) to the word list dictionary (.ci/spellcheck/.pyspelling.wordlist.txt)
  2. Skip single occurrence of unknown word or even whole phrase - just wrap the text with <spell> tag (for example, <spell>Unknown word or phrase</spell>). Note that <spell> is a custom tag and it doesn't affect the Markdown formatting and style (unlike backticks in preformatted text and code blocks).

Getting started

  1. Create a fork, a copy of the repository, by clicking on the Fork button on the top right of the OpenVINO Notebooks GitHub page
  2. Install the recommended packages for a development environment with pip install -r .ci/dev-requirements.txt inside the openvino_env environment. This installs all the packages mentioned in the Validation section.

    Note: PySpelling dependency from .ci/dev-requirements.txt requires Aspell for spell checking that should be installed manually. For installation instructions please refer to the PySpelling documentation.

  3. Create a branch in this fork, from the main branch. Name the branch however you like.
  4. Double-check the points in the Design Decisions and Validation sections.
  5. Check that your notebook works in the CI
    • Go to the GitHub page of your fork, click on Actions, select treon on the left. There will be a message This workflow has a workflow_dispatch event trigger. and a Run workflow button. Click on the button and select the branch that you want to test.
  6. Test if the notebook works in Binder and Google Colab and if so, add Launch Binder and Launch Colab badges to the README.md files.

Once your notebook passes in the CI and you have verified that everything looks good, make a Pull Request!

Pull Requests (PRs)

  1. If some time has passed since you made the fork, sync your fork via GitHub UI or update your form manually - rebase or merge openvinotoolkit/openvino_notebooks repository main branch to the main branch in your fork.
  2. Create your PR against the openvinotoolkit/openvino_notebooks repository main branch.
  3. Please create a description of what the notebook does with your PR. Screenshots are appreciated!
  4. On making or updating a Pull Request, the tests in the CI will run again. Please keep an eye on them. If the tests fail and you think the issue is not related to your PR, please make a comment on your PR.

Help

If you need help at any time, please open a discussion! If you think one of the guidelines is too strict, or should not apply to you, feel free to ask about that too.