A Streamlit component integrating Label Studio Frontend in Streamlit applications
date: "2024-07-09"
author:
name: "Arvindra Sehmi"
url: "https://www.linkedin.com/in/asehmi/" | "https://sehmi-conscious.medium.com/about"
mail: "vin [at] thesehmis.com"
avatar: "https://twitter.com/asehmi/profile_image?size=original"
related: [Introduction to Streamlit and Streamlit Components](https://auth0.com/blog/introduction-to-streamlit-and-streamlit-components/)
Label Studio is an open source data labeling tool providing flexible data annotation. Label Studio comprises Label Studio Backend (LSB) server and ML service and Label Studio Frontend (LSF) which is based on React and mobx-state-tree and distributed as an NPM package. LSF can be integrated in third-party applications without using LSB to provide data annotation support to users. LSF can be customized and extended to build custom UIs or used with pre-built labeling templates.
This Streamlit application leverages Streamlit Components extensibilty with the simple architecture of Component Zero discussed in my article, Introduction to Streamlit and Streamlit Components. Using Component Zero as a template it was straight-forward to take the code snippet in Label Studio's Frontend integration guide and build the Streamlit component.
The standalone LSF application has been deprecated. Hence, I'm not sure if the package repository is being updated from the main LS GitHub repository. Nevertheless, I updated the app_config.json file to point to the newer examples, and added image polygon and image segmentation examples.
Moving forward, you'd have to create a build of the editor (i.e. LSF) from the main LS repo, and point to these build JS and CSS files instead of taking them from the package repo.
The application can be seen running in the Streamlit Cloud at the link below:
$ cd streamlit-label-studio-frontend
$ pip install -r requirements.txtRun the included app for a quick example of how the LSF Streamlit component I built works:
$ streamlit run app.pyThe Streamlit user interface is used to load any supported annotation task configurations which Label Studio Frontend uses to display annotation tools appropriate for the annotation task at hand. Once annotations are done, the annotations data can be submitted via the custom LSF Streamlit component to Streamlit, where it is displayed. You can of course extend the Streamlit application and store or export this data as you see fit. You can leverage LSB API from Streamlit, though that is out of scope for this demo app.
The demo app is easily customised through externalized configuration files.
Configuration for labeling tasks in made in app_configs.json. There are three sections required for annotation: "user", "interfaces" and "task_configs".
In "task_configs" you define "task" objects explicitly using the task "object" key, or via a local file using the task "file" key, or from an external site using the task "url" key. Similarly, you can define Label Studio annotation UI elements using "config" objects.
Below are some examples:
./app_configs.json
{
"user" : {
"pk": 1,
"firstName": "Arvindra",
"lastName": "Sehmi"
},
"interfaces": [
"panel",
"update",
"submit",
"skip",
"controls",
"infobar",
"topbar",
"instruction",
"side-column",
"annotations:history",
"annotations:tabs",
"annotations:menu",
"annotations:current",
"annotations:add-new",
"annotations:delete",
"annotations:view-all",
"predictions:tabs",
"predictions:menu",
"auto-annotation",
"edit-history"
],
"task_configs": [
{
"name": "My Bounding Box Single",
"description": "My Bounding Box Single Image Annotation",
"annotation_type": "image",
"config": {
"file": "./label_studio_configs/xml/bounding_box_config.xml"
},
"task": {
"object": {
"annotations": [],
"predictions": [],
"id": 1,
"data": {
"image": "https://htx-pub.s3.us-east-1.amazonaws.com/examples/images/nick-owuor-astro-nic-visuals-wDifg5xc9Z4-unsplash.jpg"
}
}
}
},
{
"name": "My Bounding Box Multi",
"description": "My Bounding Box Multiple Images Annotation",
"annotation_type": "image",
"config": {
"file": "./label_studio_configs/xml/example_bounding_box_config.xml"
},
"task": {
"file": "./label_studio_configs/json/example_tasks.json"
}
},
{
"name": "Bounding Box Example",
"description": "Bounding Box Image Annotation Example",
"annotation_type": "image",
"config": {
"url": "https://raw.githubusercontent.com/HumanSignal/label-studio/develop/web/libs/editor/src/examples/image_bbox/config.xml"
},
"task": {
"url": "https://raw.githubusercontent.com/HumanSignal/label-studio/develop/web/libs/editor/src/examples/image_bbox/tasks.json"
}
}
]
}./label_studio_configs/xml/bounding_box_config.xml
<View>
<View style="padding: 25px; box-shadow: 2px 2px 8px #AAA;">
<Image name="img" value="$image" width="100%" brightnessControl="true" contrastControl="true" zoomControl="true" rotateControl="true">
</Image>
</View>
<RectangleLabels name="tag" toName="img">
<Label value="Galaxy"></Label>
<Label value="Star"></Label>
<Label value="Comet"></Label>
<Label value="Moon"></Label>
<Label value="Sky"></Label>
<Label value="Tree"></Label>
<Label value="Road"></Label>
<Label value="Fence"></Label>
</RectangleLabels>
</View>./label_studio_configs/json/example_tasks.json
[
{
"annotations": [],
"predictions": [
{
"model_version": "model 1",
"created_ago": "3 hours",
"result": [
{
"from_name": "tag",
"id": "t5sp3TyXPo",
"source": "$image",
"to_name": "img",
"type": "rectanglelabels",
"value": {
"height": 11.612284069097889,
"rectanglelabels": [
"Moonwalker"
],
"rotation": 0,
"width": 39.6,
"x": 13.2,
"y": 34.702495201535505
}
}
]
},
{
"model_version": "model 2",
"created_ago": "4 hours",
"result": [
{
"from_name": "tag",
"id": "t5sp3TyXPo",
"source": "$image",
"to_name": "img",
"type": "rectanglelabels",
"value": {
"height": 33.61228406909789,
"rectanglelabels": [
"Moonwalker"
],
"rotation": 0,
"width": 39.6,
"x": 13.2,
"y": 54.702495201535505
}
}
]
}
],
"id": 1,
"data": {
"image": [
"https://htx-pub.s3.us-east-1.amazonaws.com/examples/images/nick-owuor-astro-nic-visuals-wDifg5xc9Z4-unsplash.jpg",
"https://htx-pub.s3.us-east-1.amazonaws.com/examples/images/nick-owuor-astro-nic-visuals-wDifg5xc9Z4-unsplash.jpg"
]
}
},
{
"annotations": [],
"predictions": [],
"id": 2,
"data": {
"image": "https://htx-pub.s3.us-east-1.amazonaws.com/examples/images/nick-owuor-astro-nic-visuals-wDifg5xc9Z4-unsplash.jpg"
}
}
]More configurations can be found in the official Label Studio examples
Because of the way Streamlit references the base directory of components, local images must be stored relative to the component's base.
NOTE: The Label Studio Frontend component base is ./frontend, so images should be placed in ./frontend/images, and referenced in tasks from the relative path ./images.
For example:
"task_configs": [
{
"name": "My Bounding Box Single (Local Image)",
"description": "My Bounding Box Single Image Annotation (Local Image)",
"annotation_type": "image",
"config": {
"file": "./label_studio_configs/xml/bounding_box_config.xml"
},
"task": {
"object": {
"annotations": [],
"predictions": [],
"id": 1,
"data": {
"image": "./images/nick-owuor-astro-nic-visuals-wDifg5xc9Z4-unsplash.jpg"
}
}
}
}
]If you use vs-code, then I have included my .vscode launch settings. You can adjust the --server.port settings for Streamlit in launch.json.
A major benefit I found building a static HTML Streamlit component was that it can be debugged very easily using a browser's developer tools. This is in contrast to the obscure debugging experience you encounter using more sophisticated web frameworks.
If you enjoyed this app, please consider starring this repository.
Thanks!
Arvindra