Getting started

You can use the Label Studio Python SDK to make annotating data a more integrated part of your data science and machine learning pipelines. This software development kit (SDK) lets you call the Label Studio API directly from scripts using predefined classes and methods.

With the Label Studio Python SDK, you can perform the following tasks in a Python script:

• Authenticate to the Label Studio API
• Create a Label Studio project, including setting up a labeling configuration.
• Import tasks.
• Manage pre-annotated tasks and model predictions.
• Connect to a cloud storage provider, such as Amazon S3, Microsoft Azure, or Google Cloud Services (GCS), to retrieve unlabeled tasks and store annotated tasks.
• Manage labeling jobs by creating custom filters and ordering for tasks based on parameters that you specify.
• Export annotations from Label Studio.

Install

Install the Label Studio SDK using pip:

pip install label-studio-sdk

Authentication

In your Python script, do the following:

• Import the SDK.
• Define your API key and Label Studio URL (API key is available at Account page).
• Connect to the API.

1
# Define the URL where Label Studio is accessible and the API key for your user account
2
LABEL_STUDIO_URL = 'http://localhost:8080'
3
API_KEY = 'd6f8a2622d39e9d89ff0dfef1a80ad877f4ee9e3'
4
5
# Import the SDK and the client module
6
from label_studio_sdk.client import LabelStudio
7
8
# Connect to the Label Studio API and check the connection
9
ls = LabelStudio(base_url=LABEL_STUDIO_URL, api_key=API_KEY)

Create a Project

After you connect to the Label Studio API, you can create a project in Label Studio using the SDK. Specify the project title and the labeling configuration. Choose your labeling configuration based on the type of labeling that you wish to perform. See the available templates for Label Studio projects, or set a blank configuration with <View></View>.

For example, create a text classification project in your Python code:

1
project = ls.projects.create(
2
    title='Text Classification',
3
    label_config='''
4
    <View>
5
        <Text name="text" value="$text" />
6
        <Choices name="label" toName="text" choice="single">
7
            <Choice value="Positive" />
8
            <Choice value="Negative" />
9
        </Choices>
10
    </View>'''
11
)

For more about what you can do with the project module of the SDK, see the project module SDK reference.

Import Tasks

You can import tasks from your script using the Label Studio Python SDK client.

For a specific project created, you can import tasks in Label Studio JSON format or connect to cloud storage providers and import image, audio, or video files directly.

Add Model Predictions

You can add predictions to existing tasks in Label Studio in your Python script.

For an existing simple image classification project, you can do the following to add predictions of “Dog” for image tasks that you retrieve:

1
tasks = ls.tasks.list(project=project.id, include='id')
2
for task in tasks:
3
    ls.predictions.create(
4
        task=task.id,
5
        result=[{'from_name': 'label', 'to_name': 'text', 'type': 'choices', 'value': {'choices': ['Positive']}}],
6
        score=0.99,
7
        model_version='my-super-ai'
8
    )

For another example, see the Jupyter notebook example of importing pre-annotated data.

The image is specified in the image key using a public URL, and the prediction is referenced in an arbitrary pet key, which is then specified in the preannotated_from_fields() method.

For more examples, see the Jupyter notebook example of importing pre-annotated data.

Managing Labeling Jobs

You can also use the SDK to control how tasks appear in the data manager to annotators or reviewers. You can create custom filters and ordering for the tasks based on parameters that you specify with the SDK. This lets you have more granular control over which tasks in your dataset get labeled or reviewed, and in which order.

Create a batch of tasks to annotate

For example, you can create a filter to prepare tasks to be annotated. For example, if you want annotators to focus on tasks in the first 1000 tasks in a dataset that contain the word “possum” in the field “text” in the task data, do the following:

1
from label_studio_sdk.data_manager import Filters, Column, Type, Operator
2
3
filters = Filters.create(Filters.AND, [
4
    Filters.item(
5
        Column.id,
6
        Operator.GREATER_OR_EQUAL,
7
        Type.Number,
8
        Filters.value(1)
9
    ),
10
        Filters.item(
11
        Column.id,
12
        Operator.LESS_OR_EQUAL,
13
        Type.Number,
14
        Filters.value(1000)
15
    ),
16
    Filters.item(
17
        Column.data("text"),
18
        Operator.CONTAINS,
19
        Type.String,
20
        Filters.value("Hello")
21
    )
22
])

Most often it is very useful to create a view with annotated tasks to review:

1
from label_studio_sdk.data_manager import Filters, Column, Type, Operator
2
3
filters = Filters.create(Filters.AND, [
4
    Filters.item(
5
        Column.completed_at,
6
        Operator.EMPTY,
7
        Type.Boolean,
8
        Filters.value(False)
9
    )
10
])

To create a filtered tasks view, use the following code:

1
view = ls.views.create(
2
    project=project.id,
3
    data={
4
        'title': 'Tasks Sample',
5
        'filters': filters
6
    }
7
)
8
tab = ls.views.get(id=view.id)

If will be displayed in the data manager as tab with Tasks Sample name.

Export Annotations

Run the following code to export annotations from project’s tab you created in the previous step:

1
tasks = ls.tasks.list(view=tab.id, fields='all')
2
for task in tasks:
3
    # You can access annotations in Label Studio JSON format
4
    print(task.annotations)
5
    # And also annotation drafts and predictions
6
    print(task.predictions)
7
    print(task.drafts)

Read more about export formats in the Label Studio SDK documentation.

Handling Errors

If you encounter an error while using the Label Studio Python SDK, you can catch the error and handle it in your script.

1
from label_studio_sdk.core.api_error import ApiError
2
3
try:
4
    for annotated_task in annotated_tasks:
5
        print(annotated_task.annotations)
6
except ApiError as e:
7
    print(e)

Annotations are exported in the format specified in the Label Studio JSON format.