lightly-studio 0.3.3

LightlyStudio is a lightweight, fast, and easy-to-use data exploration tool for data scientists and engineers.

The open-source tool curating datasets

---

🚀 Welcome to LightlyStudio!

We at Lightly created LightlyStudio, an open-source tool designed to supercharge your data curation workflows for computer vision datasets. Explore your data, visualize annotations and crops, tag samples, and export curated lists to improve your machine learning pipelines. And much more!

LightlyStudio runs entirely locally on your machine, keeping your data private. It consists of a Python library for indexing your data and a web-based UI for visualization and curation.

✨ Core Workflow

Using LightlyStudio typically involves these steps:

1. Index Your Dataset: Run a Python script using the lightly_studio library to process your local dataset (images and annotations) and save metadata into a local lightly_studio.db file.
2. Launch the UI: The script then starts a local web server.
3. Explore & Curate: Use the UI to visualize images, annotations, and object crops. Filter and search your data (experimental text search available). Apply tags to interesting samples (e.g., "mislabeled", "review").
4. Export Curated Data: Export information (like filenames) for your tagged samples from the UI to use downstream.
5. Stop the Server: Close the terminal running the script (Ctrl+C) when done.

Visualize your dataset samples with annotations in the grid view.

Switch to the annotation view to inspect individual object crops easily.

Inspect individual samples in detail, viewing all annotations and metadata.

🎯 Features

• Local Web GUI: Explore and curate your dataset in your browser. Works completely offline, your data never leaves your machine.
• Flexible Input Formats: Load your image dataset from a folder, or with annotations from a number of popular formats like e.g. COCO or YOLO.
• Metadata: Attach your custom metadata to every sample.
• Tags: Mark subsets of your dataset for later use.
• Embeddings: Run similarity search queries on your data.
• Selection: Run advanced selection algorithms to tag a subset of your data.

💻 Installation

Ensure you have Python 3.8 or higher. We strongly recommend using a virtual environment.

The library is OS-independent and works on Windows, Linux, and macOS.

# 1. Create and activate a virtual environment (Recommended)
# On Linux/macOS:
python3 -m venv venv
source venv/bin/activate

# On Windows:
python -m venv venv
.\venv\Scripts\activate

# 2. Install LightlyStudio
pip install lightly-studio

Quickstart

Download example datasets by cloning the example repository:

git clone https://github.com/lightly-ai/dataset_examples dataset_examples

YOLO Object Detection

To run an example using a YOLO dataset, create a file named example_yolo.py with the following contents in the same directory that contains the dataset_examples/ folder:

# example_yolo.py

from pathlib import Path
import lightly_studio as ls

# Create a dataset and add the samples from the yolo format
dataset = ls.Dataset.create()
dataset.add_samples_from_yolo(
    data_yaml="dataset_examples/road_signs_yolo/data.yaml",
    input_split="test",
)

# Start the UI application on the port 8001.
ls.start_gui()

Run the script:

python example_yolo.py

When you are done, stop the app by pressing Ctrl+C in the terminal.

The YOLO format details:

road_signs_yolo/
├── train/
│   ├── images/
│   │   ├── image1.jpg
│   │   ├── image2.jpg
│   │   └── ...
│   └── labels/
│       ├── image1.txt
│       ├── image2.txt
│       └── ...
├── valid/  (optional)
│   ├── images/
│   │   └── ...
│   └── labels/
│       └── ...
└── data.yaml

Each label file should contain YOLO format annotations (one per line):

<class> <x_center> <y_center> <width> <height>

Where coordinates are normalized between 0 and 1.

COCO Instance Segmentation

To run an instance segmentation example using a COCO dataset, create a file named example_coco.py with the following contents in the same directory that contains the dataset_examples/ folder:

# example_coco.py

from pathlib import Path
import lightly_studio as ls

# Create a dataset and add the samples from the coco format
dataset = ls.Dataset.create()
dataset.add_samples_from_coco(
    annotations_json="dataset_examples/coco_subset_128_images/instances_train2017.json",
    images_path="dataset_examples/coco_subset_128_images/images",
    annotation_type=ls.AnnotationType.INSTANCE_SEGMENTATION,
)

# Start the UI application on the port 8001.
ls.start_gui()

Run the script:

python example_coco.py

When you are done, stop the app by pressing Ctrl+C in the terminal.

The COCO format details:

coco_subset_128_images/
├── images/
│   ├── image1.jpg
│   ├── image2.jpg
│   └── ...
└── instances_train2017.json        # Single JSON file containing all annotations

COCO uses a single JSON file containing all annotations. The format consists of three main components:

• Images: Defines metadata for each image in the dataset.
• Categories: Defines the object classes.
• Annotations: Defines object instances.

🔍 How It Works

1. Your Python script uses the lightly_studio Dataset.
2. The dataset.add_samples_from_<source> reads your images and annotations, calculates embeddings, and saves metadata to a local lightly_studio.db file (using DuckDB).
3. lightly_studio.start_gui() starts a local Backend API server.
4. This server reads from lightly_studio.db and serves data to the UI Application running in your browser (http://localhost:8001).
5. Images are streamed directly from your disk for display in the UI.

🎯 Python Interface

Dataset

Load Images From A Folder

import lightly_studio as ls

dataset = ls.Dataset.create()
dataset.add_samples_from_path(path="/path/to/image_dataset")

ls.start_gui()

☁️ Cloud Storage Support

Installation with Cloud Storage Support

pip install lightly-studio[cloud-storage]

Example: Loading Dataset from Cloud Storage

import lightly_studio as ls

dataset = ls.Dataset.create()

# Load dataset from S3
dataset.add_samples_from_path(path="s3://my-bucket/path/to/images/")

# You can use glob pattern in the file path
dataset.add_samples_from_path(path="s3://my-bucket/path/to/images/**/*.jpg") # matches all .jpg files recursively

# Load dataset from gcs
dataset.add_samples_from_path(path="gs://path/to/images/")

ls.start_gui()

Note: Currently, cloud storage support is limited to loading images only. Annotation files (YOLO labels, COCO JSON files, etc.) cannot be loaded directly from cloud storage paths.

Authentication

Important: Cloud storage authentication must be configured before running LightlyStudio. The application relies on your existing cloud storage credentials and will not prompt for authentication.

AWS S3

You can use either of the following two options:

• Set environment variables manually: Set AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY (LightlyStudio uses s3fs under the hood to connect to S3)
• Authenticate using AWS CLI: Run aws configure (this will automatically set the environment variables that LightlyStudio can access)

Google Cloud Storage

You can use either of the following two options:

• Set environment variable manually: Set GOOGLE_APPLICATION_CREDENTIALS pointing to your service account key file (LightlyStudio uses gcsfs under the hood to connect to GCS)
• Authenticate using gcloud CLI: Run gcloud auth application-default login (this will automatically set the environment variables that LightlyStudio can access)

Load Images With Annotations

The Dataset currently supports:

• YOLOv8 Object Detection: Reads .yaml file. Supports bounding boxes.
• COCO Object Detection: Reads .json annotations. Supports bounding boxes.
• COCO Instance Segmentation: Reads .json annotations. Supports instance masks in RLE (Run-Length Encoding) format.

# Load a dataset in YOLO format
import lightly_studio as ls

dataset = ls.Dataset.create()
dataset.add_samples_from_yolo(
    data_yaml="my_yolo_dataset/data.yaml",
    input_split="val",
)

ls.start_gui()

# Load an object detection/instance segmentation dataset in COCO format
import lightly_studio as ls

dataset = ls.Dataset.create()
dataset.add_samples_from_coco(
    annotations_json="my_coco_dataset/detections_train.json",
    images_path="my_coco_dataset/images",
    # If using instance segmentation, uncomment the next line.
    # annotation_type=ls.AnnotationType.INSTANCE_SEGMENTATION,
)

ls.start_gui()

Load an Existing Dataset

It is also possible to load an existing dataset by

import lightly_studio as ls

dataset = ls.Dataset.load_or_create()

This will load the dataset if it does exist in the .db file, else it will create a new dataset.

Samples

The dataset consists of samples. Every sample corresponds to an image. Dataset samples can be fetched and accessed as follows, for a full list of attributes see sample.

# Get all dataset samples
samples = list(dataset)

# Access sample attributes
s = samples[0]
s.sample_id        # Sample ID
s.file_name        # Image file name
s.file_path_abs    # Full image file path
s.tags             # The list of sample tags
s.metadata["key"]  # dict-like access for metadata

# Set sample attributes
s.tags = {"tag1", "tag2"}
s.metadata["key"] = 123

# Adding/removing tags
s.add_tag("some_tag")
s.remove_tag("some_tag")

...

Dataset Query

You can efficiently fetch filtered dataset samples with a DatasetQuery() object. To get a query for a existing dataset:

query = dataset.query()

By defining the match, order_by, and slice for a query, the intended filtering is set. If one of them is not required, they can be skipped.

When the query is used to fetch samples, the order of execution is:

1. match
2. order_by
3. slice

Example Query Usage

from lightly_studio.core.dataset_query.boolean_expression import OR
from lightly_studio.core.dataset_query.order_by import OrderByField
from lightly_studio.core.dataset_query.sample_field import SampleField

query = dataset.match(
        OR(
            SampleField.file_name == "a",
            SampleField.file_name == "b",
        )
    ).order_by(
        OrderByField(SampleField.width).desc()
    ).slice(offset=10, limit=10)

query.add_tag("query_result")

Advanced Example:

from lightly_studio.core.dataset_query.boolean_expression import AND, OR, NOT
from lightly_studio.core.dataset_query.order_by import OrderByField
from lightly_studio.core.dataset_query.sample_field import SampleField

query = dataset.match(
    OR(
        SampleField.file_name == "a",
        SampleField.file_name == "b",
        AND(
            SampleField.width > 10,
            SampleField.width < 20,
            NOT(SampleField.tags.contains("dog")),
        ),
    )
).order_by(
    OrderByField(SampleField.width).desc()
).slice(offset=10, limit=10)

query.add_tag("query_result")

for sample in query:
    print(sample.tags)

Define the Query: match

The filtering for a query can be set by:

query = query.match(expression)

To create an expression for filtering on certain sample fields, the SampleField.<field_name> <operator> <value> syntax can be used. Available field names can be seen in SampleField.

SampleField Examples:

from lightly_studio.core.dataset_query.sample_field import SampleField

# Ordinal fields: <, <=, >, >=, ==, !=

expr = SampleField.height >= 10            # All samples with images that are taller than 9 pixels
expr = SampleField.width == 10             # All samples with images that are exactly 10 pixels wide
expr = SampleField.created_at > datetime   # All samples created after datetime (actual datetime object)

# String fields: ==, !=
expr = SampleField.file_name == "some"     # All samples with "some" as file name
expr = SampleField.file_path_abs != "other" # All samples that are not having "other" as file_path

# Tags: contains()
expr = SampleField.tags.contains("dog")    # All samples that contain the tag "dog"

# Assign any of the previous expressions to a query:
query = query.match(expr)

The filtering on individual fields can flexibly be combined to create more complex match expression. For this, the boolean operators AND, OR, and NOT are available. Boolean operators can arbitrarily be nested.

Boolean Examples:

from lightly_studio.core.dataset_query.boolean_expression import AND, OR, NOT
from lightly_studio.core.dataset_query.sample_field import SampleField

# All samples with images that are between 10 and 20 pixels wide
expr = AND(
    SampleField.width > 10,
    SampleField.width < 20
)

# All samples with file names that are either "a" or "b"
expr = OR(
    SampleField.file_name == "a",
    SampleField.file_name == "b"
)

# All samples which do not contain a tag "dog"
expr = NOT(SampleField.tags.contains("dog"))

# All samples for a nested expression
expr = OR(
    SampleField.file_name == "a",
    SampleField.file_name == "b",
    AND(
        SampleField.width > 10,
        SampleField.width < 20,
        NOT(
            SampleField.tags.contains("dog")
        ),
    ),
)

# Assign any of the previous expressions to a query:
query = query.match(expr)

Define the Query: order_by

Setting the sorting of a query can done by

query = query.order_by(expression)

The order expression can be defined by OrderByField(SampleField.<field_name>).<order_direction>().

OrderByField Examples:

from lightly_studio.core.dataset_query.order_by import OrderByField
from lightly_studio.core.dataset_query.sample_field import SampleField

# Sort the query by the width of the image in ascending order
expr = OrderByField(SampleField.width)
expr = OrderByField(SampleField.width).asc()

# Sort the query by the height of the image in descending order
expr = OrderByField(SampleField.file_name).desc()

# Assign any of the previous expressions to a query:
query = query.order_by(expr)

Define the Query: slice

Setting the slicing of a query can done by:

query = query.slice(offset, limit)
# OR
query = query[offset:stop]

Both are different syntax for the same operation.

Slice Examples:

# Slice 2:5
query = query.slice(offset=2, limit=3)
query = query[2:5]

# Slice :5
query = query.slice(limit=5)
query = query[:5]

# Slice 5:
query = query.slice(offset=5)
query = query[5:]

Access the Samples

To access the filtered samples two possibilities are available: iterating over the query object or calling the to_list() method.

Iterating over the query:

query = dataset.query().match(match_expression).order_by(order_by_expression).slice(offset,limit)

samples = []
for sample in query:
    samples.append(sample)

Get all samples as list:

query = dataset.query().match(match_expression).order_by(order_by_expression).slice(offset,limit)

samples = query.to_list()

In some use cases, one might want to assign a tag to the samples that are the result of a query:

query.add_tag("tag_name")

Export Samples

Currently, exporting to the COCO object detection format is supported. The following example shows how to export the samples in the query to a COCO JSON file:

from pathlib import Path

query.export().to_coco_object_detections(Path("coco_export.json"))

Examples

Add Custom Metadata

Attach values to custom fields for every sample.

import lightly_studio as ls

# Load your dataset
dataset = ls.Dataset.create()
dataset.add_samples_from_path(path="/path/to/image_dataset")

# Attach metadata
for sample in dataset:
    sample.metadata["my_metadata"] = f"Example metadata field for {sample.file_name}"
    sample.metadata["my_dict"] = {"my_int_key": 10, "my_bool_key": True}

# View metadata in GUI
ls.start_gui()

Tags

You can easily mark subsets of your data with tags.

import lightly_studio as ls

# Load your dataset
dataset = ls.Dataset.create()
dataset.add_samples_from_path(path="/path/to/image_dataset")

# Tag the first 10 samples:
query = dataset.query()[:10]
query.add_tag("some_tag")

Find existing tags and tagged samples as follows.

import lightly_studio as ls

# Load your dataset
dataset = ls.Dataset.create()
dataset.add_samples_from_path(path="/path/to/image_dataset")

# Get all samples that contain the tag "dog"
query = dataset.query().match(SampleField.tags.contains("dog"))
samples = query.to_list()

Selection

LightlyStudio offers as a premium feature advanced methods for subselecting dataset samples.

Prerequisites: The selection functionality requires a valid LightlyStudio license key. Set the LIGHTLY_STUDIO_LICENSE_KEY environment variable before using selection features:

export LIGHTLY_STUDIO_LICENSE_KEY="license_key_here"

Alternatively, set it inside your Python script:

import os
os.environ["LIGHTLY_STUDIO_LICENSE_KEY"] = "license_key_here"

Diversity Selection

Diversity selection can be configured directly from a DatasetQuery. The example below showcases a simple case of selecting diverse samples.

import lightly_studio as ls

# Load your dataset
dataset = ls.Dataset.load_or_create()
dataset.add_samples_from_path(path="/path/to/image_dataset")

# Select a diverse subset of 10 samples.
dataset.query().selection().diverse(
    n_samples_to_select=10,
    selection_result_tag_name="diverse_selection",
)

ls.start_gui()

Metadata Weighting Selection

You can select samples based on the values of a metadata field. The example below showcases a simple case of selecting samples with the highest metadata value.

import lightly_studio as ls

# Load your dataset
dataset = ls.Dataset.load_or_create()
dataset.add_samples_from_path(path="/path/to/image_dataset")
# Compute and store 'typicality' metadata.
dataset.compute_typicality_metadata(metadata_name="typicality")

# Select the 5 samples with the highest 'typicality' scores.
dataset.query().selection().metadata_weighting(
    n_samples_to_select=5,
    selection_result_tag_name="metadata_weighting_selection",
    metadata_key="typicality",
)

Selection Based on Multiple Strategies

You can configure multiple strategies, the selection takes into account all of them at the same time, weighted by the strength parameter.

import lightly_studio as ls
from lightly_studio.selection.selection_config import (
    MetadataWeightingStrategy,
    EmbeddingDiversityStrategy,
)

# Load your dataset
dataset = ls.Dataset.load_or_create()
dataset.add_samples_from_path(path="/path/to/image_dataset")
# Compute typicality and store it as `typicality` metadata
dataset.compute_typicality_metadata(metadata_name="typicality")

# Select 10 samples by combining typicality and diversity, diversity having double the strength.
dataset.query().selection().multi_strategies(
    n_samples_to_select=10,
    selection_result_tag_name="multi_strategy_selection",
    selection_strategies=[
        MetadataWeightingStrategy(metadata_key="typicality", strength=1.0),
        EmbeddingDiversityStrategy(embedding_model_name="my_model_name", strength=2.0),
    ],
)

Exporting Selected Samples

The selected sample paths can be exported via the GUI, or by a script:

import lightly_studio as ls
from lightly_studio.core.dataset_query.sample_field import SampleField

dataset = ls.Dataset.load("my-dataset")
selected_samples = (
    dataset.match(SampleField.tags.contains("diverse_selection")).to_list()
)

with open("export.txt", "w") as f:
    for sample in selected_samples:
        f.write(f"{sample.file_path_abs}\n")

📚 FAQ

Are the datasets persistent?

Yes, the information about datasets is persistent and stored in the db file. You can see it after the dataset is processed. If you rerun the loader it will create a new dataset representing the same dataset, keeping the previous dataset information untouched.

Can I change the database path?

Not yet. The database is stored in the working directory by default.

Can I launch in another Python script or do I have to do it in the same script?

It is possible to use only one script at the same time because we lock the db file for the duration of the script.

Can I change the API backend port?

Yes. To change the port set the LIGHTLY_STUDIO_PORT variable to your preffered value. If at runtime the port is unavailable it will try to set it to a random value.