Thanks for considering contribute to Forte, a project of the ASYML family.
This file outlines the guidelines for contributing to Forte and ASYML projects. While the guideline cannot cover all scenarios, we ask everyone to be reasonable and make your bets judgments, and feel free to propose changes to this document via a pull request.
This project and everyone participating in it is governed by the Contributor Covenant Code of Conduct. By participating, you are expected to uphold this code. Please report unacceptable behavior to [email protected].
There are many ways you can contribute to ASYML projects. The goal of our projects is to modularize Machine Learning and NLP questions, and to make NLP/ML problems as standard engineering problems. Each project solves slightly different problems. Pick the problem you are most interested and get started!
- Texar(Texar-Pytorch): Modularize a complex ML model into smaller components at different levels.
- Forte: Decompose and abstract complex NLP problems into multiple modules, and standardize the interface between the sub-problems and ML interface.
- Forte Wrappers: Create Forte native composable modules by creating adapters to the great ML tools built by the commuinity.
- Stave: Provide visualization and annotation for NLP tasks, by providing generic UI elements based on the abstraction.
We adopt a namespace packaging strategy to coordinate different Forte projects
- forte: The root package contains the top-level Forte implementations, such as
pipeline
andpipeline-component
. - forte.data: contains main data relevant components, mainly implements the data pack system and the ontology system.
- forte.dataset: contains readers to popular datasets.
- forte.processors: Processors are core components that perform NLP tasks. This module currently contains some base processor implementations and simple processor examples. More processor examples can be found in the Forte Wrappers repository.
- forte.models: Contains our in-house development of some NLP models.
- forte.common: configuration, exceptions and other sharable code.
We use the namespace packaging here to make sure different fortex
projects can be installed
at the same folder without conflicting with each other. In general, fortex
is the shared namespace
for different Forte-based projects.
fortex.xxx
: Mainly used to host adapters of third party tools. Each tool is installed in its own namespace to avoid dependency conflicts. Each directory contains a standalone project and can be installed independently. The project will be installed asfortex.xxx
and underfortex/xxx
folder in the site-packages. For example,fortex.nltk
will be installed undersite_packages/fortex/nltk
.
We have a similar package convention for the ontology code.
- The
ft.onto
namespace contains the core/basic ontology types defined by Forte, data types in this namespace are mainly generic ML data concepts, such as "Sentence", "Token". Mainly the Forte developers will use this namespace. Since no namespace packaging is used here, if anotherft
package is used, it may cause conflicts. So we advise you not to. - The
ftx
namespace supports namespace packaging:- We use
ftx.onto
namespace to show extra types for demo/example purposes. - We also develop in other
ftx.xxx
namespace types for certain domains wherexxx
represents that domain.
- We use
Bugs are tracked as GitHub issues. Search the issues to make sure the problem is not reported before.
To report a bug, create an issue provide the following information by filling in the bug report template.
In the bug template, make sure you include enough information for reproducing the problem:
- Use a descriptive title to identify the problem.
- Describe the steps to reproduce the problem, ideally a minimum code/command that can reproduce the problem.
- Describe the environment as detail as possible.
- Describe the actual behavior, and the expected behavior.
Enhancements are also tracked as issues. Similarly, Search the issues to make sure the enhancement is not suggested before. To suggest the enhancement, create an issue by filling in the feature enhancement template.
Following the feature template, fill in the information in more details:
- A clear and concise description of what the problem is.
- Describe the solution you'd like, with a clear and concise description of what you want to happen.
- Describe alternatives you've considered.
- Include as much context as possible.
When you have fixed a bug or implemented a new feature, you can create a pull request for review.
-
Use a PR Template to structure your PR, and here:
- The first line should always start with
This PR fixes [issue link]
orThis PR partially addresses [issue link]
where[issue link]
can be replaced with a#ISSUE_ID
associated to a specific issue. This allows Github to automatically link your pull request to the corresponding issue. If this pull request will close the issue we usefixes
otherwise we can saypartially addresses
. - Description of changes: You should include a description of the changes that you make. If the pull request is aimed to fix an issue, you can explain the approaches to address the problem.
- Possible influences of this PR: List all the potential side-effects of your update. Examples include influences on compatibility, performance, API signature, etc.
- Test Conducted: Describe the test cases to verify the changes in pull request. You should always create unit tests for your updates in the pull request and make sure they can cover all of the conditional branches, especially the ones related to corner cases where bugs usually stem from. We will use Covergage to gauge the effectiveness of tests and you can refer to Codecov report to see which lines are not visited by your test cases.
- The first line should always start with
-
Start your PR as draft and try to pass the Github Action CI check.
-
Mark the PR to be ready-for-review once you are satisfied with it.
We use standard issue labels such as priority, bug, enhancement, etc. We have a few
topic labels to identify the type of the issue. Currently the topics are data
(
problems in our data system), docs
(problems about documentation), examples
(
problems about the examples), interface
(the interfaces between different modules)
, model
(machine learning models), ontology
(the ontology system). We may have more
topic labels in the future.
The programming language for Forte is Python. We follow
the Google Python Style guide. The
project code is examined using pylint
, flake8
, mypy
, black
and sphinx-build
which will be run
automatically in CI. It's recommended that you should run these tests locally before submitting your pull request to save time. Refer to the github workflow here for detailed steps to carry out the tests. Basically what you need to do is to install the requirements (check out the Install dependencies
sections) and run the commands (refer to the steps in Format check with Black
, Lint with flake8
, Lint with pylint
, Lint main code with mypy when torch version is not 1.5.0
, Build Docs
, etc.).
We also recommend using tools pre-commit
that automates the checking process before each commit since checking format is a repetitive process. We have the configuration file .pre-commit-config.yaml
that lists several plugins including black
to check format in the project root folder. Developers only need to install the package by pip install pre-commit
. All the package versions in the .pre-commit-config.yaml
must be consistent with package versions in workflow configuration. For example, black
package version should be set to the same.
All public methods require docstring and type annotation. It is recommended to add docstring for all functions. The docstrings should follow the Comments and Docstrings
section of Google Python Style guide. We will include a pylint plugin called docparams to validate the parameters of docstrings:
- parameters
and their types- types are only required in function signatures and sphinx will build parameter type hyperlinks based on that.
- return value and its type
- exceptions raised
You should take special care of the indentations in your documentation. Make sure the indents are consistent and follow the Google Style guide. All sections other than the heading should maintain a hanging indent of two or four spaces. Refer to the examples here for what is expected and what are the requirements for different sections like args
, lists
, returns
, etc. Invalid indentations might trigger errors in sphinx-build
and will cause confusing rendering of the documentation. You can run sphinx-build
locally to see whether the generated docs look reasonable.
Another aspect that should be noted is the format of links or cross-references of python objects. Make sure to follow the sphinx cross-referencing syntax. The references will be checked by sphinx-build nit-picky mode which raises warnings for all the missing and unresolvable links.
We plan to deliver high quality tutorials via Jupyter notebooks.
Notebooks are written under docs/notebook_tutorial
folder, and we keep notebooks there for several reasons. First, it's friendly for new users to learn forte with a runnable notebook. Second, it can be rendered directly by the sphinx documentation pacakge by including the relative path to notebook in docs/index_toc.rst
. It is straightforward for users to make references on how to use forte with the context of application. Third, we write notebook testing under tests/forte/notebook
to ensure the notebook is runnable as API changes.
Tutorials may use large assets like screenshots and illustrations. To keep them away from the main branch that may make a regular clone too slow, we now organize these assets in a separated "assets" branch to host the assets.
Jupyter notebook written under docs/notebook_tutorial
will be rendered in the sphinx documentation using package nbsphinx
. You need to make sure notebook can be rendered normally in sphinx documentation. After writing notebook under , run this command under docs folder. It might give you some sphinx warnings and you need to fix them.
As notebook is rendered in the sphinx documentation, we might want to include hyperlinks to other sphinx pages in the documentation. For example, if we want to mention another rst
file, we can write the hyperlinks in the markdown way with the relative path to the rst
file such as [reader](../toc/reader.rst)
.
As notebook includes code that might break over time when API changes. Plus, we want to test code efficiently.
Therefore, we test notebook by using package testbook
.
User can refer to notebook_test_tutorial.py for how to test notebook.
As notebook will be running in github CI, we need to consider its package dependencies and add required packages in matrix.notebook-details.dep
.
As current notebooks requires fortex packages, so we limit torch version == 1.5.0 while testing notebooks.
Notebooks will not be running automatically after committing files to the repository. Developer needs to keep notebook cell outputs that are needed for the purpose of illustration.
- Limit the first line to 72 characters or less
- Reference issues and pull requests in the second line
- For documentation only changes, use [skip ci] or [ci skip] in your commit messages to skip travis build.