chore: import upstream snapshot with attribution
This commit is contained in:
wehub-resource-sync
@@ -0,0 +1 @@
|
||||
|
||||
@@ -0,0 +1,20 @@
|
||||
# Minimal makefile for Sphinx documentation
|
||||
#
|
||||
|
||||
# You can set these variables from the command line, and also
|
||||
# from the environment for the first two.
|
||||
SPHINXOPTS ?=
|
||||
SPHINXBUILD ?= sphinx-build
|
||||
SOURCEDIR = source
|
||||
BUILDDIR = build
|
||||
|
||||
# Put it first so that "make" without argument is like "make help".
|
||||
help:
|
||||
@$(SPHINXBUILD) -M help "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)
|
||||
|
||||
.PHONY: help Makefile
|
||||
|
||||
# Catch-all target: route all unknown targets to Sphinx using the new
|
||||
# "make mode" option. $(O) is meant as a shortcut for $(SPHINXOPTS).
|
||||
%: Makefile
|
||||
@$(SPHINXBUILD) -M $@ "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)
|
||||
+307
@@ -0,0 +1,307 @@
|
||||
# Looking for `v1.0` of the `cleanlab` docs?
|
||||
|
||||
Please refer to the [`v1.0.1` documentation](https://docs.cleanlab.ai/v1.0.1/);
|
||||
the code for `v1.0` is identical to the code for `v1.0.1`.
|
||||
|
||||
# Looking for rendered docs?
|
||||
|
||||
See <https://docs.cleanlab.ai/> if you want to browse the documentation (including for past versions).
|
||||
|
||||
# CI/CD for `cleanlab` docs
|
||||
|
||||
In the `cleanlab` repository, we've configured GitHub Actions to perform the following automatically:
|
||||
|
||||
1. When a commit is pushed to the `master` branch, a new version of the `master` docs will be built and deployed to the `cleanlab-docs` repository.
|
||||
|
||||
2. When a release is published, a new version of the docs with the corresponding release tag will be built and deployed as a new folder in the `cleanlab-docs` repository. Redirection to the `stable` version of the docs will be changed to this newly released one, accessible via a link on the docs' site sidebar. All the older versions will remain available in the `cleanlab-docs` repo, accessible by manually entering the subdirectory in the URL.
|
||||
|
||||
3. When a user manually runs the workflow, one of the above will happen depending on the user's selection to run from a `branch` or `tag`.
|
||||
|
||||
If you'd like to build our docs locally or remotely yourself, or want to know more about the steps taken in the GitHub Pages workflow, read on!
|
||||
|
||||
|
||||
# Build the `cleanlab` docs **locally**
|
||||
|
||||
1. [Fork](https://docs.github.com/en/get-started/quickstart/fork-a-repo#forking-a-repository) and [clone](https://docs.github.com/en/get-started/quickstart/fork-a-repo#cloning-your-forked-repository) the `cleanlab` repository.
|
||||
|
||||
2. Install the required packages to build the docs:
|
||||
|
||||
```
|
||||
uv pip install -r docs/requirements.txt
|
||||
```
|
||||
|
||||
Or with pip:
|
||||
|
||||
```
|
||||
pip install -r docs/requirements.txt
|
||||
```
|
||||
|
||||
For Macs with Apple silicon: replace tensorflow in `docs/requirements.txt` with: `tensorflow-macos==2.9.2` and `tensorflow-metal==0.5.1`
|
||||
|
||||
|
||||
3. Install [Pandoc](https://pandoc.org/installing.html).
|
||||
|
||||
4. If you don't already have it, install [wget](https://www.gnu.org/software/wget/). This can be done with `brew` on macOS: `brew install wget`
|
||||
|
||||
5. **[Optional]** [Create a new branch](https://www.atlassian.com/git/tutorials/using-branches), make your code changes, and then `git commit` them. **ONLY COMMITTED CHANGES WILL BE REFLECTED IN THE DOCS BUILD WITH `sphinx-multiversion`.** Instead use `sphinx-build` if you don't want to commit some test changes but still want to see their corresponding docs.
|
||||
|
||||
6. Build the docs with either
|
||||
1. [`sphinx-multiversion`](https://sphinx-contrib.github.io/multiversion/):
|
||||
|
||||
* If you're building from a **branch** (usually the `master` branch):
|
||||
|
||||
```
|
||||
sphinx-multiversion docs/source cleanlab-docs -D smv_branch_whitelist=YOUR_BRANCH_NAME -D smv_tag_whitelist=None
|
||||
```
|
||||
|
||||
* If you're building from a **tag** (usually the tag of the stable release):
|
||||
|
||||
```
|
||||
sphinx-multiversion docs/source cleanlab-docs -D smv_branch_whitelist=None -D smv_tag_whitelist=YOUR_TAG_NAME
|
||||
```
|
||||
|
||||
Note: To also build docs for another branch or tag, run the above command again changing only the `YOUR_BRANCH_NAME` or `YOUR_TAG_NAME` placeholder.
|
||||
|
||||
2. [`sphinx-build`](https://www.sphinx-doc.org/en/master/man/sphinx-build.html):
|
||||
|
||||
* If you want to test out some changes without comitting them, then you can build from your current working directory tree (where you have any un-committed changes locally saved):
|
||||
|
||||
```
|
||||
sphinx-build docs/source cleanlab-docs
|
||||
```
|
||||
This won't properly produce/display other versions of the docs, but that shouldn't matter if you are just trying to test some local edits to the current version. If some notebooks are giving you trouble (eg. due to runtime or dependencies), you can simply delete those .ipynb files before calling `sphinx-build`.
|
||||
|
||||
**Fast build**: Executing the Jupyter Notebooks (i.e., the `.ipynb` files) that make up some portion of the docs, such as the tutorials, takes a long time. If you want to skip rendering these, set the environment variable `SKIP_NOTEBOOKS=1`. You can either set this using `export SKIP_NOTEBOOKS=1` or do this inline with `SKIP_NOTEBOOKS=1 sphinx-multiversion ...`.
|
||||
|
||||
**Skipping specific notebooks**: If you want to skip rendering a few specific notebooks during your local build, the best way to do this is to temporarily move the files outside the `cleanlab` folder (so `nbsphinx` would not find it), then build the docs, before finally moving the files back (to ensure they will not be deleted when pushed to GitHub)
|
||||
|
||||
Example workflow for skipping notebooks, given our current working directory is the `cleanlab` root folder and we want to ignore a specific notebook:
|
||||
|
||||
1. create an empty folder outside of cleanlab folder
|
||||
```
|
||||
mkdir ../ignore_notebooks
|
||||
```
|
||||
|
||||
2. move the notebook to ignore from local build to the newly created folder
|
||||
```
|
||||
mv docs/source/tutorials/path/to/notebook.ipynb ../ignore_notebooks
|
||||
```
|
||||
|
||||
3. build the docs locally, using `sphinx-build` as it does not require you to commit your changes
|
||||
```
|
||||
sphinx-build docs/source cleanlab-docs
|
||||
```
|
||||
|
||||
4. move the notebook back to its original location
|
||||
```
|
||||
mv ../ignore_notebooks/notebook.ipynb docs/source/tutorials/path/to/
|
||||
```
|
||||
|
||||
|
||||
While building the docs with `sphinx-multiversion`, your terminal might output:
|
||||
* `unknown config value 'smv_branch_whitelist' in override, ignoring`, and
|
||||
* `unknown config value 'smv_tag_whitelist' in override, ignoring`.
|
||||
|
||||
This is because the `smv_branch_whitelist` and `smv_tag_whitelist` config values are only used by `sphinx-multiversion`, but may also be checked by `sphinx` or other extensions that do not use them. Hence, these can be safely ignored as long as the docs are built correctly.
|
||||
|
||||
7. **[Optional]** To show dynamic versioning and version warning banners:
|
||||
|
||||
* Copy the `docs/_templates/versioning.js` file to the `cleanlab-docs/` directory.
|
||||
|
||||
* In the copied `versioning.js` file:
|
||||
|
||||
* find `placeholder_version_number` and replace it with the latest release tag name, and
|
||||
|
||||
* find `placeholder_commit_hash` and replace it with the `master` branch commit hash.
|
||||
|
||||
8. **[Optional]** To redirect site visits from `/` or `/stable` to the stable version of the docs:
|
||||
|
||||
* Create a copy of the `docs/_templates/redirect-to-stable.html` file and rename it as `index.html`.
|
||||
|
||||
* In this `index.html` file, find `stable_url` and replace it with `/cleanlab-docs/YOUR_LATEST_RELEASE_TAG_NAME/index.html`.
|
||||
|
||||
* Copy this `index.html` to:
|
||||
|
||||
* `cleanlab-docs/`, and
|
||||
|
||||
* `cleanlab-docs/stable/`.
|
||||
|
||||
9. The docs for each branch and/or tag can be found in the `cleanlab-docs/` directory, open any of the `index.html` in your browser to view the docs:
|
||||
|
||||
```
|
||||
cleanlab-docs
|
||||
| index.html (redirects to stable release of the docs)
|
||||
| versioning.js (for dynamic versioning and version warning banner)
|
||||
|
|
||||
└───YOUR_BRANCH_NAME (e.g. master)
|
||||
│ index.html
|
||||
│ ...
|
||||
│
|
||||
└───YOUR_TAG_NAME_1 (e.g. your stable release tag name)
|
||||
│ index.html
|
||||
│ ...
|
||||
│
|
||||
└───YOUR_TAG_NAME_2 (e.g. an old release tag name)
|
||||
│ index.html
|
||||
│ ...
|
||||
│
|
||||
└───stable
|
||||
│ index.html (redirects to stable release of the docs)
|
||||
│
|
||||
└───...
|
||||
```
|
||||
|
||||
Note: If you're building the docs from a working directory tree, the docs will be found at the top of the `cleanlab-docs/` directory:
|
||||
|
||||
```
|
||||
cleanlab-docs
|
||||
| index.html (docs for the working directory tree)
|
||||
| ...
|
||||
|
|
||||
└───...
|
||||
```
|
||||
|
||||
This may overwrite some of the files in `cleanlab-docs/`, like `index.html` from the previous step.
|
||||
|
||||
# Build the `cleanlab` docs **remotely** on GitHub
|
||||
|
||||
1. [Fork](https://docs.github.com/en/get-started/quickstart/fork-a-repo#forking-a-repository) the `cleanlab` repository.
|
||||
|
||||
2. [Create a new repository](https://docs.github.com/en/pages/getting-started-with-github-pages/creating-a-github-pages-site#creating-a-repository-for-your-site) named `cleanlab-docs` and a new branch named `master`.
|
||||
|
||||
3. In the `cleanlab-docs` repo, [configure GitHub Pages](https://docs.github.com/en/pages/getting-started-with-github-pages/creating-a-github-pages-site#creating-a-repository-for-your-site); under the **Source** section, select the `master` branch and `/(root)` folder. Take note of the URL where your site is published.
|
||||
|
||||
4. [Generate SSH deploy key](https://github.com/peaceiris/actions-gh-pages#user-content-️-create-ssh-deploy-key) and add them to your repos as such:
|
||||
|
||||
* In the `cleanlab-docs` repo, go to **Settings > Deploy Keys > Add deploy key** and add your **public key** with the **Allow write access**
|
||||
* In the `cleanlab` repo, go to **Settings > Secrets > New repository secrets** and add your **private key** named `ACTIONS_DEPLOY_KEY`
|
||||
|
||||
5. In the `cleanlab` repo, check that you have the **GitHub Pages** workflow under the repo's **Actions** tab. This should be created automatically from `.github\workflows\gh-pages.yaml`. This workflow can be activated by any of the 3 triggers below:
|
||||
|
||||
* A push to the `master` branch in the `cleanlab` repo.
|
||||
* Publish of a new release in the `cleanlab` repo.
|
||||
* Manually run from the **Run workflow** option and select either the `master` branch or one of the release tag.
|
||||
|
||||
6. Activate the workflow with any of the 3 triggers listed above and wait for it to complete.
|
||||
|
||||
7. Navigate to the URL where your GitHub Pages site is published in step 3. The default URL should have the format *https://repository_owner.github.io/cleanlab-docs/*.
|
||||
|
||||
# Manually adding build artifacts to the `cleanlab-docs` repo
|
||||
|
||||
GitHub Actions automatically builds and deploys the docs' build artifacts when [triggered](#cicd-for-cleanlab-docs). If you delete and recreate a release tag, the docs for this tag will be rebuilt and redeployed, hence overwriting the existing artifacts with the new ones.
|
||||
|
||||
On rare occasions, you may want to update the docs without deleting and recreating the release tag, for example, when you want to fix a typo in the docs, but you've already deployed your tag to PyPI or Conda. This can be done by manually adding specific docs' build artifacts to the `cleanlab/cleanlab-docs` repo. These steps are for users who have `push` permission to `cleanlab/cleanlab` and `cleanlab/cleanlab-docs` repo.
|
||||
|
||||
1. If you haven't already done so, [clone](https://docs.github.com/en/repositories/creating-and-managing-repositories/cloning-a-repository) the `cleanlab/cleanlab` repo.
|
||||
|
||||
2. [Create and checkout a new branch](https://git-scm.com/docs/git-checkout).
|
||||
|
||||
3. Make the necessary code changes.
|
||||
|
||||
4. Perform [git add](https://git-scm.com/docs/git-add) and [git commit](https://git-scm.com/docs/git-commit) for the changes.
|
||||
|
||||
5. [git push](https://git-scm.com/docs/git-push) to the `cleanlab/cleanlab` repo. As this is pushed from a non-`master` branch, GitHub Actions will only build but not deploy the docs' build artifacts.
|
||||
|
||||
6. Navigate to [github.com/cleanlab/cleanlab](https://www.github.com/cleanlab/cleanlab) in your browser, select the "Actions" tab, under "Workflow", click "GitHub Pages", then select the workflow that was triggered by the previous step.
|
||||
|
||||
7. Ensure that the workflow has completed running.
|
||||
|
||||
8. Scroll to the bottom of the page, under "Artifacts", click "docs-html" to download the docs' build artifacts.
|
||||
|
||||
9. Unzip "docs-html.zip" and open the "docs-html" folder.
|
||||
|
||||
10. Identify the files you would like to replace, i.e., the corresponding files creating the pages on [docs.cleanlab.ai](https://docs.cleanlab.ai).
|
||||
|
||||
11. Replace these files in [github.com/cleanlab/cleanlab-docs](https://www.github.com/cleanlab/cleanlab-docs) by uploading the new ones to the corresponding version folder in the `master` branch of the `cleanlab/cleanlab-docs` repo.
|
||||
|
||||
> :warning: Any build artifacts manually added to `cleanlab/cleanlab-docs` that do not live in the `master` branch of the `cleanlab/cleanlab` repo will be lost in future versions of cleanlab docs. So any edit made in the v2.0.0 docs which you also want to have in the v2.0.1, v2.0.2, etc. docs needs to be introduced as a PR to the `cleanlab/cleanlab` repo as well.
|
||||
|
||||
> :warning: Currently, if updating stable/old version (say `vXXX`) of tutorials from latest master branch version, the install of cleanlab package in notebooks/colabs will be wrong. To remedy this, you need to update the cleanlab version in all `.ipynb` files inside folders: **cleanlab-docs/vXXX/tutorials/** and **cleanlab-docs/vXXX/_sources/**. The tutorial `.html` pages will also have wrong colab links as well. Currently have to update the `.html` files in **cleanlab-docs/vXXX/tutorials/** to replace these colab links with the proper links (replace `/master/` in the link with `/vXXX/` for the version you are building docs for).
|
||||
|
||||
# Behind-the-scenes of the GitHub Pages workflow
|
||||
|
||||
We've configured GitHub Actions to run the GitHub Pages workflow (gh-pages.yaml) to build and deploy our docs' static files. Here's a breakdown of what this workflow does in the background:
|
||||
|
||||
## Spin up and configure the CI/CD server
|
||||
|
||||
1. Spin up a Ubuntu server.
|
||||
|
||||
2. Install Pandoc, a document converter required by `nbsphinx` to generate static sites from notebooks (`.ipynb`).
|
||||
|
||||
3. Check-out the `cleanlab` repository.
|
||||
|
||||
4. Setup Python and cache dependencies.
|
||||
|
||||
5. Install dependencies for the docs from `docs/requirements.txt`.
|
||||
|
||||
## Build the docs' static site files
|
||||
|
||||
6. Run Sphinx with the `sphinx-multiversion` wrapper to build the doc's static site files. These files will be outputted to the `cleanlab-docs/` directory.
|
||||
|
||||
## Generate the `versioning.js` file used to store the latest release tag name and commit hash
|
||||
|
||||
7. Get the latest release tag name and insert it in the `versioning.js` file. The `index.html` of each doc version will read this as a variable and display it beside the **stable** hyperlink.
|
||||
|
||||
8. Insert the latest commit hash in the `versioning.js` file. The `index.html` of each doc version will read this as a variable and display it beside the **developer** hyperlink.
|
||||
|
||||
9. Copy the `versioning.js` file to the `cleanlab-docs/` folder.
|
||||
|
||||
## If the workflow is **triggered by a new release**, generate the redirecting HTML which redirects site visits to the stable version
|
||||
|
||||
10. Insert the relative path to the stable docs in the `redirect-to-stable.html` file AKA the *redirecting HTML*.
|
||||
|
||||
11. Create a copy of the `redirect-to-stable.html` file to `cleanlab-docs/index.html` and `cleanlab-docs/index.html`.
|
||||
|
||||
## Deploy the static files
|
||||
|
||||
12. Deploy `cleanlab-docs/` folder to the `cleanlab/cleanlab-docs` repo's `master branch`.
|
||||
|
||||
|
||||
# Tips for editing docs/tutorials
|
||||
|
||||
## Tutorials
|
||||
|
||||
Each tutorial is a Jupyter notebook (unexecuted .ipynb file) that will be executed during CI for the version displayed at docs.cleanlab.ai using [nbsphinx](https://github.com/cleanlab/cleanlab/blob/31c939ff9aa487e9670b1a0f3f711a1d78448a91/docs/source/conf.py). Some basic [linting](https://github.com/cleanlab/nblint-action) is also applied to ensure proper notebook formatting such as no trailing newlines at the end of cells. Here are some tips when adding a new tutorial notebook:
|
||||
|
||||
1. Make sure to clear all Cell outputs before you `git commit` a tutorial. The outputs of cells should never be tracked in git, these outputs are automatically constructed for displaying on docs.cleanlab.ai during the CI which executes all notebooks in the folder **docs/source/**.
|
||||
|
||||
2. For cells which contain code that should **not** be executed during CI, make sure the cell-type is Markdown and use proper syntax to make contents look like code.
|
||||
|
||||
3. To suppress certain Jupyter cells that should not be shown on docs.cleanlab.ai web version of tutorial:
|
||||
```
|
||||
"metadata": {
|
||||
"nbsphinx": "hidden"
|
||||
}
|
||||
```
|
||||
This includes cells that install dependencies and cells that run tests to verify the notebook has executed correctly. These cells will still be visible when the notebook is run in Colab or locally in Jupyter, so make sure to add a comment explaining their purpose at the top.
|
||||
|
||||
4. If developing Notebook in virtualenv, make sure at the end to change the end of the raw .ipynb file to have the following:
|
||||
```
|
||||
"metadata": {
|
||||
"kernelspec": {
|
||||
"display_name": "Python 3 (ipykernel)",
|
||||
"language": "python",
|
||||
"name": "python3"
|
||||
}
|
||||
```
|
||||
instead of containing your own virtualenv in there. CI will FAIL if you instead list your own virtualenv here!
|
||||
|
||||
5. When adding dependencies to a tutorial:
|
||||
- Make sure to update **docs/requirements.txt** which lists all extra dependencies installed during CI to build the docs.
|
||||
- Add a comment in hidden cell not displayed on docs.cleanlab.ai stating which version of dependencies you used.
|
||||
- Think carefully whether each dependency is really necessary and if its future versions will be stable / compatible with future versions of existing dependencies.
|
||||
|
||||
6. Don't forget to update **docs/source/index.rst** with a short title and **docs/source/tutorials/index.rst** to ensure your tutorial properly linked. Otherwise it will not appear on docs.cleanlab.ai!
|
||||
|
||||
7. Ask yourself:
|
||||
- How can I make this tutorial run faster without sacrificing educational value? Perhaps use smaller subsample of the dataset, smaller/pretrained model, etc.
|
||||
- What sections of this tutorial are least vital? Consider creating a separate [Examples](https://github.com/cleanlab/examples) notebook that features those.
|
||||
|
||||
All of our tutorials are quickstart guides that should run quite fast. Longer/comprehensive notebooks are better added in [Examples](https://github.com/cleanlab/examples).
|
||||
|
||||
|
||||
## API Documentation
|
||||
|
||||
1. Verify your new docstrings adhere to our [documentation format guidelines](../DEVELOPMENT.md#documentation-style)
|
||||
|
||||
2. To ensure documentation for new source code files is linked from the main page, don't forget to update: **docs/source/index.rst**
|
||||
@@ -0,0 +1,35 @@
|
||||
@ECHO OFF
|
||||
|
||||
pushd %~dp0
|
||||
|
||||
REM Command file for Sphinx documentation
|
||||
|
||||
if "%SPHINXBUILD%" == "" (
|
||||
set SPHINXBUILD=sphinx-build
|
||||
)
|
||||
set SOURCEDIR=source
|
||||
set BUILDDIR=build
|
||||
|
||||
if "%1" == "" goto help
|
||||
|
||||
%SPHINXBUILD% >NUL 2>NUL
|
||||
if errorlevel 9009 (
|
||||
echo.
|
||||
echo.The 'sphinx-build' command was not found. Make sure you have Sphinx
|
||||
echo.installed, then set the SPHINXBUILD environment variable to point
|
||||
echo.to the full path of the 'sphinx-build' executable. Alternatively you
|
||||
echo.may add the Sphinx directory to PATH.
|
||||
echo.
|
||||
echo.If you don't have Sphinx installed, grab it from
|
||||
echo.https://www.sphinx-doc.org/
|
||||
exit /b 1
|
||||
)
|
||||
|
||||
%SPHINXBUILD% -M %1 %SOURCEDIR% %BUILDDIR% %SPHINXOPTS% %O%
|
||||
goto end
|
||||
|
||||
:help
|
||||
%SPHINXBUILD% -M help %SOURCEDIR% %BUILDDIR% %SPHINXOPTS% %O%
|
||||
|
||||
:end
|
||||
popd
|
||||
@@ -0,0 +1,31 @@
|
||||
jinja2==3.1.3
|
||||
sphinx==7.1.2
|
||||
sphinx-tabs==3.4.5
|
||||
nbconvert==7.13.1
|
||||
nbsphinx==0.9.3
|
||||
autodocsumm==0.2.12
|
||||
furo==2023.09.10
|
||||
ipython==8.0.1
|
||||
ipykernel==6.29.0
|
||||
ipywidgets==8.1.1
|
||||
sphinx-multiversion==0.2.4
|
||||
sphinx-copybutton==0.5.2
|
||||
sphinxcontrib-katex==0.9.9
|
||||
sphinxcontrib-gtagjs==0.2.1
|
||||
sphinx-jinja==2.0.2
|
||||
sphinx-autodoc-typehints==1.25.2
|
||||
sphinxext-opengraph==0.9.1
|
||||
requests==2.28.2
|
||||
sentence-transformers>=2.3.0
|
||||
speechbrain==0.5.13
|
||||
faiss-cpu==1.8.0
|
||||
datasets>=2.14.6
|
||||
xgboost
|
||||
matplotlib==3.7.4 # TODO: uncap version
|
||||
huggingface_hub==0.26.5 # TODO: uncap version
|
||||
torch==2.1.2 # TODO: uncap version
|
||||
torchvision==0.16.2 # TODO: uncap version
|
||||
torchaudio==2.1.2 # TODO: uncap version
|
||||
timm==0.9.12 # TODO: uncap version
|
||||
seaborn==0.13.2 # TODO: uncap version
|
||||
scikit-learn==1.5.2 # TODO: uncap version
|
||||
@@ -0,0 +1,6 @@
|
||||
Without this directory, Sphinx will warn:
|
||||
|
||||
> WARNING: html_static_path entry '_static' does not exist
|
||||
|
||||
This file exists to suppress that warning and because empty directories cannot
|
||||
be checked in to Git.
|
||||
@@ -0,0 +1,41 @@
|
||||
details {
|
||||
margin-bottom: 0.75rem;
|
||||
margin-top: 0.5rem;
|
||||
}
|
||||
|
||||
details summary {
|
||||
cursor: pointer;
|
||||
}
|
||||
|
||||
details summary > * {
|
||||
display: inline;
|
||||
}
|
||||
|
||||
details[open] summary {
|
||||
padding-bottom: 0.75rem;
|
||||
border-bottom: 2px dashed #ccc;
|
||||
}
|
||||
|
||||
details[open] {
|
||||
border-bottom: 2px dashed #ccc;
|
||||
}
|
||||
|
||||
h1 {
|
||||
font-size: 2em;
|
||||
}
|
||||
|
||||
h2 {
|
||||
font-size: 1.5em;
|
||||
}
|
||||
|
||||
h3 {
|
||||
font-size: 1.17em;
|
||||
}
|
||||
|
||||
h5 {
|
||||
font-size: .83em;
|
||||
}
|
||||
|
||||
h6 {
|
||||
font-size: .75em;
|
||||
}
|
||||
@@ -0,0 +1,26 @@
|
||||
<a style="padding-bottom: 0px;" class="sidebar-brand{% if logo %} centered{% endif %}" href="{{ pathto(master_doc) }}">
|
||||
{% block brand_content %}
|
||||
{%- if logo_url %}
|
||||
<div class="sidebar-logo-container">
|
||||
<img class="sidebar-logo" src="{{ logo_url }}" alt="Logo" />
|
||||
</div>
|
||||
{%- endif %}
|
||||
{%- if theme_light_logo and theme_dark_logo %}
|
||||
<div class="sidebar-logo-container">
|
||||
<img class="sidebar-logo only-light" src="{{ pathto('_static/' + theme_light_logo, 1) }}" alt="Light Logo" />
|
||||
<img class="sidebar-logo only-dark" src="{{ pathto('_static/' + theme_dark_logo, 1) }}" alt="Dark Logo" />
|
||||
</div>
|
||||
{%- endif %}
|
||||
{% if not theme_sidebar_hide_name %}
|
||||
<span style="margin-bottom:0px" class="sidebar-brand-text">
|
||||
{{ docstitle if docstitle else project }}
|
||||
</span>
|
||||
{%- endif %}
|
||||
{% endblock brand_content %}
|
||||
</a>
|
||||
|
||||
<div style="margin-left: 12px; margin-top: 4px;">
|
||||
<a style="margin-top: 6px;" class="github-button" href="https://github.com/cleanlab/cleanlab" data-size="large" data-show-count="true"
|
||||
aria-label="Star cleanlab/cleanlab on GitHub">Star</a>
|
||||
</div>
|
||||
|
||||
@@ -0,0 +1,66 @@
|
||||
{% extends "!page.html" %} {% block content %}
|
||||
|
||||
<noscript>
|
||||
<div class="admonition warning">
|
||||
<p class="admonition-title">Warning</p>
|
||||
<p>
|
||||
Parts of this site uses JavaScript, but your browser does not
|
||||
support it.
|
||||
</p>
|
||||
</div>
|
||||
</noscript>
|
||||
|
||||
{% if current_version %}
|
||||
|
||||
<!-- Start of Version Warning Banner -->
|
||||
|
||||
<p id="doc_ver_warning"></p>
|
||||
|
||||
<script type="text/javascript" src="{{ DOCS_SITE_URL }}versioning.js"></script>
|
||||
<script type="text/javascript">
|
||||
window.addEventListener("load", () => {
|
||||
const version_number = Version.version_number;
|
||||
const path_arr = window.location.pathname.split("/");
|
||||
|
||||
if (path_arr.includes("master")) {
|
||||
document.getElementById(
|
||||
"doc_ver_warning"
|
||||
).innerHTML = `<div class="admonition warning">
|
||||
<p class="admonition-title">Warning</p>
|
||||
<p>This version of the documentation corresponds to the master branch of <code class="docutils literal notranslate"><span class="pre">cleanlab</span></code> source code from <a href="https://github.com/cleanlab/cleanlab/">GitHub</a>. To see the documentation for the latest <code class="docutils literal notranslate"><span class="pre">pip</span></code>-installed version, click <a href="{{ DOCS_SITE_URL }}">here</a>.</p>
|
||||
</div>`;
|
||||
} else if (
|
||||
!path_arr.includes(version_number) &&
|
||||
!path_arr.includes("stable")
|
||||
) {
|
||||
document.getElementById("doc_ver_warning").innerHTML =
|
||||
`<div class="admonition warning">
|
||||
<p class="admonition-title">Warning</p>
|
||||
<p>This documentation is for an old version (<code class="docutils literal notranslate"><span class="pre">{{ current_version.name }}</span></code>) of <code class="docutils literal notranslate"><span class="pre">cleanlab</span></code>. To see the documentation for the latest stable version (<code class="docutils literal notranslate"><span class="pre">` +
|
||||
version_number +
|
||||
`</span></code>), click <a href="{{ DOCS_SITE_URL }}">here</a>.</p>
|
||||
</div>`;
|
||||
} else {
|
||||
document.getElementById("doc_ver_warning").remove();
|
||||
}
|
||||
});
|
||||
</script>
|
||||
|
||||
<!-- End of Version Warning Banner -->
|
||||
|
||||
{% endif %} {{ super() }} {% endblock %} {% block footer %} {{ super() }}
|
||||
|
||||
<script type="text/javascript">
|
||||
window.addEventListener("load", () => {
|
||||
let elements = document.getElementsByClassName("left-details");
|
||||
|
||||
elements[0].insertAdjacentHTML(
|
||||
"afterbegin",
|
||||
`<code class="docutils literal notranslate"><span class="pre">cleanlab</span></code> is distributed on <a href="https://pypi.org/project/cleanlab/">PyPI</a> and <a href="https://anaconda.org/conda-forge/cleanlab">conda</a>.`
|
||||
);
|
||||
});
|
||||
</script>
|
||||
|
||||
{% endblock %} {% block regular_scripts %}{{ super() }}
|
||||
<script async defer src="https://buttons.github.io/buttons.js"></script>
|
||||
{% endblock %}
|
||||
@@ -0,0 +1,9 @@
|
||||
<!DOCTYPE html>
|
||||
<html>
|
||||
<head>
|
||||
<title>Redirecting to the latest release/tag</title>
|
||||
<meta charset="utf-8" />
|
||||
<meta http-equiv="refresh" content="0; url=stable_url" />
|
||||
<link rel="canonical" href="https://docs.cleanlab.ai/" />
|
||||
</head>
|
||||
</html>
|
||||
@@ -0,0 +1,65 @@
|
||||
{% if versions %}
|
||||
|
||||
<!-- Start of versioning -->
|
||||
|
||||
<div class="sidebar-tree">
|
||||
<p class="caption" role="heading">
|
||||
<span class="caption-text">{{ _('Versions') }}</span>
|
||||
</p>
|
||||
<ul>
|
||||
<li class="toctree-l1">
|
||||
<a
|
||||
id="version_number"
|
||||
class="reference internal"
|
||||
href="{{ DOCS_SITE_URL }}"
|
||||
>stable</a
|
||||
>
|
||||
</li>
|
||||
<li class="toctree-l1">
|
||||
<a
|
||||
id="commit_hash"
|
||||
class="reference internal"
|
||||
href="{{DOCS_SITE_URL}}master/"
|
||||
>developer</a
|
||||
>
|
||||
</li>
|
||||
{% for version in RELEASE_VERSIONS %}
|
||||
<li class="toctree-l1">
|
||||
<a
|
||||
id="{{ version }}"
|
||||
class="reference internal"
|
||||
href="{{DOCS_SITE_URL}}{{version}}/"
|
||||
>{{version}}</a
|
||||
>
|
||||
</li>
|
||||
{% endfor %}
|
||||
</ul>
|
||||
</div>
|
||||
|
||||
<br>
|
||||
<br>
|
||||
|
||||
<script
|
||||
type="text/javascript"
|
||||
src="{{ DOCS_SITE_URL }}versioning.js"
|
||||
></script>
|
||||
|
||||
<script type="text/javascript">
|
||||
window.addEventListener("load", () => {
|
||||
const version_number = Version.version_number;
|
||||
const commit_hash = Version.commit_hash;
|
||||
|
||||
document.getElementById("version_number").innerHTML =
|
||||
"stable <code class='docutils literal notranslate'><span class='pre'> (" +
|
||||
version_number +
|
||||
")</span></code>";
|
||||
document.getElementById("commit_hash").innerHTML =
|
||||
"master <code class='docutils literal notranslate'><span class='pre'> (" +
|
||||
commit_hash.slice(0, 7) +
|
||||
"…)</span></code>";
|
||||
});
|
||||
</script>
|
||||
|
||||
<!-- End of versioning -->
|
||||
|
||||
{% endif %}
|
||||
@@ -0,0 +1,4 @@
|
||||
var Version = {
|
||||
version_number: "placeholder_version_number",
|
||||
commit_hash: "placeholder_commit_hash",
|
||||
};
|
||||
@@ -0,0 +1,12 @@
|
||||
benchmarking
|
||||
============
|
||||
|
||||
.. automodule:: cleanlab.benchmarking
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
|
||||
.. toctree::
|
||||
|
||||
noise_generation
|
||||
@@ -0,0 +1,8 @@
|
||||
noise_generation
|
||||
================
|
||||
|
||||
.. automodule:: cleanlab.benchmarking.noise_generation
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,8 @@
|
||||
classification
|
||||
==============
|
||||
|
||||
.. automodule:: cleanlab.classification
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,8 @@
|
||||
count
|
||||
=====
|
||||
|
||||
.. automodule:: cleanlab.count
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,8 @@
|
||||
data_valuation
|
||||
==============
|
||||
|
||||
.. automodule:: cleanlab.data_valuation
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,9 @@
|
||||
datalab
|
||||
=======
|
||||
|
||||
.. automodule:: cleanlab.datalab.datalab
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
:ignore-module-all:
|
||||
@@ -0,0 +1,10 @@
|
||||
.. tip::
|
||||
|
||||
This type of issue has the issue name `"{{issue_name}}"`.
|
||||
|
||||
Run a check for this particular kind of issue by calling :py:meth:`Datalab.find_issues() <cleanlab.datalab.datalab.Datalab.find_issues>` like so:
|
||||
|
||||
.. code-block:: python
|
||||
|
||||
# `lab` is a Datalab instance
|
||||
lab.find_issues(..., issue_types = {"{{issue_name}}": {}})
|
||||
@@ -0,0 +1,225 @@
|
||||
.. _issue_manager_creating_your_own:
|
||||
|
||||
Creating Your Own Issues Manager
|
||||
================================
|
||||
|
||||
|
||||
|
||||
This guide walks through the process of creating your own
|
||||
:py:class:`IssueManager <cleanlab.datalab.internal.issue_manager.issue_manager.IssueManager>`
|
||||
to detect a custom-defined type of issue alongside the pre-defined issue types in
|
||||
:py:class:`Datalab <cleanlab.datalab.datalab.Datalab>`.
|
||||
|
||||
.. seealso::
|
||||
|
||||
- :py:meth:`register <cleanlab.datalab.internal.issue_manager_factory.register>`:
|
||||
You can either use this function at runtime to register a new issue manager:
|
||||
|
||||
.. code-block:: python
|
||||
|
||||
from cleanlab.datalab.internal.issue_manager_factory import register
|
||||
register(MyIssueManager) # Defaults to task="classification"
|
||||
# register(MyIssueManagerForRegression, task="regression") # Alternative for regression tasks
|
||||
|
||||
or add as a decorator to the class definition (currently only works for classification tasks):
|
||||
|
||||
.. code-block:: python
|
||||
|
||||
@register
|
||||
class MyIssueManager(IssueManager):
|
||||
...
|
||||
|
||||
Prerequisites
|
||||
-------------
|
||||
|
||||
As a starting point for this guide, we'll import the necessary things for the next section and create a dummy dataset.
|
||||
|
||||
.. note::
|
||||
|
||||
.. include:: ../optional_dependencies.rst
|
||||
|
||||
.. code-block:: python
|
||||
|
||||
|
||||
import numpy as np
|
||||
import pandas as pd
|
||||
from cleanlab import IssueManager
|
||||
|
||||
# Create a dummy dataset
|
||||
N = 20
|
||||
data = pd.DataFrame(
|
||||
{
|
||||
"text": [f"example {i}" for i in range(N)],
|
||||
"label": np.random.randint(0, 2, N),
|
||||
},
|
||||
)
|
||||
|
||||
|
||||
Implementing IssueManagers
|
||||
--------------------------
|
||||
|
||||
.. _basic_issue_manager:
|
||||
|
||||
Basic Issue Check
|
||||
~~~~~~~~~~~~~~~~~
|
||||
|
||||
|
||||
To create a basic issue manager, inherit from the
|
||||
:py:class:`IssueManager <cleanlab.datalab.internal.issue_manager.issue_manager.IssueManager>` class,
|
||||
assign a name to the class as the class-variable, `issue_name`, and implement the
|
||||
:py:meth:`find_issues <cleanlab.datalab.internal.issue_manager.issue_manager.IssueManager.find_issues>` method.
|
||||
|
||||
The :py:meth:`find_issues <cleanlab.datalab.internal.issue_manager.issue_manager.IssueManager.find_issues>`
|
||||
method should mark each example in the dataset as an issue or not with a boolean array.
|
||||
It should also provide a score for each example in the dataset that quantifies the quality of the example
|
||||
with regards to the issue.
|
||||
|
||||
.. code-block:: python
|
||||
|
||||
class Basic(IssueManager):
|
||||
# Assign a name to the issue
|
||||
issue_name = "basic"
|
||||
def find_issues(self, **kwargs) -> None:
|
||||
# Compute scores for each example
|
||||
scores = np.random.rand(len(self.datalab.data))
|
||||
|
||||
# Construct a dataframe where examples are marked for issues
|
||||
# and the score for each example is included.
|
||||
self.issues = pd.DataFrame(
|
||||
{
|
||||
f"is_{self.issue_name}_issue" : scores < 0.1,
|
||||
self.issue_score_key : scores,
|
||||
},
|
||||
)
|
||||
|
||||
# Score the dataset as a whole based on this issue type
|
||||
self.summary = self.make_summary(score = scores.mean())
|
||||
|
||||
|
||||
.. _intermediate_issue_manager:
|
||||
|
||||
Intermediate Issue Check
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
|
||||
To create an intermediate issue:
|
||||
|
||||
- Perform the same steps as in the :ref:`basic issue check <basic_issue_manager>` section.
|
||||
- Populate the `info` attribute with a dictionary of information about the identified issues.
|
||||
|
||||
The information can be included in a report generated by :py:class:`Datalab <cleanlab.datalab.datalab.Datalab>`,
|
||||
if you add any of the keys to the `verbosity_levels` class-attribute.
|
||||
Optionally, you can also add a description of the type of issue this issue manager handles to the `description` class-attribute.
|
||||
|
||||
.. code-block:: python
|
||||
|
||||
class Intermediate(IssueManager):
|
||||
issue_name = "intermediate"
|
||||
# Add a dictionary of information to include in the report
|
||||
verbosity_levels = {
|
||||
0: [],
|
||||
1: ["std"],
|
||||
2: ["raw_scores"],
|
||||
}
|
||||
# Add a description of the issue
|
||||
description = "Intermediate issues are a bit more involved than basic issues."
|
||||
def find_issues(self, *, intermediate_arg: int, **kwargs) -> None:
|
||||
N = len(self.datalab.data)
|
||||
raw_scores = np.random.rand(N)
|
||||
std = raw_scores.std()
|
||||
threshold = min(0, raw_scores.mean() - std)
|
||||
sin_filter = np.sin(intermediate_arg * np.arange(N) / N)
|
||||
kernel = sin_filter ** 2
|
||||
scores = kernel * raw_scores
|
||||
self.issues = pd.DataFrame(
|
||||
{
|
||||
f"is_{self.issue_name}_issue" : scores < threshold,
|
||||
self.issue_score_key : scores,
|
||||
},
|
||||
)
|
||||
self.summary = self.make_summary(score = scores.mean())
|
||||
|
||||
# Useful information that will be available in the Datalab instance
|
||||
self.info = {
|
||||
"std": std,
|
||||
"raw_scores": raw_scores,
|
||||
"kernel": kernel,
|
||||
}
|
||||
|
||||
Advanced Issue Check
|
||||
~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
There could be different types of issues detected in a dataset. A local issue which affects individual data points in a dataset and can be tracked via `Datalab.issues` dataframe (to see which data points are exhibiting this type of issue). Alternatively, a global issue which affects the overall dataset but is not easily attributable to individual data points (hard to say one data point exhibits the issue but another does not). Even for global issues, we recommend trying to assign a per data point score (and boolean) if possible, see the Non-IID IssueManager as an example of this. Note that a global issue must have num_issues greater than 0 in its `issue_summary`, otherwise it won't show up in `Datalab.report()` by default.
|
||||
|
||||
|
||||
Use with Datalab
|
||||
----------------
|
||||
|
||||
We can create a
|
||||
:py:class:`Datalab <cleanlab.datalab.datalab.Datalab>`
|
||||
instance and run issue checks with the custom issue managers we created like so:
|
||||
|
||||
|
||||
.. code-block:: python
|
||||
|
||||
from cleanlab.datalab.internal.issue_manager_factory import register
|
||||
from cleanlab import Datalab
|
||||
|
||||
|
||||
# Register the issue manager
|
||||
for issue_manager in [Basic, Intermediate]:
|
||||
register(issue_manager)
|
||||
|
||||
# Instantiate a datalab instance
|
||||
datalab = Datalab(data, label_name="label")
|
||||
|
||||
# Run the issue check
|
||||
issue_types = {"basic": {}, "intermediate": {"intermediate_arg": 2}}
|
||||
datalab.find_issues(issue_types=issue_types)
|
||||
|
||||
# Print report
|
||||
datalab.report(verbosity=0)
|
||||
|
||||
|
||||
The report will look something like this:
|
||||
|
||||
.. code-block:: text
|
||||
|
||||
Here is a summary of the different kinds of issues found in the data:
|
||||
|
||||
issue_type score num_issues
|
||||
basic 0.477762 2
|
||||
intermediate 0.286455 0
|
||||
|
||||
(Note: A lower score indicates a more severe issue across all examples in the dataset.)
|
||||
|
||||
|
||||
------------------------------------------- basic issues -------------------------------------------
|
||||
|
||||
Number of examples with this issue: 2
|
||||
Overall dataset quality in terms of this issue: 0.4778
|
||||
|
||||
Examples representing most severe instances of this issue:
|
||||
is_basic_issue basic_score
|
||||
13 True 0.003042
|
||||
8 True 0.058117
|
||||
11 False 0.121908
|
||||
15 False 0.169312
|
||||
17 False 0.229044
|
||||
|
||||
|
||||
--------------------------------------- intermediate issues ----------------------------------------
|
||||
|
||||
About this issue:
|
||||
Intermediate issues are a bit more involved than basic issues.
|
||||
|
||||
Number of examples with this issue: 0
|
||||
Overall dataset quality in terms of this issue: 0.2865
|
||||
|
||||
Examples representing most severe instances of this issue:
|
||||
is_intermediate_issue intermediate_score kernel
|
||||
0 False 0.000000 0.0
|
||||
1 False 0.007059 0.009967
|
||||
3 False 0.010995 0.087332
|
||||
2 False 0.016296 0.03947
|
||||
11 False 0.019459 0.794251
|
||||
@@ -0,0 +1,29 @@
|
||||
Generating Cluster IDs
|
||||
======================
|
||||
|
||||
The underperforming group issue manager provides the option for passing pre-computed
|
||||
cluster IDs to `find_issues`. These cluster IDs can be obtained by clustering
|
||||
the features using algorithms such as K-Means, DBSCAN, HDBSCAN etc. Note that
|
||||
|
||||
* K-Means requires specifying the number of clusters explicitly.
|
||||
* DBSCAN is sensitive to the choice of `eps` (radius) and `min_samples` (minimum samples for each cluster).
|
||||
|
||||
|
||||
Example:
|
||||
|
||||
.. code-block:: python
|
||||
|
||||
import datalab
|
||||
from sklearn.cluster import KMeans
|
||||
features, labels = your_data() # Get features and labels
|
||||
pred_probs = get_pred_probs() # Get prediction probabilities for all samples
|
||||
# Group features into 8 clusters
|
||||
clusterer = KMeans(n_clusters=5)
|
||||
clusterer.fit(features)
|
||||
cluster_ids = clusterer.labels_
|
||||
lab = Datalab(data={"features": features, "y": labels}, label_name="y")
|
||||
issue_types = {"underperforming_group": {"cluster_ids": cluster_ids}}
|
||||
lab.find_issues(features=features, pred_probs=pred_probs, issue_types=issue_types)
|
||||
|
||||
|
||||
|
||||
@@ -0,0 +1,41 @@
|
||||
Datalab guides
|
||||
==============
|
||||
|
||||
Guides for using Datalab and understanding the issues it detects.
|
||||
|
||||
.. note::
|
||||
|
||||
.. include:: ../optional_dependencies.rst
|
||||
|
||||
|
||||
Types of issues
|
||||
---------------
|
||||
|
||||
Guides to use Datalab with greater control, selecting what issues to search for and what nondefault settings to use for detecting them.
|
||||
|
||||
.. toctree::
|
||||
:maxdepth: 3
|
||||
|
||||
issue_type_description
|
||||
|
||||
Customizing issue types
|
||||
-----------------------
|
||||
|
||||
Guides (for developers) to create a custom issue type that Datalab audits for together with its built-in issue types.
|
||||
|
||||
.. toctree::
|
||||
:maxdepth: 3
|
||||
|
||||
custom_issue_manager
|
||||
|
||||
|
||||
Cleanlab Studio (Easy Mode)
|
||||
---------------------------
|
||||
|
||||
`Cleanlab Studio <https://cleanlab.ai/blog/data-centric-ai/>`_ is a fully automated platform that can detect the same data issues as this package, as well as `many more types of issues <https://help.cleanlab.ai/guide/concepts/cleanlab_columns/>`_, all without you having to do any Machine Learning (or even write any code). Beyond being 100x faster to use and producing more useful results, `Cleanlab Studio <https://cleanlab.ai/blog/data-centric-ai/>`_ also provides an intelligent data correction interface for you to quickly fix the issues detected in your dataset (a single data scientist can fix millions of data points thanks to AI suggestions).
|
||||
|
||||
`Cleanlab Studio <https://cleanlab.ai/blog/data-centric-ai/>`_ offers a powerful AutoML system (with Foundation models) that is useful for more than improving data quality. With a few clicks, you can: find + fix issues in your dataset, identify the best type of ML model and train/tune it, and deploy this model to serve accurate predictions for new data. Also use the same AutoML to auto-label large datasets (a single user can label millions of data points thanks to powerful Foundation models). `Try Cleanlab Studio for free! <https://cleanlab.ai/signup/>`_
|
||||
|
||||
.. image:: https://raw.githubusercontent.com/cleanlab/assets/master/cleanlab/ml-with-cleanlab-studio.png
|
||||
:width: 800
|
||||
:alt: Stages of modern AI pipeline that can now be automated with Cleanlab Studio
|
||||
@@ -0,0 +1,857 @@
|
||||
Datalab Issue Types
|
||||
*******************
|
||||
|
||||
|
||||
Types of issues Datalab can detect
|
||||
===================================
|
||||
|
||||
This page describes the various types of issues that Datalab can detect in a dataset.
|
||||
For each type of issue, we explain: what it says about your data if detected, why this matters, and what parameters you can optionally specify to control the detection of this issue.
|
||||
|
||||
In case you didn't know: you can alternatively use `Cleanlab Studio <https://cleanlab.ai/blog/data-centric-ai/>`_ to detect the same data issues as this package, plus `many more types of issues <https://help.cleanlab.ai/guide/concepts/cleanlab_columns/>`_, all without having to do any Machine Learning (or even write any code).
|
||||
|
||||
|
||||
.. include:: table.rst
|
||||
|
||||
|
||||
Estimates for Each Issue Type
|
||||
------------------------------
|
||||
|
||||
Datalab produces three estimates for **each** type of issue (called say `<ISSUE_NAME>` here):
|
||||
|
||||
|
||||
1. A numeric quality score `<ISSUE_NAME>_score` (between 0 and 1) estimating how severe this issue is exhibited in each example from a dataset. Examples with higher scores are less likely to suffer from this issue. Access these via: the :py:attr:`Datalab.issues <cleanlab.datalab.datalab.Datalab.issues>` attribute or the method :py:meth:`Datalab.get_issues(\<ISSUE_NAME\>) <cleanlab.datalab.datalab.Datalab.get_issues>`.
|
||||
2. A Boolean `is_<ISSUE_NAME>_issue` flag for each example from a dataset. Examples where this has value `True` are those estimated to exhibit this issue. Access these via: the :py:attr:`Datalab.issues <cleanlab.datalab.datalab.Datalab.issues>` attribute or the method :py:meth:`Datalab.get_issues(\<ISSUE_NAME\>) <cleanlab.datalab.datalab.Datalab.get_issues>`.
|
||||
3. An overall dataset quality score (between 0 and 1), quantifying how severe this issue is overall across the entire dataset. Datasets with higher scores do not exhibit this issue as badly overall. Access these via: the :py:attr:`Datalab.issue_summary <cleanlab.datalab.datalab.Datalab.issue_summary>` attribute or the method :py:meth:`Datalab.get_issue_summary(\<ISSUE_NAME\>) <cleanlab.datalab.datalab.Datalab.get_issue_summary>`.
|
||||
|
||||
**Example (for the outlier issue type)**
|
||||
|
||||
.. code-block:: python
|
||||
|
||||
issue_name = "outlier" # how to reference the outlier issue type in code
|
||||
issue_score = "outlier_score" # name of column with quality scores for the outlier issue type, atypical datapoints receive lower scores
|
||||
is_issue = "is_outlier_issue" # name of Boolean column flagging which datapoints are considered outliers in the dataset
|
||||
|
||||
**Dataset vs. data point level issues**
|
||||
|
||||
Some issues are primarily about the overall dataset (e.g. non-IID, class imbalance, underperforming group), whereas others are primarily about individual examples (e.g. label issue, outlier, near duplicate, null, etc). The former issue types should be first investigated via the global score from :py:meth:`Datalab.get_issue_summary <cleanlab.datalab.datalab.Datalab.get_issue_summary>`, as the per-example results for such issues from :py:meth:`Datalab.get_issues <cleanlab.datalab.datalab.Datalab.get_issues>` require more expertise to interpret.
|
||||
|
||||
Inputs to Datalab
|
||||
-----------------
|
||||
|
||||
Datalab estimates various issues based on the four inputs below.
|
||||
Each input is optional, if you do not provide it, Datalab will skip checks for those types of issues that require this input.
|
||||
|
||||
1. ``label_name`` - a field in the dataset that the stores the annotated class for each example in a multi-class classification dataset.
|
||||
2. ``pred_probs`` - predicted class probabilities output by your trained model for each example in the dataset (these should be out-of-sample, eg. produced via cross-validation).
|
||||
3. ``features`` - numeric vector representations of the features for each example in the dataset. These may be embeddings from a (pre)trained model, or just a numerically-transformed version of the original data features.
|
||||
4. ``knn_graph`` - K nearest neighbor graph represented as a sparse matrix of dissimilarity values between examples in the dataset. If both `knn_graph` and `features` are provided, the `knn_graph` takes precedence, and if only `features` is provided, then a `knn_graph` is internally constructed based on the (either euclidean or cosine) distance between different examples’ features.
|
||||
|
||||
|
||||
Label Issue
|
||||
-----------
|
||||
|
||||
Examples whose given label is estimated to be potentially incorrect (e.g. due to annotation error) are flagged as having label issues.
|
||||
Datalab estimates which examples appear mislabeled as well as a numeric label quality score for each, which quantifies the likelihood that an example is correctly labeled.
|
||||
|
||||
For now, Datalab can only detect label issues in multi-class classification datasets, regression datasets, and multi-label classification datasets.
|
||||
The cleanlab library has alternative methods you can use to detect label issues in other types of datasets (multi-annotator, token classification, etc.).
|
||||
|
||||
Label issues are calculated based on provided `pred_probs` from a trained model. If you do not provide this argument, but you do provide `features`, then a K Nearest Neighbor model will be fit to produce `pred_probs` based on your `features`. Otherwise if neither `pred_probs` nor `features` is provided, then this type of issue will not be considered.
|
||||
For the most accurate results, provide out-of-sample `pred_probs` which can be obtained for a dataset via `cross-validation <https://docs.cleanlab.ai/stable/tutorials/pred_probs_cross_val.html>`_.
|
||||
|
||||
Having mislabeled examples in your dataset may hamper the performance of supervised learning models you train on this data.
|
||||
For evaluating models or performing other types of data analytics, mislabeled examples may lead you to draw incorrect conclusions.
|
||||
To handle mislabeled examples, you can either filter out the data with label issues or try to correct their labels.
|
||||
|
||||
Learn more about the method used to detect label issues in our paper: `Confident Learning: Estimating Uncertainty in Dataset Labels <https://arxiv.org/abs/1911.00068>`_
|
||||
|
||||
.. testsetup:: *
|
||||
|
||||
import numpy as np
|
||||
from cleanlab import Datalab
|
||||
from sklearn.linear_model import LogisticRegression
|
||||
from sklearn.model_selection import cross_val_predict
|
||||
|
||||
# Load a dataset
|
||||
np.random.seed(0)
|
||||
|
||||
X = np.random.rand(100, 10)
|
||||
X[-1] = X[-2] # Create an exact-duplicate example
|
||||
y = np.random.randint(0, 3, 100)
|
||||
|
||||
X[y == 1] -= 0.85 # Add noise to the features of class 1
|
||||
X[y == 2] += 0.85 # Add noise to the features of class 2
|
||||
|
||||
y[-3] = {0: 1, 1: 2, 2: 0}[y[-3]] # Swap the label of the example at index -3
|
||||
|
||||
clf = LogisticRegression(random_state=0)
|
||||
pred_probs = cross_val_predict(clf, X, y, cv=3, method="predict_proba")
|
||||
|
||||
data = {"features": X, "labels": y}
|
||||
|
||||
lab = Datalab(data, label_name="labels", task="classification")
|
||||
|
||||
.. testsetup::
|
||||
|
||||
lab.find_issues(features=X, pred_probs=pred_probs)
|
||||
lab.find_issues(features=X, pred_probs=pred_probs, issue_types={"data_valuation": {}})
|
||||
|
||||
Some metadata about label issues is stored in the `issues` attribute of the Datalab object.
|
||||
Let's look at one way to access this information.
|
||||
|
||||
.. testcode::
|
||||
|
||||
lab.get_issues("label").sort_values("label_score").head(5)
|
||||
|
||||
The output will look something like this:
|
||||
|
||||
.. testoutput::
|
||||
|
||||
is_label_issue label_score given_label predicted_label
|
||||
97 True 0.064045 0 2
|
||||
58 False 0.680894 2 2
|
||||
41 False 0.746043 0 0
|
||||
4 False 0.794894 2 2
|
||||
98 False 0.802911 1 1
|
||||
|
||||
``is_label_issue``
|
||||
~~~~~~~~~~~~~~~~~~
|
||||
|
||||
A boolean column that flags examples with label issues.
|
||||
If `True`, the example is estimated to have a label issue.
|
||||
If `False`, the example is estimated to not have a label issue.
|
||||
|
||||
``label_score``
|
||||
~~~~~~~~~~~~~~~
|
||||
|
||||
A numeric column that gives the label quality score for each example.
|
||||
The score lies between 0 and 1.
|
||||
The lower the score, the less likely the given label is to be correct.
|
||||
|
||||
|
||||
``given_label``
|
||||
~~~~~~~~~~~~~~~
|
||||
|
||||
A column of the actual labels as provided in the original dataset.
|
||||
|
||||
``predicted_label``
|
||||
~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
A column of the predicted labels for each example. This column may contain different labels than the given label, especially when the example is estimated to have a label issue or when a model predicts a different label than the given label.
|
||||
|
||||
.. jinja ::
|
||||
|
||||
{% with issue_name = "label" %}
|
||||
{% include "cleanlab/datalab/guide/_templates/issue_types_tip.rst" %}
|
||||
{% endwith %}
|
||||
|
||||
|
||||
Outlier Issue
|
||||
-------------
|
||||
|
||||
Examples that are very different from the rest of the dataset (i.e. potentially out-of-distribution or rare/anomalous instances).
|
||||
|
||||
Outlier issues are calculated based on provided `features` , `knn_graph` , or `pred_probs`.
|
||||
If you do not provide one of these arguments, this type of issue will not be considered.
|
||||
This article describes how outlier issues are detected in a dataset: `https://cleanlab.ai/blog/outlier-detection/ <https://cleanlab.ai/blog/outlier-detection/>`_.
|
||||
|
||||
When based on `features` or `knn_graph`, the outlier quality of each example is scored inversely proportional to its distance to its K nearest neighbors in the dataset.
|
||||
|
||||
When based on `pred_probs`, the outlier quality of each example is scored inversely proportional to the uncertainty in its prediction.
|
||||
|
||||
Modeling data with outliers may have unexpected consequences.
|
||||
Closely inspect them and consider removing some outliers that may be negatively affecting your models.
|
||||
|
||||
|
||||
Learn more about the methods used to detect outliers in our article: `Out-of-Distribution Detection via Embeddings or Predictions <https://cleanlab.ai/blog/outlier-detection/>`_
|
||||
|
||||
Some metadata about outlier issues is stored in the `issues` attribute of the Datalab object.
|
||||
Let's look at one way to access this information.
|
||||
|
||||
.. testcode::
|
||||
|
||||
lab.get_issues("outlier").sort_values("outlier_score").head(5)
|
||||
|
||||
The output will look something like this:
|
||||
|
||||
.. testoutput::
|
||||
|
||||
is_outlier_issue outlier_score
|
||||
98 True 0.011562
|
||||
62 False 0.019657
|
||||
22 False 0.035243
|
||||
1 False 0.040907
|
||||
42 False 0.056865
|
||||
|
||||
|
||||
|
||||
``is_outlier_issue``
|
||||
~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
A boolean column, where `True` indicates that an example is identified as an outlier.
|
||||
|
||||
``outlier_score``
|
||||
~~~~~~~~~~~~~~~~~
|
||||
|
||||
A numeric column with scores between 0 and 1.
|
||||
A smaller value for an example indicates that it is less common or typical in the dataset, suggesting that it is more likely to be an outlier.
|
||||
|
||||
|
||||
.. jinja ::
|
||||
|
||||
{% with issue_name = "outlier" %}
|
||||
{% include "cleanlab/datalab/guide/_templates/issue_types_tip.rst" %}
|
||||
{% endwith %}
|
||||
|
||||
(Near) Duplicate Issue
|
||||
----------------------
|
||||
|
||||
A (near) duplicate issue refers to two or more examples in a dataset that are extremely similar to each other, relative to the rest of the dataset.
|
||||
The examples flagged with this issue may be exactly duplicated, or lie atypically close together when represented as vectors (i.e. feature embeddings).
|
||||
Near duplicated examples may record the same information with different:
|
||||
|
||||
- Abbreviations, misspellings, typos, formatting, etc. in text data.
|
||||
- Compression formats, resolutions, or sampling rates in image, video, and audio data.
|
||||
- Minor variations which naturally occur in many types of data (e.g. translated versions of an image).
|
||||
|
||||
Near Duplicate issues are calculated based on provided `features` or `knn_graph`.
|
||||
If you do not provide one of these arguments, this type of issue will not be considered.
|
||||
|
||||
Datalab defines near duplicates as those examples whose distance to their nearest neighbor (in the space of provided `features`) in the dataset is less than `c * D`, where `0 < c < 1` is a small constant, and `D` is the median (over the full dataset) of such distances between each example and its nearest neighbor.
|
||||
Scoring the numeric quality of an example in terms of the near duplicate issue type is done proportionally to its distance to its nearest neighbor.
|
||||
|
||||
Including near-duplicate examples in a dataset may negatively impact a ML model's generalization performance and lead to overfitting.
|
||||
In particular, it is questionable to include examples in a test dataset which are (nearly) duplicated in the corresponding training dataset.
|
||||
More generally, examples which happen to be duplicated can affect the final modeling results much more than other examples — so you should at least be aware of their presence.
|
||||
|
||||
Some metadata about near-duplicate issues is stored in the `issues` attribute of the Datalab object.
|
||||
Let's look at one way to access this information.
|
||||
|
||||
.. testcode::
|
||||
|
||||
lab.get_issues("near_duplicate").sort_values("near_duplicate_score").head(5)
|
||||
|
||||
The output will look something like this:
|
||||
|
||||
.. testoutput::
|
||||
|
||||
is_near_duplicate_issue near_duplicate_score near_duplicate_sets distance_to_nearest_neighbor
|
||||
36 True 0.066009 [11, 80] 0.003906
|
||||
11 True 0.066009 [36] 0.003906
|
||||
80 True 0.093245 [36] 0.005599
|
||||
27 False 0.156720 [] 0.009751
|
||||
72 False 0.156720 [] 0.009751
|
||||
|
||||
|
||||
``is_near_duplicate_issue``
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
A boolean column, where `True` indicates that an example is identified as either a near- or exact-duplicate of other examples in the dataset.
|
||||
|
||||
``near_duplicate_score``
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
A numeric column with scores between 0 and 1. The lower the score, the more likely the example is to be a near-duplicate of another example in the dataset.
|
||||
|
||||
Exact duplicates are assigned a score of 0, while near-duplicates are assigned a score close to 0.
|
||||
|
||||
``near_duplicate_sets``
|
||||
~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
A column of lists of integers. The i-th list contains the indices of examples that are considered near-duplicates of example i (not including example i).
|
||||
|
||||
``distance_to_nearest_neighbor``
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
A numeric column that represents the distance between each example and its nearest neighbor in the dataset.
|
||||
The distance is calculated based on the provided `features` or `knn_graph`, and is directly related to the `near_duplicate_score`.
|
||||
A smaller distance indicates that the example is similar to another example in the dataset.
|
||||
|
||||
.. jinja ::
|
||||
|
||||
{% with issue_name = "near_duplicate" %}
|
||||
{% include "cleanlab/datalab/guide/_templates/issue_types_tip.rst" %}
|
||||
{% endwith %}
|
||||
|
||||
Non-IID Issue
|
||||
-------------
|
||||
|
||||
Whether the overall dataset exhibits statistically significant violations of the IID assumption like: changepoints or shift, drift, autocorrelation, etc. The specific form of violation considered is whether the examples are ordered within the dataset such that almost adjacent examples tend to have more similar feature values. If you care about this check, do **not** first shuffle your dataset -- this check is entirely based on the sequential order of your data. Learn more via our blog: `https://cleanlab.ai/blog/non-iid-detection/ <https://cleanlab.ai/blog/non-iid-detection/>`_
|
||||
|
||||
The Non-IID issue is detected based on provided `features` or `knn_graph`. If you do not provide one of these arguments, this type of issue will not be considered. While the Non-IID check produces per-example information, it is primarily about assessing the overall dataset rather than assessing individual examples. So pay more attention to the overall dataset Non-IID score obtained via :py:meth:`Datalab.get_issue_summary("non_iid") <cleanlab.datalab.datalab.Datalab.get_issue_summary>` than the per-example scores.
|
||||
|
||||
The Non-IID issue is really a dataset-level check, not a per-datapoint level check (either a dataset violates the IID assumption or it doesn't). The per-datapoint scores returned for Non-IID issues merely highlight which datapoints you might focus on to better understand this dataset-level issue - there is not necessarily something specifically wrong with these specific datapoints.
|
||||
|
||||
Mathematically, the **overall** Non-IID score for the dataset is defined as the p-value of a statistical test for whether the distribution of *index-gap* values differs between group A vs. group B defined as follows. For a pair of examples in the dataset `x1, x2`, we define their *index-gap* as the distance between the indices of these examples in the ordering of the data (e.g. if `x1` is the 10th example and `x2` is the 100th example in the dataset, their index-gap is 90). We construct group A from pairs of examples which are amongst the K nearest neighbors of each other, where neighbors are defined based on the provided `knn_graph` or via distances in the space of the provided vector `features` . Group B is constructed from random pairs of examples in the dataset.
|
||||
|
||||
The Non-IID quality score for each example `x` is defined via a similarly computed p-value but with Group A constructed from the K nearest neighbors of `x` and Group B constructed from random examples from the dataset paired with `x`. Learn more about this method in our paper: `Detecting Dataset Drift and Non-IID Sampling via k-Nearest Neighbors <https://arxiv.org/abs/2305.15696>`_ (or the associated `blogpost <https://cleanlab.ai/blog/non-iid-detection/>`_).
|
||||
|
||||
The assumption that examples in a dataset are Independent and Identically Distributed (IID) is fundamental to proper modeling. Detecting all possible violations of the IID assumption is statistically impossible. This issue type only considers specific forms of violation where examples that tend to be closer together in the dataset ordering also tend to have more similar feature values. This includes scenarios where:
|
||||
|
||||
- The underlying distribution from which examples stem is evolving/drifting over time (not identically distributed).
|
||||
- An example can influence the values of future examples in the dataset (not independent).
|
||||
|
||||
For datasets with low non-IID score, you should consider why your data are not IID and act accordingly. For example, if the data distribution is drifting over time, consider employing a time-based train/test split instead of a random partition. Note that shuffling the data ahead of time will ensure a good non-IID score, but this is not always a fix to the underlying problem (e.g. future deployment data may stem from a different distribution, or you may overlook the fact that examples influence each other). We thus recommend **not** shuffling your data to be able to diagnose this issue if it exists.
|
||||
|
||||
Some metadata about non-IID issues is stored in the `issues` attribute of the Datalab object.
|
||||
Let's look at one way to access this information.
|
||||
|
||||
.. testcode::
|
||||
|
||||
lab.get_issues("non_iid").sort_values("non_iid_score").head(5)
|
||||
|
||||
The output will look something like this:
|
||||
|
||||
.. testoutput::
|
||||
|
||||
is_non_iid_issue non_iid_score
|
||||
24 False 0.681458
|
||||
37 False 0.804582
|
||||
64 False 0.810646
|
||||
80 False 0.815691
|
||||
78 False 0.834293
|
||||
|
||||
``is_non_iid_issue``
|
||||
~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
A boolean column, where `True` values indicate that the dataset exhibits statistically significant violations of the IID assumption.
|
||||
If the overall dataset does not appear to be Non-IID (p-value > 0.05), then all entries in this column will be `False`.
|
||||
If the dataset appears to be Non-IID (p-value < 0.05), then one entry will be `True`, specifically the example with the lowest `non_iid_score`.
|
||||
We do not recommend interpreting the per-example boolean values, as the Non-IID check is more about the overall dataset.
|
||||
|
||||
``non_iid_score``
|
||||
~~~~~~~~~~~~~~~~~
|
||||
|
||||
A numeric column with scores between 0 and 1, containing the Non-IID quality scores for each example.
|
||||
Learn more via our `blogpost <https://cleanlab.ai/blog/non-iid-detection/>`_.
|
||||
|
||||
Be cautious when interpreting the non-IID issue score for individual examples.
|
||||
The dataset as a whole receives a p-value for our non-IID test (obtained via :py:meth:`Datalab.get_issue_summary("non_iid") <cleanlab.datalab.datalab.Datalab.get_issue_summary>`), which better indicates whether the dataset exhibits non-IID behavior.
|
||||
|
||||
When this p-value is low, you can use the per-example non-IID scores to identify which examples to look at for better understanding this non-IID behavior.
|
||||
|
||||
.. jinja ::
|
||||
|
||||
{% with issue_name = "non_iid" %}
|
||||
{% include "cleanlab/datalab/guide/_templates/issue_types_tip.rst" %}
|
||||
{% endwith %}
|
||||
|
||||
Class Imbalance Issue
|
||||
---------------------
|
||||
|
||||
Class imbalance is diagnosed just using the `labels` provided as part of the dataset. The overall class imbalance quality score of a dataset is the proportion of examples belonging to the rarest class `q`. If this proportion `q` falls below a threshold, then we say this dataset suffers from the class imbalance issue.
|
||||
|
||||
In a dataset identified as having class imbalance, the class imbalance quality score for each example is set equal to `q` if it is labeled as the rarest class, and is equal to 1 for all other examples.
|
||||
|
||||
Class imbalance in a dataset can lead to subpar model performance for the under-represented class. Consider collecting more data from the under-represented class, or at least take special care while modeling via techniques like over/under-sampling, SMOTE, asymmetric class weighting, etc.
|
||||
|
||||
This issue-type is more about the overall dataset vs. individual data points. If severe class imbalance is detected, Datalab will flag the individual data points from the minority class.
|
||||
|
||||
Some metadata about class imbalance issues is stored in the `issues` attribute of the Datalab object.
|
||||
Let's look at one way to access this information.
|
||||
|
||||
.. testcode::
|
||||
|
||||
lab.get_issues("class_imbalance").sort_values("class_imbalance_score").head(5)
|
||||
|
||||
The output will look something like this:
|
||||
|
||||
.. testoutput::
|
||||
|
||||
is_class_imbalance_issue class_imbalance_score given_label
|
||||
27 False 0.28 2
|
||||
72 False 0.28 2
|
||||
75 False 0.28 2
|
||||
33 False 0.28 2
|
||||
68 False 0.28 2
|
||||
|
||||
``is_class_imbalance_issue``
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
A boolean column, where `True` indicates which examples belong to the minority class (rarest class) in a classification dataset that exhibits severe class imbalance. If the dataset is not considered to have severe class imbalance (i.e. proportion of examples in the rarest class is not to small relative to the number of classes in the dataset), then all values will be `False`.
|
||||
|
||||
|
||||
``class_imbalance_score``
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
A numeric column with scores between 0 and 1.
|
||||
Any example belonging to the most under-represented class is assigned a score equal to the proportion of examples in the dataset belonging to that class.
|
||||
All other examples are assigned a score of 1.
|
||||
All examples sharing the same label also share the same score.
|
||||
|
||||
``given_label``
|
||||
~~~~~~~~~~~~~~~
|
||||
|
||||
A column of the actual labels as provided in the original dataset.
|
||||
|
||||
.. jinja ::
|
||||
|
||||
{% with issue_name = "class_imbalance" %}
|
||||
{% include "cleanlab/datalab/guide/_templates/issue_types_tip.rst" %}
|
||||
{% endwith %}
|
||||
|
||||
Image-specific Issues
|
||||
---------------------
|
||||
|
||||
Datalab can identify image-specific issues in datasets, such as images that are excessively dark or bright, blurry, lack detail, or have unusual sizes.
|
||||
To detect these issues, simply specify the `image_key` argument in :py:meth:`~cleanlab.datalab.datalab.Datalab`, indicating the image column name in your dataset.
|
||||
This functionality currently works only with Hugging Face datasets. You can convert other local dataset formats into a Hugging Face dataset by following `this guide <https://huggingface.co/docs/datasets/en/loading>`_.
|
||||
More information on these image-specific issues is available in the `CleanVision package <https://github.com/cleanlab/cleanvision?tab=readme-ov-file#clean-your-data-for-better-computer-vision>`_ .
|
||||
|
||||
Spurious Correlations between image-specific properties and labels
|
||||
------------------------------------------------------------------
|
||||
|
||||
Based on the :ref:`image properties discussed earlier <Image-specific Issues>`, Datalab can also look for spurious correlations between image properties and the labels in the dataset.
|
||||
These are unintended relationships between irrelevant features in images and the given labels, which ML models may easily exploit during training without learning the relevant features.
|
||||
Once deployed, such models would consistently fail to generalize on unseen data where these spurious correlations most likely don't hold.
|
||||
|
||||
Spurious correlations may arise in the dataset due to various reasons, such as:
|
||||
|
||||
- Images for certain classes might be consistently captured under specific environmental conditions.
|
||||
- Preprocessing techniques applied to the data might introduce systematic differences across classes.
|
||||
- Objects of different classes may be systematically photographed in particular ways.
|
||||
|
||||
Spurious Correlations are checked for when Datalab is initialized for an image dataset with the `image_key` keyword argument,
|
||||
after checking for :ref:`Image-specific Issues <Image-specific Issues>` where the image properties are computed.
|
||||
|
||||
Each image property (e.g. darkness/brightness) is assigned a label uncorrelatedness score for the entire dataset. The lower the score, the more strongly the property is correlated with the class labels, across images of the dataset. This score is mathematically defined as: 1 minus the relative accuracy improvement in predicting the labels based solely on this image property value (relative to always predicting the most common overall class).
|
||||
|
||||
Consider reviewing the relationship between images with high and low values of this property and the labels if the corresponding label uncorrelatedness score is low, because ML models trained on this dataset may latch onto the spurious correlation and fail to generalize.
|
||||
|
||||
This issue type is more about the overall dataset vs. individual data points and will only be highlighted by Datalab in its report, if any such troublesome image properties are found.
|
||||
|
||||
Metadata about spurious correlations is stored in the `info` attribute of the Datalab object.
|
||||
It can be accessed like so:
|
||||
|
||||
.. code::
|
||||
|
||||
lab.get_info("spurious_correlations")["correlations_df"]
|
||||
|
||||
|
||||
The output will look something like this:
|
||||
|
||||
.. testoutput::
|
||||
|
||||
property score
|
||||
0 blurry_score 0.559
|
||||
1 dark_score 0.808
|
||||
2 light_score 0.723
|
||||
3 odd_size_score 0.957
|
||||
4 odd_aspect_ratio_score 0.835
|
||||
5 grayscale_score 0.003 # Likely to be spuriously correlated with the labels
|
||||
6 low_information_score 0.688
|
||||
|
||||
|
||||
.. warning::
|
||||
|
||||
Note that the label uncorrelatedness scores are *not* stored in the `issues` attribute of Datalab.
|
||||
|
||||
``property``
|
||||
~~~~~~~~~~~~
|
||||
|
||||
A categorical column that identifies specific image-related characteristics assessed for potential spurious correlations with the class labels. Each entry in this column represents a distinct property of the images, such as blurriness, darkness, or grayscale, which may or may not be correlated with the labels.
|
||||
|
||||
``score``
|
||||
~~~~~~~~~
|
||||
|
||||
A numeric column that gives the level of label uncorrelatedness for a given image-specific property. The score lies between 0 and 1. The lower the score for an image-property, the more correlated the image-property is with the given labels.
|
||||
|
||||
.. tip::
|
||||
|
||||
This type of issue has the issue name `"spurious_correlations"`.
|
||||
|
||||
Run a check for this particular kind of issue by calling :py:meth:`Datalab.find_issues() <cleanlab.datalab.datalab.Datalab.find_issues>` like so:
|
||||
|
||||
.. code-block:: python
|
||||
|
||||
# `lab` is a Datalab instance
|
||||
lab.find_issues(..., issue_types = {"spurious_correlations": {}})
|
||||
|
||||
|
||||
|
||||
Underperforming Group Issue
|
||||
---------------------------
|
||||
|
||||
An underperforming group refers to a cluster of similar examples (i.e. a slice) in the dataset for which the ML model predictions are poor. The examples in this underperforming group may have noisy labels or feature values, or the trained ML model may not have learned how to properly handle them (consider collecting more data from this subpopulation or up-weighting the existing data from this group).
|
||||
|
||||
This issue-type is more about the overall dataset vs. individual data points. If an underperforming group is detected, Datalab will flag the individual data points from this group.
|
||||
|
||||
Underperforming Group issues are detected based on one of:
|
||||
|
||||
- provided `pred_probs` and `features`,
|
||||
- provided `pred_probs` and `knn_graph`, or
|
||||
- provided `pred_probs` and `cluster_ids`. (This option is for advanced users, see the `FAQ <../../../tutorials/faq.html#How-do-I-specify-pre-computed-data-slices/clusters-when-detecting-the-Underperforming-Group-Issue?>`_ for more details.)
|
||||
|
||||
If you do not provide both these arguments, this type of issue will not be considered.
|
||||
|
||||
To find the underperforming group, Cleanlab clusters the data using the provided `features` and determines the cluster `c` with the lowest average model predictive performance. Model predictive performance is evaluated via the model's self-confidence of the given labels, calculated using :py:func:`rank.get_self_confidence_for_each_label <cleanlab.rank.get_self_confidence_for_each_label>`. Suppose the average predictive power across the full dataset is `r` and is `q` within a cluster of examples. This cluster is considered to be an underperforming group if `q/r` falls below a threshold. A dataset suffers from the Underperforming Group issue if there exists such a cluster within it.
|
||||
The underperforming group quality score is equal to `q/r` for examples belonging to the underperforming group, and is equal to 1 for all other examples.
|
||||
Advanced users: If you have pre-computed cluster assignments for each example in the dataset, you can pass them explicitly to :py:meth:`Datalab.find_issues <cleanlab.datalab.datalab.Datalab.find_issues>` using the `cluster_ids` key in the `issue_types` dict argument. This is useful for tabular datasets where you want to group/slice the data based on a categorical column. An integer encoding of the categorical column can be passed as cluster assignments for finding the underperforming group, based on the data slices you define.
|
||||
|
||||
Some metadata about underperforming group issues is stored in the `issues` attribute of the Datalab object.
|
||||
Let's look at one way to access this information.
|
||||
|
||||
.. testcode::
|
||||
|
||||
lab.get_issues("underperforming_group").sort_values("underperforming_group_score").head(5)
|
||||
|
||||
The output will look something like this:
|
||||
|
||||
.. testoutput::
|
||||
|
||||
is_underperforming_group_issue underperforming_group_score
|
||||
0 False 1.0
|
||||
72 False 1.0
|
||||
71 False 1.0
|
||||
70 False 1.0
|
||||
69 False 1.0
|
||||
|
||||
``is_underperforming_group_issue``
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
A boolean column, where `True` indicates which examples belong to the subgroup (i.e. cluster/slice) for which model predictions are significantly worse than for the rest of the dataset.
|
||||
If there is no such underperforming subgroup detected, then all values will be `False`.
|
||||
|
||||
``underperforming_group_score``
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
A numeric column with scores between 0 and 1. Only examples belonging to a detected underperforming group receive a score less than 1.
|
||||
Every example in the underperforming group shares the same score, which is the ratio of group's label quality score vs. the mean label quality score across the dataset.
|
||||
The lower the score, the worse the model predictions are for this particular subgroup relative to the rest of the dataset.
|
||||
|
||||
.. jinja ::
|
||||
|
||||
{% with issue_name = "underperforming_group" %}
|
||||
{% include "cleanlab/datalab/guide/_templates/issue_types_tip.rst" %}
|
||||
{% endwith %}
|
||||
|
||||
Null Issue
|
||||
----------
|
||||
|
||||
Examples identified with the null issue correspond to rows that have null/missing values across all feature columns (i.e. the entire row is missing values).
|
||||
|
||||
Null issues are detected based on provided `features`. If you do not provide `features`, this type of issue will not be considered.
|
||||
|
||||
Each example's null issue quality score equals the proportion of features values in this row that are not null/missing. The overall dataset null issue quality score
|
||||
equals the average of the individual examples' quality scores.
|
||||
|
||||
Presence of null examples in the dataset can lead to errors when training ML models. It can also
|
||||
result in the model learning incorrect patterns due to the null values.
|
||||
|
||||
Some metadata about null issues is stored in the `issues` attribute of the Datalab object.
|
||||
Let's look at one way to access this information.
|
||||
|
||||
.. testcode::
|
||||
|
||||
lab.get_issues("null").sort_values("null_score").head(5)
|
||||
|
||||
The output will look something like this:
|
||||
|
||||
.. testoutput::
|
||||
|
||||
is_null_issue null_score
|
||||
0 False 1.0
|
||||
72 False 1.0
|
||||
71 False 1.0
|
||||
70 False 1.0
|
||||
69 False 1.0
|
||||
|
||||
``is_null_issue``
|
||||
~~~~~~~~~~~~~~~~~
|
||||
|
||||
A boolean column, where `True` indicates that an example is identified as having null/missing values across all feature columns.
|
||||
Examples that just have a single non-null value across multiple feature columns are not flagged with this issue.
|
||||
|
||||
``null_score``
|
||||
~~~~~~~~~~~~~~
|
||||
|
||||
A numeric column with scores between 0 and 1. The score represents the proportion of non-null (i.e. non-missing) values in each example.
|
||||
Lower scores indicate examples with more null/missing values.
|
||||
|
||||
.. jinja ::
|
||||
|
||||
{% with issue_name = "null"%}
|
||||
{% include "cleanlab/datalab/guide/_templates/issue_types_tip.rst" %}
|
||||
{% endwith %}
|
||||
|
||||
Data Valuation Issue
|
||||
--------------------
|
||||
|
||||
The examples in the dataset with lowest data valuation scores contribute least to a trained ML model's performance (those whose value falls below a threshold are flagged with this type of issue).
|
||||
|
||||
Data valuation issues can be detected based on provided `features` or a provided `knn_graph` (or one pre-computed during the computation of other issue types). If you do not provide one of these two arguments and there isn't a `knn_graph` already stored in the Datalab object, this type of issue will not be considered.
|
||||
|
||||
The data valuation score is an approximate Data Shapley value, calculated based on the labels of the top k nearest neighbors of an example. The details of this KNN-Shapley value could be found in the papers: `Efficient Task-Specific Data Valuation for Nearest Neighbor Algorithms <https://arxiv.org/abs/1908.08619>`_ and `Scalability vs. Utility: Do We Have to Sacrifice One for the Other in Data Importance Quantification? <https://arxiv.org/abs/1911.07128>`_.
|
||||
|
||||
Some metadata about data valuation issues is stored in the `issues` attribute of the Datalab object.
|
||||
Let's look at one way to access this information.
|
||||
|
||||
.. testcode::
|
||||
|
||||
lab.get_issues("data_valuation").sort_values("data_valuation_score").head(5)
|
||||
|
||||
The output will look something like this:
|
||||
|
||||
.. testoutput::
|
||||
|
||||
is_data_valuation_issue data_valuation_score
|
||||
39 False 0.5
|
||||
32 False 0.5
|
||||
98 False 0.5
|
||||
6 False 0.5
|
||||
7 False 0.5
|
||||
|
||||
``is_data_valuation_issue``
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
A boolean column, where `True` indicates that an example does not appear to contribute positively to a model's training performance.
|
||||
|
||||
``data_valuation_score``
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
A numeric column with scores between 0 and 1. The score reflects how valuable each individual example is in terms of improving the performance of the ML model trained on this dataset.
|
||||
Examples with higher scores more positively influence the resulting model's predictive performance, contributing to better learning. One would expect the model to get worse if many such examples were removed from its training dataset.
|
||||
|
||||
.. jinja ::
|
||||
|
||||
{% with issue_name = "data_valuation"%}
|
||||
{% include "cleanlab/datalab/guide/_templates/issue_types_tip.rst" %}
|
||||
{% endwith %}
|
||||
|
||||
Identifier Column Issue
|
||||
-----------------------
|
||||
This issue type flags sequential numerical columns in the features of a dataset. A numerical sequential column is most likely an identifier column, which - in most cases - should not be used in training ML models.
|
||||
More formally, an identifier column is defined as a column i in features such that set(features[:,i]) = set(c, c+1, ..., c+n) for some integer c, where n = num_rows of features. Note we don't consider the ordering of the column in case the dataset was pre-shuffled.
|
||||
|
||||
If the condition is met in one or more columns, then we say this dataset has the identifier_column issue. The overall issue-summary-score is binary. The score equals 1 if the issue is present and 0 if the issue is not present.
|
||||
|
||||
The info attribute of Datalab shows the indices of the columns with the column_identifier issue.
|
||||
|
||||
|
||||
|
||||
.. jinja ::
|
||||
|
||||
{% with issue_name = "identifier_column"%}
|
||||
{% include "cleanlab/datalab/guide/_templates/issue_types_tip.rst" %}
|
||||
{% endwith %}
|
||||
|
||||
Optional Issue Parameters
|
||||
=========================
|
||||
|
||||
Here is the dict of possible (**optional**) parameter values that can be specified via the argument `issue_types` to :py:meth:`Datalab.find_issues <cleanlab.datalab.datalab.Datalab.find_issues>`.
|
||||
Optionally specify these to exert greater control over how issues are detected in your dataset.
|
||||
Appropriate defaults are used for any parameters you do not specify, so no need to specify all of these!
|
||||
|
||||
.. code-block:: python
|
||||
|
||||
possible_issue_types = {
|
||||
"label": label_kwargs, "outlier": outlier_kwargs,
|
||||
"near_duplicate": near_duplicate_kwargs, "non_iid": non_iid_kwargs,
|
||||
"class_imbalance": class_imbalance_kwargs, "underperforming_group": underperforming_group_kwargs,
|
||||
"null": null_kwargs, "data_valuation": data_valuation_kwargs,
|
||||
"identifier_column": identifier_column_kwargs,
|
||||
}
|
||||
|
||||
|
||||
where the possible `kwargs` dicts for each key are described in the sections below.
|
||||
|
||||
Label Issue Parameters
|
||||
----------------------
|
||||
|
||||
.. code-block:: python
|
||||
|
||||
label_kwargs = {
|
||||
"k": # number of nearest neighbors to consider when computing pred_probs from features,
|
||||
"health_summary_parameters": # dict of potential keyword arguments to method `dataset.health_summary()`,
|
||||
"clean_learning_kwargs": # dict of keyword arguments to constructor `CleanLearning()` including keys like: "find_label_issues_kwargs" or "label_quality_scores_kwargs",
|
||||
"thresholds": # `thresholds` argument to `CleanLearning.find_label_issues()`,
|
||||
"noise_matrix": # `noise_matrix` argument to `CleanLearning.find_label_issues()`,
|
||||
"inverse_noise_matrix": # `inverse_noise_matrix` argument to `CleanLearning.find_label_issues()`,
|
||||
"save_space": # `save_space` argument to `CleanLearning.find_label_issues()`,
|
||||
"clf_kwargs": # `clf_kwargs` argument to `CleanLearning.find_label_issues()`. Currently has no effect.,
|
||||
"validation_func": # `validation_func` argument to `CleanLearning.fit()`. Currently has no effect.,
|
||||
}
|
||||
|
||||
.. attention::
|
||||
|
||||
``health_summary_parameters`` and ``health_summary_kwargs`` can work in tandem to determine the arguments to be used in the call to :py:meth:`dataset.health_summary <cleanlab.dataset.health_summary>`.
|
||||
|
||||
.. note::
|
||||
|
||||
For more information, view the source code of: :py:class:`datalab.internal.issue_manager.label.LabelIssueManager <cleanlab.datalab.internal.issue_manager.label.LabelIssueManager>`.
|
||||
|
||||
Outlier Issue Parameters
|
||||
------------------------
|
||||
|
||||
.. code-block:: python
|
||||
|
||||
outlier_kwargs = {
|
||||
"threshold": # floating value between 0 and 1 that sets the sensitivity of the outlier detection algorithms, based on either features or pred_probs.
|
||||
"k": # integer representing the number of nearest neighbors for nearest neighbors search (passed as argument to `NearestNeighbors`), if necessary, Used with features,
|
||||
"t": # integer used to modulate the strength of the transformation from distances to scores that lie in the range [0, 1]. Used with features,
|
||||
"scaling_factor": # floating value used to normalize the distances before they are converted into scores. Used with features,
|
||||
"metric": # string or callable representing the distance metric used in nearest neighbors search (passed as argument to `NearestNeighbors`), if necessary, Used with features,
|
||||
"ood_kwargs": # dict of keyword arguments to constructor `OutOfDistribution()`{
|
||||
"params": {
|
||||
# NOTE: Each of the following keyword arguments can also be provided outside "ood_kwargs"
|
||||
"adjust_pred_probs": # `adjust_pred_probs` argument to constructor `OutOfDistribution()`. Used with pred_probs,
|
||||
"method": # `method` argument to constructor `OutOfDistribution()`. Used with pred_probs,
|
||||
"confident_thresholds": # `confident_thresholds` argument to constructor `OutOfDistribution()`. Used with pred_probs,
|
||||
},
|
||||
},
|
||||
}
|
||||
|
||||
.. note::
|
||||
|
||||
For more information, view the source code of: :py:class:`datalab.internal.issue_manager.outlier.OutlierIssueManager <cleanlab.datalab.internal.issue_manager.outlier.OutlierIssueManager>`.
|
||||
|
||||
Duplicate Issue Parameters
|
||||
--------------------------
|
||||
|
||||
.. code-block:: python
|
||||
|
||||
near_duplicate_kwargs = {
|
||||
"metric": # string or callable representing the distance metric used in nearest neighbors search (passed as argument to `NearestNeighbors`), if necessary,
|
||||
"k": # integer representing the number of nearest neighbors for nearest neighbors search (passed as argument to `NearestNeighbors`), if necessary,
|
||||
"threshold": # `threshold` argument to constructor of `NearDuplicateIssueManager()`. Non-negative floating value that determines the maximum distance between two examples to be considered outliers, relative to the median distance to the nearest neighbors,
|
||||
}
|
||||
|
||||
.. attention::
|
||||
|
||||
`k` does not affect the results of the (near) duplicate search algorithm. It only affects the construction of the knn graph, if necessary.
|
||||
|
||||
.. note::
|
||||
|
||||
For more information, view the source code of: :py:class:`datalab.internal.issue_manager.duplicate.NearDuplicateIssueManager <cleanlab.datalab.internal.issue_manager.duplicate.NearDuplicateIssueManager>`.
|
||||
|
||||
|
||||
Non-IID Issue Parameters
|
||||
------------------------
|
||||
|
||||
.. code-block:: python
|
||||
|
||||
non_iid_kwargs = {
|
||||
"metric": # `metric` argument to constructor of `NonIIDIssueManager`. String or callable for the distance metric used for nearest neighbors search if necessary. `metric` argument to constructor of `sklearn.neighbors.NearestNeighbors`,
|
||||
"k": # `k` argument to constructor of `NonIIDIssueManager`. Integer representing the number of nearest neighbors for nearest neighbors search if necessary. `n_neighbors` argument to constructor of `sklearn.neighbors.NearestNeighbors`,
|
||||
"num_permutations": # `num_permutations` argument to constructor of `NonIIDIssueManager`,
|
||||
"seed": # seed for numpy's random number generator (used for permutation tests),
|
||||
"significance_threshold": # `significance_threshold` argument to constructor of `NonIIDIssueManager`. Floating value between 0 and 1 that determines the overall signicance of non-IID issues found in the dataset.
|
||||
}
|
||||
|
||||
.. note::
|
||||
|
||||
For more information, view the source code of: :py:class:`datalab.internal.issue_manager.noniid.NonIIDIssueManager <cleanlab.datalab.internal.issue_manager.noniid.NonIIDIssueManager>`.
|
||||
|
||||
|
||||
Imbalance Issue Parameters
|
||||
--------------------------
|
||||
|
||||
.. code-block:: python
|
||||
|
||||
class_imbalance_kwargs = {
|
||||
"threshold": # `threshold` argument to constructor of `ClassImbalanceIssueManager`. Non-negative floating value between 0 and 1 indicating the minimum fraction of samples of each class that are present in a dataset without class imbalance.
|
||||
}
|
||||
|
||||
.. note::
|
||||
|
||||
For more information, view the source code of: :py:class:`datalab.internal.issue_manager.imbalance.ClassImbalanceIssueManager <cleanlab.datalab.internal.issue_manager.imbalance.ClassImbalanceIssueManager>`.
|
||||
|
||||
Underperforming Group Issue Parameters
|
||||
--------------------------------------
|
||||
|
||||
.. code-block:: python
|
||||
|
||||
underperforming_group_kwargs = {
|
||||
# Constructor arguments for `UnderperformingGroupIssueManager`
|
||||
"threshold": # Non-negative floating value between 0 and 1 used for determinining group of points with low confidence.
|
||||
"metric": # String or callable for the distance metric used for nearest neighbors search if necessary. `metric` argument to constructor of `sklearn.neighbors.NearestNeighbors`.
|
||||
"k": # Integer representing the number of nearest neighbors for constructing the nearest neighbour graph. `n_neighbors` argument to constructor of `sklearn.neighbors.NearestNeighbors`.
|
||||
"min_cluster_samples": # Non-negative integer value specifying the minimum number of examples required for a cluster to be considered as the underperforming group. Used in `UnderperformingGroupIssueManager.filter_cluster_ids`.
|
||||
"clustering_kwargs": # Key-value pairs representing arguments for the constructor of the clustering algorithm class (e.g. `sklearn.cluster.DBSCAN`).
|
||||
|
||||
# Argument for the find_issues() method of UnderperformingGroupIssueManager
|
||||
"cluster_ids": # A 1-D numpy array containing cluster labels for each sample in the dataset. If passed, these cluster labels are used for determining the underperforming group.
|
||||
}
|
||||
|
||||
.. note::
|
||||
|
||||
For more information, view the source code of: :py:class:`datalab.internal.issue_manager.underperforming_group.UnderperformingGroupIssueManager <cleanlab.datalab.internal.issue_manager.underperforming_group.UnderperformingGroupIssueManager>`.
|
||||
|
||||
For more information on generating `cluster_ids` for this issue manager, refer to this `FAQ Section <../../../tutorials/faq.html#How-do-I-specify-pre-computed-data-slices/clusters-when-detecting-the-Underperforming-Group-Issue?>`_.
|
||||
|
||||
Null Issue Parameters
|
||||
---------------------
|
||||
|
||||
.. code-block:: python
|
||||
|
||||
null_kwargs = {}
|
||||
|
||||
.. note::
|
||||
|
||||
For more information, view the source code of: :py:class:`datalab.internal.issue_manager.null.NullIssueManager <cleanlab.datalab.internal.issue_manager.null.NullIssueManager>`.
|
||||
|
||||
Data Valuation Issue Parameters
|
||||
-------------------------------
|
||||
|
||||
.. code-block:: python
|
||||
|
||||
data_valuation_kwargs = {
|
||||
"k": # Number of nearest neighbors used to calculate data valuation scores,
|
||||
"threshold": # Examples with scores below this threshold will be flagged with a data valuation issue
|
||||
}
|
||||
|
||||
.. note::
|
||||
For more information, view the source code of: :py:class:`datalab.internal.issue_manager.data_valuation.DataValuationIssueManager <cleanlab.datalab.internal.issue_manager.data_valuation.DataValuationIssueManager>`.
|
||||
|
||||
Identifier Column Parameters
|
||||
----------------------------
|
||||
|
||||
.. code-block:: python
|
||||
|
||||
identifier_column_kwargs = {}
|
||||
|
||||
.. note::
|
||||
|
||||
For more information, view the source code of: :py:class:`datalab.internal.issue_manager.identifier_column.IdentifierColumnIssueManager <cleanlab.datalab.internal.issue_manager.identifier_column.IdentifierColumnIssueManager>`.
|
||||
|
||||
Image Issue Parameters
|
||||
----------------------
|
||||
|
||||
To customize optional parameters for specific image issue types, you can provide a dictionary format corresponding to each image issue. The following codeblock demonstrates how to specify optional parameters for all image issues. However, it's important to note that providing optional parameters for specific image issues is not mandatory. If no specific parameters are provided, defaults will be used for those issues.
|
||||
|
||||
.. code-block:: python
|
||||
|
||||
image_issue_types_kwargs = {
|
||||
"dark": {"threshold": 0.32}, # `threshold` argument for dark issue type. Non-negative floating value between 0 and 1, lower value implies fewer samples will be marked as issue and vice versa.
|
||||
"light": {"threshold": 0.05}, # `threshold` argument for light issue type. Non-negative floating value between 0 and 1, lower value implies fewer samples will be marked as issue and vice versa.
|
||||
"blurry": {"threshold": 0.29}, # `threshold` argument for blurry issue type. Non-negative floating value between 0 and 1, lower value implies fewer samples will be marked as issue and vice versa.
|
||||
"low_information": {"threshold": 0.3}, # `threshold` argument for low_information issue type. Non-negative floating value between 0 and 1, lower value implies fewer samples will be marked as issue and vice versa.
|
||||
"odd_aspect_ratio": {"threshold": 0.35}, # `threshold` argument for odd_aspect_ratio issue type. Non-negative floating value between 0 and 1, lower value implies fewer samples will be marked as issue and vice versa.
|
||||
"odd_size": {"threshold": 10.0}, # `threshold` argument for odd_size issue type. Non-negative integer value between starting from 0, unlike other issues, here higher value implies fewer samples will be selected.
|
||||
}
|
||||
|
||||
.. note::
|
||||
|
||||
For more information, view the cleanvision `docs <https://cleanvision.readthedocs.io/en/latest/tutorials/tutorial.html#5.-Check-for-an-issue-with-a-different-threshold>`_.
|
||||
|
||||
|
||||
Spurious Correlations Issue Parameters
|
||||
--------------------------------------
|
||||
|
||||
.. code-block:: python
|
||||
|
||||
spurious_correlations_kwargs = {
|
||||
"threshold": 0.3, # Non-negative floating value between 0 and 1, lower value implies fewer image properties may have a low enough label uncorrelatedness score to be marked as issue and vice versa.
|
||||
}
|
||||
|
||||
Cleanlab Studio (Easy Mode)
|
||||
---------------------------
|
||||
|
||||
`Cleanlab Studio <https://cleanlab.ai/blog/data-centric-ai/>`_ is a fully automated platform that can detect the same data issues as this package, as well as `many more types of issues <https://help.cleanlab.ai/guide/concepts/cleanlab_columns/>`_, all without you having to do any Machine Learning (or even write any code). Beyond being 100x faster to use and producing more useful results, `Cleanlab Studio <https://cleanlab.ai/blog/data-centric-ai/>`_ also provides an intelligent data correction interface for you to quickly fix the issues detected in your dataset (a single data scientist can fix millions of data points thanks to AI suggestions).
|
||||
|
||||
`Cleanlab Studio <https://cleanlab.ai/blog/data-centric-ai/>`_ offers a powerful AutoML system (with Foundation models) that is useful for more than improving data quality. With a few clicks, you can: find + fix issues in your dataset, identify the best type of ML model and train/tune it, and deploy this model to serve accurate predictions for new data. Also use the same AutoML to auto-label large datasets (a single user can label millions of data points thanks to powerful Foundation models). `Try Cleanlab Studio for free! <https://cleanlab.ai/signup/>`_
|
||||
|
||||
.. image:: https://raw.githubusercontent.com/cleanlab/assets/master/cleanlab/ml-with-cleanlab-studio.png
|
||||
:width: 800
|
||||
:alt: Stages of modern AI pipeline that can now be automated with Cleanlab Studio
|
||||
@@ -0,0 +1,186 @@
|
||||
.. tabs::
|
||||
|
||||
.. tab:: Classification task
|
||||
|
||||
.. list-table::
|
||||
:widths: 20 10 20 50
|
||||
:header-rows: 1
|
||||
|
||||
* - Issue Name
|
||||
- Default
|
||||
- Column Name
|
||||
- Required keyword arguments in :py:meth:`Datalab.find_issues <cleanlab.datalab.datalab.Datalab.find_issues>`
|
||||
* - :ref:`label<Label Issue>`
|
||||
- Yes
|
||||
-
|
||||
:ref:`is_label_issue<\`\`is_label_issue\`\`>`
|
||||
|
||||
:ref:`label_score<\`\`label_score\`\`>`
|
||||
|
||||
:ref:`given_label<\`\`given_label\`\`>`
|
||||
|
||||
:ref:`predicted_label<\`\`predicted_label\`\`>`
|
||||
- `pred_probs` or `features`
|
||||
* - :ref:`outlier<Outlier Issues>`
|
||||
- Yes
|
||||
-
|
||||
:ref:`is_outlier_issue<\`\`is_outlier_issue\`\`>`
|
||||
|
||||
:ref:`outlier_score<\`\`outlier_score\`\`>`
|
||||
- `pred_probs` or `features` or `knn_graph`
|
||||
* - :ref:`near_duplicate<(Near) Duplicate Issue>`
|
||||
- Yes
|
||||
-
|
||||
:ref:`is_near_duplicate_issue<\`\`is_near_duplicate_issue\`\`>`
|
||||
|
||||
:ref:`near_duplicate_score<\`\`near_duplicate_score\`\`>`
|
||||
|
||||
:ref:`near_duplicate_sets<\`\`near_duplicate_sets\`\`>`
|
||||
|
||||
:ref:`distance_to_nearest_neighbor<\`\`distance_to_nearest_neighbor\`\`>`
|
||||
- `features` or `knn_graph`
|
||||
* - :ref:`non_iid<Non-IID Issue>`
|
||||
- Yes
|
||||
-
|
||||
:ref:`is_non_iid_issue<\`\`is_non_iid_issue\`\`>`
|
||||
|
||||
:ref:`non_iid_score<\`\`non_iid_score\`\`>`
|
||||
- `pred_probs` or `features` or `knn_graph`
|
||||
* - :ref:`class_imbalance<Class Imbalance Issue>`
|
||||
- Yes
|
||||
-
|
||||
:ref:`is_class_imbalance_issue<\`\`is_class_imbalance_issue\`\`>`
|
||||
|
||||
:ref:`class_imbalance_score<\`\`class_imbalance_score\`\`>`
|
||||
- None [#f1]_
|
||||
* - :ref:`underperforming_group<Underperforming Group Issue>`
|
||||
- Yes
|
||||
-
|
||||
:ref:`is_underperforming_group_issue<\`\`is_underperforming_group_issue\`\`>`
|
||||
|
||||
:ref:`underperforming_group_score<\`\`underperforming_group_score\`\`>`
|
||||
- (`pred_probs`, `features`) or (`pred_probs`, `knn_graph`) or (`pred_probs`, `cluster_ids`) [#f2]_
|
||||
* - :ref:`null<Null Issue>`
|
||||
- Yes
|
||||
-
|
||||
:ref:`is_null_issue<\`\`is_null_issue\`\`>`
|
||||
|
||||
:ref:`null_score<\`\`null_score\`\`>`
|
||||
- `features`
|
||||
* - :ref:`data_valuation<Data Valuation Issue>`
|
||||
- No
|
||||
-
|
||||
:ref:`is_data_valuation_issue<\`\`is_data_valuation_issue\`\`>`
|
||||
|
||||
:ref:`data_valuation_score<\`\`data_valuation_score\`\`>`
|
||||
- `knn_graph`
|
||||
|
||||
.. rubric:: Notes
|
||||
|
||||
.. [#f1] Only runs if `label_name` is provided in `Datalab()` constructor
|
||||
.. [#f2] `cluster_ids` currently needs to be passed via `issue_types`
|
||||
|
||||
.. tab:: Regression task
|
||||
|
||||
.. list-table::
|
||||
:widths: 20 10 20 50
|
||||
:header-rows: 1
|
||||
|
||||
* - Issue Name
|
||||
- Default
|
||||
- Column Name
|
||||
- Required keyword arguments in :py:meth:`Datalab.find_issues <cleanlab.datalab.datalab.Datalab.find_issues>`
|
||||
* - :ref:`label<Label Issue>`
|
||||
- Yes
|
||||
-
|
||||
:ref:`is_label_issue<\`\`is_label_issue\`\`>`
|
||||
|
||||
:ref:`label_score<\`\`label_score\`\`>`
|
||||
|
||||
:ref:`given_label<\`\`given_label\`\`>`
|
||||
|
||||
:ref:`predicted_label<\`\`predicted_label\`\`>`
|
||||
- `pred_probs` [#f3]_ or `features` or (`features`, `model`) [#f4]_
|
||||
* - :ref:`outlier<Outlier Issues>`
|
||||
- Yes
|
||||
-
|
||||
:ref:`is_outlier_issue<\`\`is_outlier_issue\`\`>`
|
||||
|
||||
:ref:`outlier_score<\`\`outlier_score\`\`>`
|
||||
- `features` or `knn_graph`
|
||||
* - :ref:`near_duplicate<(Near) Duplicate Issue>`
|
||||
- Yes
|
||||
-
|
||||
:ref:`is_near_duplicate_issue<\`\`is_near_duplicate_issue\`\`>`
|
||||
|
||||
:ref:`near_duplicate_score<\`\`near_duplicate_score\`\`>`
|
||||
- `features` or `knn_graph`
|
||||
* - :ref:`non_iid<Non-IID Issue>`
|
||||
- Yes
|
||||
-
|
||||
:ref:`is_non_iid_issue<\`\`is_non_iid_issue\`\`>`
|
||||
|
||||
:ref:`non_iid_score<\`\`non_iid_score\`\`>`
|
||||
- `features` or `knn_graph`
|
||||
* - :ref:`null<Null Issue>`
|
||||
- Yes
|
||||
-
|
||||
:ref:`is_null_issue<\`\`is_null_issue\`\`>`
|
||||
|
||||
:ref:`null_score<\`\`null_score\`\`>`
|
||||
- `features`
|
||||
|
||||
.. rubric:: Notes
|
||||
|
||||
.. [#f3] :abbr:`pred_probs (Predicted Probabilities)` gets reinterpreted as a `predictions` argument internally
|
||||
.. [#f4] `model` currently needs to be passed as `issue_types = {"label": {"clean_learning_kwargs": {"model": your_regression_model}}}`
|
||||
|
||||
.. tab:: Multilabel task
|
||||
|
||||
.. list-table::
|
||||
:widths: 20 10 20 50
|
||||
:header-rows: 1
|
||||
|
||||
* - Issue Name
|
||||
- Default
|
||||
- Column Name
|
||||
- Required keyword arguments in :py:meth:`Datalab.find_issues <cleanlab.datalab.datalab.Datalab.find_issues>`
|
||||
* - :ref:`label<Label Issue>`
|
||||
- Yes
|
||||
-
|
||||
:ref:`is_label_issue<\`\`is_label_issue\`\`>`
|
||||
|
||||
:ref:`label_score<\`\`label_score\`\`>`
|
||||
|
||||
:ref:`given_label<\`\`given_label\`\`>`
|
||||
|
||||
:ref:`predicted_label<\`\`predicted_label\`\`>`
|
||||
- `pred_probs` or `features`
|
||||
* - :ref:`outlier<Outlier Issues>`
|
||||
- Yes
|
||||
-
|
||||
:ref:`is_outlier_issue<\`\`is_outlier_issue\`\`>`
|
||||
|
||||
:ref:`outlier_score<\`\`outlier_score\`\`>`
|
||||
- `features` or `knn_graph`
|
||||
* - :ref:`near_duplicate<(Near) Duplicate Issue>`
|
||||
- Yes
|
||||
-
|
||||
:ref:`is_near_duplicate_issue<\`\`is_near_duplicate_issue\`\`>`
|
||||
|
||||
:ref:`near_duplicate_score<\`\`near_duplicate_score\`\`>`
|
||||
- `features` or `knn_graph`
|
||||
* - :ref:`non_iid<Non-IID Issue>`
|
||||
- Yes
|
||||
-
|
||||
:ref:`is_non_iid_issue<\`\`is_non_iid_issue\`\`>`
|
||||
|
||||
:ref:`non_iid_score<\`\`non_iid_score\`\`>`
|
||||
- `features` or `knn_graph`
|
||||
* - :ref:`null<Null Issue>`
|
||||
- Yes
|
||||
-
|
||||
:ref:`is_null_issue<\`\`is_null_issue\`\`>`
|
||||
|
||||
:ref:`null_score<\`\`null_score\`\`>`
|
||||
- `features`
|
||||
@@ -0,0 +1,31 @@
|
||||
datalab
|
||||
=======
|
||||
|
||||
.. automodule:: cleanlab.datalab
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
|
||||
Getting Started
|
||||
---------------
|
||||
|
||||
.. include:: optional_dependencies.rst
|
||||
|
||||
Guides
|
||||
------
|
||||
|
||||
.. toctree::
|
||||
:maxdepth: 2
|
||||
|
||||
guide/index
|
||||
|
||||
|
||||
API Reference
|
||||
-------------
|
||||
|
||||
.. toctree::
|
||||
:maxdepth: 2
|
||||
|
||||
datalab
|
||||
internal/index
|
||||
@@ -0,0 +1,9 @@
|
||||
imagelab
|
||||
=============
|
||||
|
||||
|
||||
.. automodule:: cleanlab.datalab.internal.adapter.imagelab
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,10 @@
|
||||
adapter
|
||||
=============
|
||||
|
||||
.. warning::
|
||||
Methods in this ``adapter`` module are intended for internal use within the ``cleanlab`` package. They are not guaranteed to be stable between different versions.
|
||||
|
||||
.. toctree::
|
||||
:maxdepth: 2
|
||||
|
||||
imagelab
|
||||
@@ -0,0 +1,9 @@
|
||||
data
|
||||
====
|
||||
|
||||
.. automodule:: cleanlab.datalab.internal.data
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
:ignore-module-all:
|
||||
@@ -0,0 +1,9 @@
|
||||
data_issues
|
||||
===========
|
||||
|
||||
.. automodule:: cleanlab.datalab.internal.data_issues
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
:ignore-module-all:
|
||||
@@ -0,0 +1,9 @@
|
||||
factory
|
||||
=======
|
||||
|
||||
.. automodule:: cleanlab.datalab.internal.issue_manager_factory
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
:ignore-module-all:
|
||||
@@ -0,0 +1,26 @@
|
||||
internal
|
||||
========
|
||||
|
||||
.. warning::
|
||||
Methods in this ``internal`` module are intended for internal use within the ``cleanlab`` package. They are not guaranteed to be stable between different versions.
|
||||
|
||||
.. automodule:: cleanlab.datalab.internal
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
|
||||
|
||||
|
||||
.. toctree::
|
||||
:maxdepth: 2
|
||||
|
||||
adapter/index
|
||||
data
|
||||
data_issues
|
||||
issue_finder
|
||||
factory
|
||||
model_outputs
|
||||
issue_manager/index
|
||||
report
|
||||
task
|
||||
@@ -0,0 +1,12 @@
|
||||
issue_finder
|
||||
============
|
||||
|
||||
.. note:: This module is not intended to be used directly by users. It is used by the :mod:`cleanlab.datalab.datalab` module.
|
||||
Specifically, it is used by the :py:meth:`Datalab.find_issues <cleanlab.datalab.datalab.Datalab.find_issues>` method.
|
||||
|
||||
.. automodule:: cleanlab.datalab.internal.issue_finder
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
:ignore-module-all:
|
||||
@@ -0,0 +1,5 @@
|
||||
.. warning::
|
||||
|
||||
This issue manager isn't set up for direct Datalab use yet.
|
||||
|
||||
Register it first using `~cleanlab.datalab.internal.issue_manager_factory.register`.
|
||||
@@ -0,0 +1,9 @@
|
||||
data_valuation
|
||||
===========
|
||||
|
||||
|
||||
.. automodule:: cleanlab.datalab.internal.issue_manager.data_valuation
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,9 @@
|
||||
duplicate
|
||||
=========
|
||||
|
||||
|
||||
.. automodule:: cleanlab.datalab.internal.issue_manager.duplicate
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,9 @@
|
||||
imbalance
|
||||
=========
|
||||
|
||||
|
||||
.. automodule:: cleanlab.datalab.internal.issue_manager.imbalance
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,29 @@
|
||||
issue_manager
|
||||
=============
|
||||
|
||||
.. warning::
|
||||
Methods in this ``issue_manager`` module are bleeding edge and may have sharp edges. They are not guaranteed to be stable between different ``cleanlab`` versions.
|
||||
|
||||
|
||||
Registered issue managers
|
||||
-------------------------
|
||||
|
||||
These are the issue managers that Datalab has registered.
|
||||
|
||||
.. toctree::
|
||||
Base issue_manager module <issue_manager>
|
||||
label
|
||||
outlier
|
||||
duplicate
|
||||
noniid
|
||||
imbalance
|
||||
underperforming_group
|
||||
null
|
||||
data_valuation
|
||||
|
||||
ML task-specific issue managers
|
||||
---------------------------------
|
||||
|
||||
.. toctree::
|
||||
regression/index
|
||||
multilabel/index
|
||||
@@ -0,0 +1,8 @@
|
||||
issue_manager
|
||||
=============
|
||||
|
||||
.. automodule:: cleanlab.datalab.internal.issue_manager.issue_manager
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,8 @@
|
||||
label
|
||||
=====
|
||||
|
||||
.. automodule:: cleanlab.datalab.internal.issue_manager.label
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,8 @@
|
||||
multilabel
|
||||
==========
|
||||
|
||||
.. warning::
|
||||
Methods in this ``multilabel`` module are bleeding edge and may have sharp edges. They are not guaranteed to be stable between different ``cleanlab`` versions.
|
||||
|
||||
.. toctree::
|
||||
label
|
||||
@@ -0,0 +1,8 @@
|
||||
label
|
||||
=====
|
||||
|
||||
.. automodule:: cleanlab.datalab.internal.issue_manager.multilabel.label
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,9 @@
|
||||
noniid
|
||||
=======
|
||||
|
||||
|
||||
.. automodule:: cleanlab.datalab.internal.issue_manager.noniid
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,10 @@
|
||||
null
|
||||
====
|
||||
|
||||
.. include:: _notices/not_registered.rst
|
||||
|
||||
.. automodule:: cleanlab.datalab.internal.issue_manager.null
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,9 @@
|
||||
outlier
|
||||
=======
|
||||
|
||||
|
||||
.. automodule:: cleanlab.datalab.internal.issue_manager.outlier
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,8 @@
|
||||
regression
|
||||
==========
|
||||
|
||||
.. warning::
|
||||
Methods in this ``regression`` module are bleeding edge and may have sharp edges. They are not guaranteed to be stable between different ``cleanlab`` versions.
|
||||
|
||||
.. toctree::
|
||||
label
|
||||
@@ -0,0 +1,8 @@
|
||||
label
|
||||
=====
|
||||
|
||||
.. automodule:: cleanlab.datalab.internal.issue_manager.regression.label
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,9 @@
|
||||
underperforming_group
|
||||
=========
|
||||
|
||||
|
||||
.. automodule:: cleanlab.datalab.internal.issue_manager.underperforming_group
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,7 @@
|
||||
model_outputs
|
||||
=============
|
||||
|
||||
.. automodule:: cleanlab.datalab.internal.model_outputs
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,12 @@
|
||||
report
|
||||
======
|
||||
|
||||
.. note:: This module is not intended to be used directly by users. It is used by the :mod:`cleanlab.datalab.datalab` module.
|
||||
Specifically, it is used by the :py:meth:`Datalab.report <cleanlab.datalab.datalab.Datalab.report>` method.
|
||||
|
||||
.. automodule:: cleanlab.datalab.internal.report
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
:ignore-module-all:
|
||||
@@ -0,0 +1,11 @@
|
||||
task
|
||||
====
|
||||
|
||||
.. note:: This module is not intended to be used directly by users. It is used by the :mod:`cleanlab.datalab.datalab` module.
|
||||
|
||||
.. automodule:: cleanlab.datalab.internal.task
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
:ignore-module-all:
|
||||
@@ -0,0 +1,11 @@
|
||||
Using Datalab requires additional dependencies beyond the rest of the ``cleanlab`` package. To install them, run:
|
||||
|
||||
.. code-block:: console
|
||||
|
||||
$ pip install "cleanlab[datalab]"
|
||||
|
||||
For the developmental version of the package, install from source:
|
||||
|
||||
.. code-block:: console
|
||||
|
||||
$ pip install "git+https://github.com/cleanlab/cleanlab.git#egg=cleanlab[datalab]"
|
||||
@@ -0,0 +1,72 @@
|
||||
dataset
|
||||
=======
|
||||
|
||||
.. automodule:: cleanlab.dataset
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
|
||||
.. testsetup:: *
|
||||
|
||||
import cleanlab
|
||||
import numpy as np
|
||||
from cleanlab.benchmarking import noise_generation
|
||||
|
||||
SEED = 0
|
||||
|
||||
def get_data_labels_from_dataset(
|
||||
means=[[3, 2], [7, 7], [0, 8], [0, 10]],
|
||||
covs=[
|
||||
[[5, -1.5], [-1.5, 1]],
|
||||
[[1, 0.5], [0.5, 4]],
|
||||
[[5, 1], [1, 5]],
|
||||
[[3, 1], [1, 1]],
|
||||
],
|
||||
sizes=[100, 50, 50, 50],
|
||||
avg_trace=0.8,
|
||||
seed=SEED, # set to None for non-reproducible randomness
|
||||
):
|
||||
np.random.seed(seed=SEED)
|
||||
|
||||
K = len(means) # number of classes
|
||||
data = []
|
||||
labels = []
|
||||
test_data = []
|
||||
test_labels = []
|
||||
|
||||
for idx in range(K):
|
||||
data.append(
|
||||
np.random.multivariate_normal(
|
||||
mean=means[idx], cov=covs[idx], size=sizes[idx]
|
||||
)
|
||||
)
|
||||
test_data.append(
|
||||
np.random.multivariate_normal(
|
||||
mean=means[idx], cov=covs[idx], size=sizes[idx]
|
||||
)
|
||||
)
|
||||
labels.append(np.array([idx for i in range(sizes[idx])]))
|
||||
test_labels.append(np.array([idx for i in range(sizes[idx])]))
|
||||
X_train = np.vstack(data)
|
||||
y_train = np.hstack(labels)
|
||||
X_test = np.vstack(test_data)
|
||||
y_test = np.hstack(test_labels)
|
||||
|
||||
# Compute p(y=k) the prior distribution over true labels.
|
||||
py_true = np.bincount(y_train) / float(len(y_train))
|
||||
|
||||
noise_matrix_true = noise_generation.generate_noise_matrix_from_trace(
|
||||
K,
|
||||
trace=avg_trace * K,
|
||||
py=py_true,
|
||||
valid_noise_matrix=True,
|
||||
seed=SEED,
|
||||
)
|
||||
|
||||
# Generate our noisy labels using the noise_marix.
|
||||
s = noise_generation.generate_noisy_labels(y_train, noise_matrix_true)
|
||||
s_test = noise_generation.generate_noisy_labels(y_test, noise_matrix_true)
|
||||
ps = np.bincount(s) / float(len(s)) # Prior distribution over noisy labels
|
||||
|
||||
return X_train, s
|
||||
@@ -0,0 +1,8 @@
|
||||
coteaching
|
||||
==========
|
||||
|
||||
.. automodule:: cleanlab.experimental.coteaching
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,22 @@
|
||||
experimental
|
||||
============
|
||||
|
||||
.. warning::
|
||||
Methods in this ``experimental`` module are bleeding edge and may have sharp edges. They are not guaranteed to be stable between different ``cleanlab`` versions.
|
||||
|
||||
Useful methods/models adapted for use with cleanlab.
|
||||
|
||||
Some of these files include various models that can be used with cleanlab to find issues in specific types of data. These require dependencies on deep learning and other machine learning packages that are not official cleanlab dependencies. You must install these dependencies on your own if you wish to use them.
|
||||
|
||||
.. automodule:: cleanlab.experimental
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
|
||||
.. toctree::
|
||||
label_issues_batched
|
||||
span_classification
|
||||
mnist_pytorch
|
||||
coteaching
|
||||
cifar_cnn
|
||||
@@ -0,0 +1,8 @@
|
||||
label_issues_batched
|
||||
====================
|
||||
|
||||
.. automodule:: cleanlab.experimental.label_issues_batched
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,8 @@
|
||||
mnist_pytorch
|
||||
=============
|
||||
|
||||
.. automodule:: cleanlab.experimental.mnist_pytorch
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,8 @@
|
||||
span_classification
|
||||
===================
|
||||
|
||||
.. automodule:: cleanlab.experimental.span_classification
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,8 @@
|
||||
filter
|
||||
=======
|
||||
|
||||
.. automodule:: cleanlab.filter
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,23 @@
|
||||
internal
|
||||
========
|
||||
|
||||
.. warning::
|
||||
These ``internal`` utility methods are intended for internal use within the ``cleanlab`` package. They are not guaranteed to be stable between different versions.
|
||||
|
||||
.. automodule:: cleanlab.internal
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
|
||||
.. toctree::
|
||||
|
||||
util
|
||||
latent_algebra
|
||||
label_quality_utils
|
||||
multilabel_utils
|
||||
multilabel_scorer
|
||||
neighbor/index
|
||||
outlier
|
||||
token_classification_utils
|
||||
validation
|
||||
@@ -0,0 +1,8 @@
|
||||
label_quality_utils
|
||||
===================
|
||||
|
||||
.. automodule:: cleanlab.internal.label_quality_utils
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,8 @@
|
||||
latent_algebra
|
||||
==============
|
||||
|
||||
.. automodule:: cleanlab.internal.latent_algebra
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,8 @@
|
||||
multiannotator_utils
|
||||
==========================
|
||||
|
||||
.. automodule:: cleanlab.internal.multiannotator_utils
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,8 @@
|
||||
multilabel_scorer
|
||||
=================
|
||||
|
||||
.. automodule:: cleanlab.internal.multilabel_scorer
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,8 @@
|
||||
multilabel_utils
|
||||
================
|
||||
|
||||
.. automodule:: cleanlab.internal.multilabel_utils
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,21 @@
|
||||
neighbor
|
||||
========
|
||||
|
||||
The `neighbor` modules provide functionality for performing nearest neighbor search and pairwise distance calculations in those searches.
|
||||
|
||||
This submodule consists of the following modules:
|
||||
|
||||
- `neighbor.knn_graph`: Contains functions for setting up a nearest neighbor search index and constructing knn graphs.
|
||||
- `neighbor.search`: Contains a helper function that wraps the default implementation of nearest neighbor searches.
|
||||
- `neighbor.metric`: Contains functions for selecting distance metrics for nearest neighbor searches.
|
||||
|
||||
.. automodule:: cleanlab.internal.neighbor
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
|
||||
.. toctree::
|
||||
knn_graph
|
||||
metric
|
||||
search
|
||||
@@ -0,0 +1,8 @@
|
||||
knn_graph
|
||||
=========
|
||||
|
||||
.. automodule:: cleanlab.internal.neighbor.knn_graph
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,8 @@
|
||||
metric
|
||||
======
|
||||
|
||||
.. automodule:: cleanlab.internal.neighbor.metric
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,8 @@
|
||||
search
|
||||
======
|
||||
|
||||
.. automodule:: cleanlab.internal.neighbor.search
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,8 @@
|
||||
outlier
|
||||
=======
|
||||
|
||||
.. automodule:: cleanlab.internal.outlier
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,8 @@
|
||||
token_classification_utils
|
||||
==========================
|
||||
|
||||
.. automodule:: cleanlab.internal.token_classification_utils
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,8 @@
|
||||
util
|
||||
====
|
||||
|
||||
.. automodule:: cleanlab.internal.util
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,8 @@
|
||||
validation
|
||||
==========
|
||||
|
||||
.. automodule:: cleanlab.internal.validation
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,16 @@
|
||||
models
|
||||
======
|
||||
|
||||
.. warning::
|
||||
Methods in this ``models`` module are not guaranteed to be stable between different ``cleanlab`` versions.
|
||||
|
||||
Useful models adapted for use with cleanlab.
|
||||
|
||||
Some of these files include various models that can be used with cleanlab to find issues in specific types of data. These require dependencies on deep learning and other machine learning packages that are not official cleanlab dependencies. You must install these dependencies on your own if you wish to use them.
|
||||
|
||||
|
||||
.. automodule:: cleanlab.models
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,8 @@
|
||||
multiannotator
|
||||
==============
|
||||
|
||||
.. automodule:: cleanlab.multiannotator
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,8 @@
|
||||
dataset
|
||||
=======
|
||||
|
||||
.. automodule:: cleanlab.multilabel_classification.dataset
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,8 @@
|
||||
filter
|
||||
======
|
||||
|
||||
.. automodule:: cleanlab.multilabel_classification.filter
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,22 @@
|
||||
multilabel_classification
|
||||
=========================
|
||||
|
||||
Methods to detect data and label issues in multi-label classification datasets.
|
||||
|
||||
In multi-class classification, each example in the dataset belongs to exactly 1 out of K classes (e.g. if classifying animals as: {dog, cat, rat}).
|
||||
|
||||
In multi-label classification, each example in the dataset can belong to 1 or more classes (out of K possible classes), or none of the classes at all (e.g. if classifying animals as: {predator, pet, reptile}).
|
||||
|
||||
|
||||
|
||||
.. automodule:: cleanlab.multilabel_classification
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
|
||||
.. toctree::
|
||||
|
||||
filter
|
||||
rank
|
||||
dataset
|
||||
@@ -0,0 +1,8 @@
|
||||
rank
|
||||
====
|
||||
|
||||
.. automodule:: cleanlab.multilabel_classification.rank
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,9 @@
|
||||
filter
|
||||
====
|
||||
|
||||
.. automodule:: cleanlab.object_detection.filter
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
|
||||
@@ -0,0 +1,16 @@
|
||||
object_detection
|
||||
====================
|
||||
|
||||
Methods to detect label issues in object detection datasets.
|
||||
|
||||
.. automodule:: cleanlab.object_detection
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
|
||||
.. toctree::
|
||||
|
||||
rank
|
||||
filter
|
||||
summary
|
||||
@@ -0,0 +1,9 @@
|
||||
rank
|
||||
====
|
||||
|
||||
.. automodule:: cleanlab.object_detection.rank
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
|
||||
@@ -0,0 +1,9 @@
|
||||
summary
|
||||
====
|
||||
|
||||
.. automodule:: cleanlab.object_detection.summary
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
|
||||
@@ -0,0 +1,8 @@
|
||||
outlier
|
||||
==============
|
||||
|
||||
.. automodule:: cleanlab.outlier
|
||||
:autosummary:
|
||||
:members:
|
||||
:show-inheritance:
|
||||
:exclude-members: DEFAULT_PARAM_DICT, OOD_PARAMS, OUTLIER_PARAMS
|
||||
@@ -0,0 +1,8 @@
|
||||
rank
|
||||
====
|
||||
|
||||
.. automodule:: cleanlab.rank
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,15 @@
|
||||
regression
|
||||
====================
|
||||
|
||||
Methods to detect data and label issues in regression datasets.
|
||||
|
||||
.. automodule:: cleanlab.regression
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
|
||||
.. toctree::
|
||||
|
||||
rank
|
||||
learn
|
||||
@@ -0,0 +1,8 @@
|
||||
regression.learn
|
||||
================
|
||||
|
||||
.. automodule:: cleanlab.regression.learn
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,8 @@
|
||||
regression.rank
|
||||
===============
|
||||
|
||||
.. automodule:: cleanlab.regression.rank
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,8 @@
|
||||
filter
|
||||
====
|
||||
|
||||
.. automodule:: cleanlab.segmentation.filter
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,16 @@
|
||||
segmentation
|
||||
============
|
||||
|
||||
Methods to detect label issues in segmentation datasets.
|
||||
|
||||
.. automodule:: cleanlab.segmentation
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
|
||||
.. toctree::
|
||||
|
||||
rank
|
||||
filter
|
||||
summary
|
||||
@@ -0,0 +1,8 @@
|
||||
rank
|
||||
====
|
||||
|
||||
.. automodule:: cleanlab.segmentation.rank
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,8 @@
|
||||
summary
|
||||
=======
|
||||
|
||||
.. automodule:: cleanlab.segmentation.summary
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,8 @@
|
||||
filter
|
||||
======
|
||||
|
||||
.. automodule:: cleanlab.token_classification.filter
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,16 @@
|
||||
token_classification
|
||||
====================
|
||||
|
||||
Methods to detect data and label issues in token classification datasets.
|
||||
|
||||
.. automodule:: cleanlab.token_classification
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
|
||||
.. toctree::
|
||||
|
||||
filter
|
||||
rank
|
||||
summary
|
||||
@@ -0,0 +1,8 @@
|
||||
rank
|
||||
====
|
||||
|
||||
.. automodule:: cleanlab.token_classification.rank
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,8 @@
|
||||
summary
|
||||
=======
|
||||
|
||||
.. automodule:: cleanlab.token_classification.summary
|
||||
:autosummary:
|
||||
:members:
|
||||
:undoc-members:
|
||||
:show-inheritance:
|
||||
@@ -0,0 +1,296 @@
|
||||
# Configuration file for the Sphinx documentation builder.
|
||||
#
|
||||
# This file only contains a selection of the most common options. For a full
|
||||
# list see the documentation:
|
||||
# https://www.sphinx-doc.org/en/master/usage/configuration.html
|
||||
|
||||
# -- Path setup --------------------------------------------------------------
|
||||
|
||||
# If extensions (or modules to document with autodoc) are in another directory,
|
||||
# add these directories to sys.path here. If the directory is relative to the
|
||||
# documentation root, use os.path.abspath to make it absolute, like shown here.
|
||||
|
||||
import os
|
||||
import sys
|
||||
import datetime
|
||||
import shutil
|
||||
|
||||
sys.path.insert(0, os.path.abspath("../../cleanlab"))
|
||||
|
||||
# doctest setup
|
||||
# If extensions (or modules to document with autodoc) are in another directory,
|
||||
# add these directories to sys.path here.
|
||||
import pathlib
|
||||
|
||||
sys.path.insert(0, pathlib.Path(__file__).parents[2].resolve().as_posix())
|
||||
|
||||
# Open Graph extension
|
||||
ogp_site_url = "https://docs.cleanlab.ai"
|
||||
ogp_image = (
|
||||
"https://raw.githubusercontent.com/cleanlab/assets/master/cleanlab/clos-preview-card.png"
|
||||
)
|
||||
|
||||
|
||||
# -- Project information -----------------------------------------------------
|
||||
|
||||
project = "cleanlab"
|
||||
copyright = f"{datetime.datetime.now().year}, Cleanlab Inc."
|
||||
author = "Cleanlab Inc."
|
||||
|
||||
# -- General configuration ---------------------------------------------------
|
||||
|
||||
# Add any Sphinx extension module names here, as strings. They can be
|
||||
# extensions coming with Sphinx (named 'sphinx.ext.*') or your custom
|
||||
# ones.
|
||||
extensions = [
|
||||
"sphinx.ext.napoleon",
|
||||
"nbsphinx",
|
||||
"sphinx.ext.autodoc",
|
||||
"sphinx.ext.autosectionlabel",
|
||||
"autodocsumm",
|
||||
"sphinx.ext.viewcode",
|
||||
"sphinx.ext.todo",
|
||||
"sphinx_tabs.tabs",
|
||||
"sphinx_multiversion",
|
||||
"sphinx_copybutton",
|
||||
"sphinxcontrib.katex",
|
||||
"sphinxcontrib.gtagjs",
|
||||
"sphinx_jinja",
|
||||
"sphinx_autodoc_typehints",
|
||||
"sphinx.ext.doctest",
|
||||
"sphinxext.opengraph",
|
||||
]
|
||||
|
||||
numpy_show_class_members = True
|
||||
|
||||
# Don't add .txt suffix to source files:
|
||||
html_sourcelink_suffix = ""
|
||||
|
||||
# Add any paths that contain templates here, relative to this directory.
|
||||
templates_path = ["_templates"]
|
||||
|
||||
# List of patterns, relative to source directory, that match files and
|
||||
# directories to ignore when looking for source files.
|
||||
# This pattern also affects html_static_path and html_extra_path.
|
||||
exclude_patterns = ["_build"]
|
||||
|
||||
autosummary_generate = True
|
||||
|
||||
# set the default role of `name` to make cross references
|
||||
default_role = "py:obj"
|
||||
|
||||
# -- Options for apidoc extension ----------------------------------------------
|
||||
|
||||
# apidoc_module_dir = "cleanlab/cleanlab"
|
||||
|
||||
# -- Options for todo extension ----------------------------------------------
|
||||
|
||||
# If true, `todo` and `todoList` produce output, else they produce nothing.
|
||||
todo_include_todos = True
|
||||
|
||||
|
||||
# -- Options for Napoleon extension -------------------------------------------
|
||||
|
||||
napoleon_google_docstring = False
|
||||
napoleon_numpy_docstring = True
|
||||
napoleon_include_init_with_doc = False
|
||||
napoleon_include_private_with_doc = False
|
||||
napoleon_include_special_with_doc = True
|
||||
napoleon_use_admonition_for_examples = False
|
||||
napoleon_use_admonition_for_notes = False
|
||||
napoleon_use_admonition_for_references = False
|
||||
napoleon_use_ivar = False
|
||||
napoleon_use_param = True
|
||||
napoleon_use_rtype = False
|
||||
napoleon_preprocess_types = True
|
||||
napoleon_type_aliases = None
|
||||
napoleon_attr_annotations = True
|
||||
|
||||
# -- Options for autodoc extension -------------------------------------------
|
||||
|
||||
# This value selects what content will be inserted into the main body of an autoclass
|
||||
# directive
|
||||
#
|
||||
# http://www.sphinx-doc.org/en/master/usage/extensions/autodoc.html#directive-autoclass
|
||||
autoclass_content = "class"
|
||||
|
||||
|
||||
# Default options to an ..autoXXX directive.
|
||||
autodoc_default_options = {
|
||||
"autosummary": True,
|
||||
"members": None,
|
||||
"inherited-members": None,
|
||||
"show-inheritance": None,
|
||||
"special-members": "__call__",
|
||||
}
|
||||
|
||||
# Subclasses should show parent classes docstrings if they don't override them.
|
||||
autodoc_inherit_docstrings = True
|
||||
|
||||
# Order functions displayed by the order of source code
|
||||
autodoc_member_order = "bysource"
|
||||
|
||||
# -- Options for copybutton extension -----------------------------------------
|
||||
|
||||
# Strip input prompts when copying code blocks. Supports:
|
||||
# - Python Repl + continuation prompt
|
||||
# - Bash prompt
|
||||
# - ipython + continuation prompt
|
||||
# - jupyter-console + continuation prompt
|
||||
copybutton_prompt_text = r">>> |\.\.\. |\$ |In \[\d*\]: | {2,5}\.\.\.: | {5,8}: "
|
||||
copybutton_prompt_is_regexp = True
|
||||
|
||||
# -- Options for katex extension -------------------------------------------
|
||||
|
||||
if os.getenv("CI") or shutil.which("katex") is not None:
|
||||
# requires that the machine have `katex` installed: `npm install -g katex`
|
||||
katex_prerender = True
|
||||
|
||||
# -- Options for gtagjs extension -------------------------------------------
|
||||
|
||||
gtagjs_ids = ["G-EV8RVEFX82"]
|
||||
|
||||
# -- Variables Setting ---------------------------------------------------
|
||||
|
||||
# Determine doc site URL (DOCS_SITE_URL)
|
||||
# Check if it's running in production repo
|
||||
if os.getenv("GITHUB_REPOSITORY") == "cleanlab/cleanlab":
|
||||
DOCS_SITE_URL = "/"
|
||||
else:
|
||||
DOCS_SITE_URL = "/cleanlab-docs/"
|
||||
|
||||
gh_env_file = os.getenv("GITHUB_ENV")
|
||||
if gh_env_file is not None:
|
||||
with open(gh_env_file, "a") as f:
|
||||
f.write(f"\nDOCS_SITE_URL={DOCS_SITE_URL}") # Set to Environment Var
|
||||
|
||||
GITHUB_REPOSITORY_OWNER = os.getenv("GITHUB_REPOSITORY_OWNER") or "cleanlab"
|
||||
GITHUB_REF_NAME = os.getenv("GITHUB_REF_NAME") or "master"
|
||||
|
||||
# Pass additional variables to Jinja templates
|
||||
html_context = {
|
||||
"DOCS_SITE_URL": DOCS_SITE_URL,
|
||||
# Add new tags to RELEASE_VERSIONS before release
|
||||
# fmt: off
|
||||
"RELEASE_VERSIONS": [
|
||||
"v2.7.1",
|
||||
"v2.7.0",
|
||||
"v2.6.6",
|
||||
"v2.6.5",
|
||||
"v2.6.4",
|
||||
"v2.6.3",
|
||||
"v2.6.2",
|
||||
"v2.6.1",
|
||||
"v2.6.0",
|
||||
"v2.5.0",
|
||||
"v2.4.0",
|
||||
"v2.3.1",
|
||||
"v2.3.0",
|
||||
"v2.2.0",
|
||||
"v2.1.0",
|
||||
"v2.0.0",
|
||||
"v1.0.1",
|
||||
],
|
||||
# fmt: on
|
||||
}
|
||||
|
||||
# -- nbsphinx Configuration ---------------------------------------------------
|
||||
|
||||
# This is processed by Jinja2 and inserted before each notebook
|
||||
nbsphinx_prolog = (
|
||||
"""
|
||||
{% set docname = env.doc2path(env.docname, base=None) %}
|
||||
|
||||
.. raw:: html
|
||||
|
||||
<style>
|
||||
.nbinput .prompt,
|
||||
.nboutput .prompt {
|
||||
display: none;
|
||||
}
|
||||
|
||||
.output_area {
|
||||
max-height: 300px;
|
||||
overflow: auto;
|
||||
}
|
||||
|
||||
.dataframe {
|
||||
background: #D7D7D7;
|
||||
}
|
||||
|
||||
th {
|
||||
color:black;
|
||||
}
|
||||
</style>
|
||||
|
||||
<script type="text/javascript">
|
||||
window.addEventListener('load', () => {
|
||||
const h1_element = document.getElementsByTagName("h1");
|
||||
h1_element[0].insertAdjacentHTML("afterend", `
|
||||
<p>
|
||||
<a style="background-color:white;color:black;padding:4px 12px;text-decoration:none;display:inline-block;border-radius:8px;box-shadow:0 2px 4px 0 rgba(0, 0, 0, 0.2), 0 3px 10px 0 rgba(0, 0, 0, 0.19)" href="https://colab.research.google.com/github/"""
|
||||
+ GITHUB_REPOSITORY_OWNER
|
||||
+ """/cleanlab-docs/blob/master/"""
|
||||
+ GITHUB_REF_NAME
|
||||
+ """/{{ docname|e }}" target="_blank">
|
||||
<img src="https://colab.research.google.com/img/colab_favicon_256px.png" alt="" style="width:40px;height:40px;vertical-align:middle">
|
||||
<span style="vertical-align:middle">Run in Google Colab</span>
|
||||
</a>
|
||||
</p>
|
||||
`);
|
||||
})
|
||||
|
||||
</script>
|
||||
"""
|
||||
)
|
||||
|
||||
# Change this to "always" before running in the doc's CI/CD server
|
||||
if os.getenv("CI"):
|
||||
nbsphinx_execute = "always"
|
||||
if os.getenv("SKIP_NOTEBOOKS", "0") != "0":
|
||||
nbsphinx_execute = "never"
|
||||
|
||||
# -- Options for HTML output -------------------------------------------------
|
||||
|
||||
# The theme to use for HTML and HTML Help pages. See the documentation for
|
||||
# a list of builtin themes.
|
||||
#
|
||||
html_theme = "furo"
|
||||
html_favicon = "https://raw.githubusercontent.com/cleanlab/assets/a4483476d449f2f05a4c7cde329e72358099cc07/cleanlab/cleanlab_favicon.svg"
|
||||
html_title = "cleanlab"
|
||||
html_logo = (
|
||||
"https://raw.githubusercontent.com/cleanlab/assets/master/cleanlab/cleanlab_logo_only.png"
|
||||
)
|
||||
html_theme_options = {
|
||||
"footer_icons": [
|
||||
{
|
||||
"name": "GitHub",
|
||||
"url": "https://github.com/cleanlab/cleanlab",
|
||||
"html": """
|
||||
<svg stroke="currentColor" fill="currentColor" stroke-width="0" viewBox="0 0 16 16">
|
||||
<path fill-rule="evenodd" d="M8 0C3.58 0 0 3.58 0 8c0 3.54 2.29 6.53 5.47 7.59.4.07.55-.17.55-.38 0-.19-.01-.82-.01-1.49-2.01.37-2.53-.49-2.69-.94-.09-.23-.48-.94-.82-1.13-.28-.15-.68-.52-.01-.53.63-.01 1.08.58 1.23.82.72 1.21 1.87.87 2.33.66.07-.52.28-.87.51-1.07-1.78-.2-3.64-.89-3.64-3.95 0-.87.31-1.59.82-2.15-.08-.2-.36-1.02.08-2.12 0 0 .67-.21 2.2.82.64-.18 1.32-.27 2-.27.68 0 1.36.09 2 .27 1.53-1.04 2.2-.82 2.2-.82.44 1.1.16 1.92.08 2.12.51.56.82 1.27.82 2.15 0 3.07-1.87 3.75-3.65 3.95.29.25.54.73.54 1.48 0 1.07-.01 1.93-.01 2.2 0 .21.15.46.55.38A8.013 8.013 0 0 0 16 8c0-4.42-3.58-8-8-8z"></path>
|
||||
</svg>
|
||||
""",
|
||||
"class": "",
|
||||
},
|
||||
],
|
||||
}
|
||||
|
||||
# Add any paths that contain custom static files (such as style sheets) here,
|
||||
# relative to this directory. They are copied after the builtin static files,
|
||||
# so a file named "default.css" will overwrite the builtin "default.css".
|
||||
html_static_path = ["_static"]
|
||||
|
||||
html_css_files = ["css/custom.css"]
|
||||
|
||||
html_sidebars = {
|
||||
"**": [
|
||||
"brand.html",
|
||||
"sidebar/search.html",
|
||||
"sidebar/scroll-start.html",
|
||||
"sidebar/navigation.html",
|
||||
"sidebar/ethical-ads.html",
|
||||
"versioning.html",
|
||||
"sidebar/scroll-end.html",
|
||||
],
|
||||
}
|
||||
@@ -0,0 +1,208 @@
|
||||
:og:title: Open-Source Documentation | Cleanlab
|
||||
:og:description: Get started with Cleanlab Open-Source, learn about capabilities, and follow tutorials to improve your own Data and Models.
|
||||
|
||||
cleanlab open-source documentation
|
||||
==================================
|
||||
|
||||
`cleanlab <https://github.com/cleanlab/cleanlab>`_ **automatically detects data and label issues in your ML datasets.**
|
||||
|
||||
| This helps you improve your data and train robust ML models on noisy real-world datasets. cleanlab has already found thousands of `label errors <https://labelerrors.com>`_ in ImageNet, MNIST, and other popular ML benchmarking datasets. Beyond handling label errors, this is a comprehensive open-source library implementing many data-centric AI capabilities. Start using automation to improve your data!
|
||||
|
||||
Quickstart
|
||||
==========
|
||||
|
||||
1. Install ``cleanlab``
|
||||
-----------------------
|
||||
|
||||
.. tabs::
|
||||
|
||||
.. tab:: uv
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
uv pip install cleanlab
|
||||
|
||||
To install the package with all optional dependencies:
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
uv pip install "cleanlab[all]"
|
||||
|
||||
.. tab:: pip
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
pip install cleanlab
|
||||
|
||||
To install the package with all optional dependencies:
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
pip install "cleanlab[all]"
|
||||
|
||||
|
||||
.. tab:: conda
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
conda install -c cleanlab cleanlab
|
||||
|
||||
.. tab:: source
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
uv pip install git+https://github.com/cleanlab/cleanlab.git
|
||||
|
||||
Or with pip:
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
pip install git+https://github.com/cleanlab/cleanlab.git
|
||||
|
||||
To install the package with all optional dependencies:
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
uv pip install "git+https://github.com/cleanlab/cleanlab.git#egg=cleanlab[all]"
|
||||
|
||||
|
||||
2. Check your data for all sorts of issues
|
||||
------------------------------------------
|
||||
|
||||
cleanlab automatically detects various issues in *any dataset that a classifier can be trained on*. The cleanlab package *works with any ML model* by operating on model outputs (predicted class probabilities or feature embeddings) -- it doesn't require that a particular model created those outputs. For any classification dataset, use your trained model to produce `pred_probs` (predicted class probabilities) and/or `feature_embeddings` (numeric vector representations of each datapoint). To automatically check your dataset for common real-world issues (like label errors, outliers, near duplicates, IID violations, underperforming groups, ...), simply run these few lines of code:
|
||||
|
||||
.. code-block:: python
|
||||
|
||||
from cleanlab import Datalab
|
||||
|
||||
lab = Datalab(data=your_dataset, label_name="column_name_of_labels")
|
||||
lab.find_issues(features=feature_embeddings, pred_probs=pred_probs)
|
||||
lab.report() # summarize issues in dataset, how severe they are in each data point, ...
|
||||
|
||||
While other data quality tools only catch limited types of data issues based on manually pre-defined validation rules, cleanlab applies automated data-centric AI techniques using your trained ML model to detect many more types of data issues that would otherwise be hard to catch. Don't dive into ML model improvement without first using AI to help check your data!
|
||||
|
||||
|
||||
3. Handle label errors and train robust models with noisy labels
|
||||
----------------------------------------------------------------
|
||||
|
||||
Mislabeled data is a particularly concerning issue plaguing real-world datasets. To use a scikit-learn-compatible model for classification with noisy labels, you don't need to train a model to find label issues -- you can pass the untrained model object, data, and labels into :py:meth:`CleanLearning.find_label_issues <cleanlab.classification.CleanLearning.find_label_issues>` and cleanlab will handle model training for you.
|
||||
|
||||
.. code-block:: python
|
||||
|
||||
from cleanlab.classification import CleanLearning
|
||||
|
||||
# This works with any sklearn-compatible model - just input data + labels and cleanlab will detect label issues ツ
|
||||
label_issues_info = CleanLearning(clf=sklearn_compatible_model).find_label_issues(data, labels)
|
||||
|
||||
:py:class:`CleanLearning <cleanlab.classification.CleanLearning>` also works with models from most standard ML frameworks by wrapping the model for scikit-learn compliance, e.g. pytorch (can use `skorch <https://github.com/skorch-dev/skorch>`_ package), etc.
|
||||
|
||||
:py:meth:`find_label_issues <cleanlab.classification.CleanLearning.find_label_issues>` returns a boolean mask flagging which examples have label issues and a numeric label quality score for each example quantifying our confidence that its label is correct.
|
||||
|
||||
Beyond standard classification tasks, cleanlab can also detect mislabeled examples in: `multi-label data <tutorials/multilabel_classification.ipynb>`_ (e.g. image/document tagging), `sequence prediction <tutorials/token_classification.ipynb>`_ (e.g. entity recognition), and `data labeled by multiple annotators <tutorials/multiannotator.ipynb>`_ (e.g. crowdsourcing).
|
||||
|
||||
.. important::
|
||||
Cleanlab performs better if the ``pred_probs`` from your model are **out-of-sample**. Details on how to compute out-of-sample predicted probabilities for your entire dataset are :ref:`here <pred_probs_cross_val>`.
|
||||
|
||||
cleanlab's :py:class:`CleanLearning <cleanlab.classification.CleanLearning>` class trains a more robust version of any existing (`scikit-learn <https://scikit-learn.org/>`_ `compatible <https://scikit-learn.org/stable/developers/develop.html>`_) classification model, `clf`, by fitting it to an automatically filtered version of your dataset with low-quality data removed. It returns a model trained only on the clean data, from which you can get predictions in the same way as your existing classifier.
|
||||
|
||||
.. code-block:: python
|
||||
|
||||
from sklearn.linear_model import LogisticRegression
|
||||
from cleanlab.classification import CleanLearning
|
||||
|
||||
cl = CleanLearning(clf=LogisticRegression()) # any sklearn-compatible classifier
|
||||
cl.fit(train_data, labels)
|
||||
|
||||
# Estimate the predictions you would have gotten if you trained without mislabeled data
|
||||
predictions = cl.predict(test_data)
|
||||
|
||||
|
||||
4. Dataset curation: fix dataset-level issues
|
||||
---------------------------------------------
|
||||
|
||||
cleanlab's `dataset <tutorials/dataset_health.ipynb>`_ module helps you deal with dataset-level issues -- :py:meth:`find overlapping classes <cleanlab.dataset.find_overlapping_classes>` (classes to merge), :py:meth:`rank class-level label quality <cleanlab.dataset.rank_classes_by_label_quality>` (classes to keep/delete), and :py:meth:`measure overall dataset health <cleanlab.dataset.overall_label_health_score>` (to track dataset quality as you make adjustments).
|
||||
|
||||
View all dataset-level issues in one line of code with :py:meth:`dataset.health_summary() <cleanlab.dataset.health_summary>`.
|
||||
|
||||
.. code-block:: python
|
||||
|
||||
from cleanlab.dataset import health_summary
|
||||
|
||||
health_summary(labels, pred_probs, class_names=class_names)
|
||||
|
||||
|
||||
5. Improve your data via many other techniques
|
||||
----------------------------------------------
|
||||
|
||||
Beyond handling label errors, cleanlab supports other data-centric AI capabilities including:
|
||||
|
||||
- Detecting outliers and out-of-distribution examples in both training and future test data `(tutorial) <tutorials/outliers.ipynb>`_
|
||||
- Analyzing data labeled by multiple annotators to estimate consensus labels and their quality `(tutorial) <tutorials/multiannotator.ipynb>`_
|
||||
- Active learning with multiple annotators to identify which data is most informative to label or re-label next `(tutorial) <https://github.com/cleanlab/examples/blob/master/active_learning_multiannotator/active_learning.ipynb>`_
|
||||
|
||||
|
||||
If you have questions, check out our `FAQ <tutorials/faq.ipynb>`_ and feel free to open a Github Issue!
|
||||
|
||||
|
||||
.. toctree::
|
||||
:hidden:
|
||||
|
||||
Quickstart <self>
|
||||
|
||||
.. toctree::
|
||||
:hidden:
|
||||
:caption: Tutorials
|
||||
|
||||
Datalab Tutorials <tutorials/datalab/index>
|
||||
Improving ML Performance <tutorials/improving_ml_performance>
|
||||
CleanLearning Tutorials <tutorials/clean_learning/index>
|
||||
Workflows of Data-Centric AI <tutorials/indepth_overview>
|
||||
Analyze Dataset-level Issues <tutorials/dataset_health>
|
||||
Outlier Detection <tutorials/outliers>
|
||||
Improving Consensus Labels for Multiannotator Data <tutorials/multiannotator>
|
||||
Multi-Label Classification <tutorials/multilabel_classification>
|
||||
Noisy Labels in Regression <tutorials/regression>
|
||||
Token Classification (text) <tutorials/token_classification>
|
||||
Image Segmentation <tutorials/segmentation>
|
||||
Object Detection <tutorials/object_detection>
|
||||
Predicted Probabilities via Cross Validation <tutorials/pred_probs_cross_val>
|
||||
FAQ <tutorials/faq>
|
||||
|
||||
.. toctree::
|
||||
:caption: API Reference
|
||||
:hidden:
|
||||
:maxdepth: 3
|
||||
|
||||
cleanlab/datalab/index
|
||||
cleanlab/classification
|
||||
cleanlab/filter
|
||||
cleanlab/rank
|
||||
cleanlab/count
|
||||
cleanlab/dataset
|
||||
cleanlab/outlier
|
||||
cleanlab/multiannotator
|
||||
cleanlab/data_valuation
|
||||
cleanlab/multilabel_classification/index
|
||||
cleanlab/regression/index
|
||||
cleanlab/token_classification/index
|
||||
cleanlab/segmentation/index
|
||||
cleanlab/object_detection/index
|
||||
cleanlab/benchmarking/index
|
||||
cleanlab/models/index
|
||||
cleanlab/experimental/index
|
||||
cleanlab/internal/index
|
||||
|
||||
.. toctree::
|
||||
:caption: Guides
|
||||
:hidden:
|
||||
|
||||
Datalab issue types <cleanlab/datalab/guide/index>
|
||||
|
||||
.. toctree::
|
||||
:caption: Links
|
||||
:hidden:
|
||||
|
||||
GitHub <https://github.com/cleanlab/cleanlab>
|
||||
PyPI <https://pypi.org/project/cleanlab/>
|
||||
Conda <https://anaconda.org/conda-forge/cleanlab>
|
||||
Blog <https://cleanlab.ai/blog/>
|
||||
@@ -0,0 +1,87 @@
|
||||
How to migrate to versions >= 2.0.0 from pre 1.0.1
|
||||
==================================================
|
||||
|
||||
If you previously used older versions of cleanlab,
|
||||
this guide helps update your existing code to work with versions >= 2.0.0 in no time!
|
||||
Below we outline the major updates and code substitutions to be aware of.
|
||||
A detailed API change-log is listed in the `v2.0.0. Release Notes <https://github.com/cleanlab/cleanlab/releases/tag/v2.0.0>`_.
|
||||
|
||||
|
||||
Function and class name changes
|
||||
-------------------------------
|
||||
|
||||
This section covers the most commonly-used functionality from Cleanlab 1.0.
|
||||
|
||||
| **Old:** ``pruning.get_noise_indices(s, psx, prune_method, sorted_index_method, ...)``
|
||||
| -->
|
||||
| **New:** :py:func:`filter.find_label_issues <cleanlab.filter.find_label_issues>` ``(labels, pred_probs, filter_by, return_indices_ranked_by, ...)``
|
||||
|
||||
Note: ``inverse_noise_matrix`` is no longer a supported input argument, but ``confident_joint`` remains (you can easily convert between these two).
|
||||
|
||||
----
|
||||
|
||||
| **Old:** ``pruning.order_label_errors(label_errors_bool, psx, labels, sorted_index_method)``
|
||||
| -->
|
||||
| **New:** :py:func:`rank.order_label_issues <cleanlab.rank.order_label_issues>` ``(label_issues_mask, labels, pred_probs, rank_by, ...)``
|
||||
|
||||
Note: You can now alternatively use :py:func:`rank.get_label_quality_score() <cleanlab.rank.get_label_quality_score>` to numerically score the labels instead of ranking them.
|
||||
|
||||
----
|
||||
|
||||
| **Old:** ``latent_estimation.num_label_errors(labels, psx, ...)``
|
||||
| -->
|
||||
| **New:** :py:func:`count.num_label_issues <cleanlab.count.num_label_issues>` ``(labels, pred_probs, ...)``
|
||||
|
||||
Note: This is the most accurate way to estimate the raw *number* of label errors in a dataset.
|
||||
|
||||
----
|
||||
|
||||
| **Old:** ``classification.LearningWithNoisyLabels(..., prune_method)``
|
||||
| -->
|
||||
| **New:** :py:class:`classification.CleanLearning <cleanlab.classification.CleanLearning>` ``(..., find_label_issues_kwargs)``
|
||||
|
||||
Note: :py:class:`CleanLearning <cleanlab.classification.CleanLearning>` can now find label errors for you, neatly organizing them in a ``pandas.DataFrame`` as well as computing the required out-of-sample predicted probabilities. You just specify which classifier, we handle the cross-validation!
|
||||
|
||||
|
||||
Module name changes
|
||||
-------------------
|
||||
|
||||
Reorganized modules:
|
||||
|
||||
- ``cleanlab.pruning`` --> :py:mod:`cleanlab.filter`
|
||||
- ``cleanlab.latent_estimation`` --> :py:mod:`cleanlab.count`
|
||||
- ``cleanlab.noise_generation`` --> :py:mod:`cleanlab.benchmarking.noise_generation`
|
||||
- ``cleanlab.baseline_methods`` --> incorporated into :py:mod:`cleanlab.filter`
|
||||
|
||||
Internal and experimental functionality, marked as such and not guaranteed to be stable between releases:
|
||||
|
||||
- ``cleanlab.models`` --> :py:mod:`cleanlab.experimental`
|
||||
- ``cleanlab.coteaching`` --> :py:mod:`cleanlab.experimental.coteaching`
|
||||
- ``cleanlab.latent_algebra`` --> :py:mod:`cleanlab.internal.latent_algebra`
|
||||
- ``cleanlab.util`` --> :py:mod:`cleanlab.internal.util`
|
||||
|
||||
|
||||
New modules
|
||||
-----------
|
||||
|
||||
- :py:mod:`cleanlab.dataset` : New methods to print summaries of overall types of label issues most common in a dataset.
|
||||
- :py:mod:`cleanlab.rank` : Moved all ranking and ordering functions from ``cleanlab.pruning`` to here. This module contains methods to score the label quality of each example and rank your data by the quality of their labels.
|
||||
- :py:mod:`cleanlab.internal` and :py:mod:`cleanlab.experimental`: Moved all advanced code and utility methods to this module, including the old ``cleanlab.latent_algebra`` module. Researchers may find useful functions in here.
|
||||
|
||||
|
||||
Removed modules
|
||||
---------------
|
||||
|
||||
- ``cleanlab.polyplex``
|
||||
|
||||
|
||||
Common argument and variable name changes
|
||||
-----------------------------------------
|
||||
|
||||
Here are some common name and terminology changes in Cleanlab 2.0:
|
||||
|
||||
- ``s`` --> ``labels`` (the given labels in the data, which are potentially noisy)
|
||||
- ``psx`` --> ``pred_probs`` (predicted probabilities output by trained classifier)
|
||||
- ``label_error`` --> ``label_issue`` (a label that is likely to be wrong)
|
||||
|
||||
See the documentation for individual functions for details on how argument names changed.
|
||||
@@ -0,0 +1,3 @@
|
||||
*.tar.gz
|
||||
pretrained_models/
|
||||
spoken_digits/
|
||||
@@ -0,0 +1,8 @@
|
||||
CleanLearning Tutorials
|
||||
=======================
|
||||
|
||||
.. toctree::
|
||||
:maxdepth: 1
|
||||
|
||||
Text Classification <text>
|
||||
Tabular Classification (Numeric/Categorical) <tabular>
|
||||
Some files were not shown because too many files have changed in this diff Show More
Reference in New Issue
Block a user