Architecture

Basics

The central part of the library is the Dataset class, which represents a dataset and allows to iterate over its elements. DatasetItem , an element of a dataset, represents a single dataset entry with annotations - an image, video sequence, audio track etc. It can contain only annotated data or meta information, only annotations, or all of this.

Basic library usage and data flow:

Extractors -> Dataset -> Converter
                 |
             Filtration
          Transformations
             Statistics
              Merging
             Inference
          Quality Checking
             Comparison
                ...

1. Data is read (or produced) by one or many Extractor s and merged into a Dataset

2. The dataset is processed in some way

3. The dataset is saved with a Converter

Datumaro has a number of dataset and annotation features:

• iteration over dataset elements

• filtering of datasets and annotations by a custom criteria

• working with subsets (e.g. train , val , test )

• computing of dataset statistics

• comparison and merging of datasets

• various annotation operations

from datumaro import Bbox, Polygon, Dataset, DatasetItem

# Import and export a dataset
dataset = Dataset.import_from('src/dir', 'voc')
dataset.export('dst/dir', 'coco')

# Create a dataset, convert polygons to masks, save in PASCAL VOC format
dataset = Dataset.from_iterable([
  DatasetItem(id='image1', annotations=[
    Bbox(x=1, y=2, w=3, h=4, label=1),
    Polygon([1, 2, 3, 2, 4, 4], label=2, attributes={'occluded': True}),
  ]),
], categories=['cat', 'dog', 'person'])
dataset.transform('polygons_to_masks')
dataset.export('dst/dir', 'voc')

The Dataset class

The Dataset class from the datumaro.components.dataset module represents a dataset, consisting of multiple DatasetItem s. Annotations are represented by members of the datumaro.components.extractor module, such as Label , Mask or Polygon. A dataset can contain items from one or multiple subsets (e.g. train , test , val etc.), the list of dataset subsets is available in dataset.subsets().

A DatasetItem is an element of a dataset. Its id is the name of the corresponding image, video frame, or other media being annotated. An item can have some attributes , associated media info and annotations.

Datasets typically have annotations, and these annotations can require additional information to be interpreted correctly. For instance, it can be class names, class hierarchy, keypoint connections, class colors for masks, class attributes. Such information is stored in dataset.categories(), which is a mapping from AnnotationType to a corresponding …Categories class. Each annotation type can have its Categories. Typically, there will be at least LabelCategories ; if there are instance masks, the dataset will contain MaskCategories etc. The “main” type of categories is LabelCategories - annotations and other categories use label indices from this object.

The main operation for a dataset is iteration over its elements ( DatasetItem s). An item corresponds to a single image, a video sequence, etc. There are also many other operations available, such as filtration ( dataset.select() ), transformation (dataset.transform()), exporting ( dataset.export() ) and others. A Dataset is an Iterable and Extractor by itself.

A Dataset can be created from scratch by its class constructor. Categories can be set immediately or later with the define_categories() method, but only once. You can create a dataset filled with initial DatasetItem s with Dataset.from_iterable(). If you need to create a dataset from one or many other extractors (or datasets), it can be done with Dataset.from_extractors().

If a dataset is created from multiple extractors with Dataset.from_extractors() , the source datasets will be joined , so their categories must match. If datasets have mismatching categories, use the more complex IntersectMerge class from datumaro.components.operations , which will merge all the labels and remap the shifted indices in annotations.

A Dataset can be loaded from an existing dataset on disk with Dataset.import_from() (for arbitrary formats) and Dataset.load() (for the Datumaro data format).

By default, Dataset works lazily, which means all the operations requiring iteration over inputs will be deferred as much as possible. If you don’t want such behavior, use the init_cache() method or wrap the code in eager_mode (from datumaro.components.dataset ), which will load all the annotations into memory. The media won’t be loaded unless the data is required, because it can quickly waste all the available memory. You can check if the dataset is cached with the is_cache_initialized attribute.

Once created, a dataset can be modified in batch mode with transforms or directly with the put() and remove() methods. Dataset instances record information about changes done, which can be obtained by get_patch(). The patch information is used automatically on saving and exporting to reduce the amount of disk writes. Changes can be flushed with flush_changes().

from datumaro import Bbox, Label, Polygon, Dataset, DatasetItem

# create a dataset directly from items
dataset1 = Dataset.from_iterable([
    DatasetItem(id='image1', annotations=[
        Bbox(x=1, y=2, w=3, h=4, label=1),
        Polygon([1, 2, 3, 2, 4, 4], label=2),
    ]),
], categories=['cat', 'dog', 'person', 'truck'])

dataset2 = Dataset(categories=dataset1.categories())
dataset2.put(DatasetItem(id='image2', annotations=[
    Label(label=3),
    Bbox(x=2, y=0, w=3, h=1, label=2)
]))

# create a dataset from other datasets
dataset = Dataset.from_extractors(dataset1, dataset2)

# keep only annotated images
dataset.select(lambda item: len(item.annotations) != 0)

# change dataset labels
dataset.transform('remap_labels',
    {
        'cat': 'dog', # rename cat to dog
        'truck': 'car', # rename truck to car
        'person': '', # remove this label
    },
    default='delete')

# iterate over elements
for item in dataset:
    print(item.id, item.annotations)

# iterate over subsets as Datasets
for subset_name, subset in dataset.subsets().items():
    for item in subset:
        print(item.id, item.annotations)

Dataset merging

There are 2 methods of merging datasets in Datumaro:

• simple merging (“joining”)

• complex merging

The simple merging (“joining”)

This approach finds the corresponding DatasetItem s in inputs, finds equal annotations and leaves only the unique set of annotations. This approach requires all the inputs to have categories with the same labels (or no labels) in the same order.

This algorithm is applied automatically in Dataset.from_extractors() and when the build targets are merged in the Project.Tree.make_dataset().

The complex merging

If datasets have mismatching categories, they can’t be merged by the simple approach, because it can lead to errors in the resulting annotations. For complex cases Datumaro provides a more sophisticated algorithm, which finds matching annotations by computing distances between them. Labels and attributes are deduced by voting, spatial annotations use the corresponding metrics like Intersection-over-Union (IoU), OKS, PDJ and others.

The categories of the input datasets are compared, the matching ones complement missing information in each other, the mismatching ones are appended after next. Label indices in annotations are shifted to the new values.

The complex algorithm is available in the IntersectMerge class from datumaro.components.operations. It must be used explicitly. This class also allows to check the inputs and the output dataset for errors and problems.

Projects

Projects are intended for complex use of Datumaro. They provide means of persistence, versioning, high-level operations for datasets and also allow to extend Datumaro via Plugins. A project provides access to build trees and revisions, data sources, models, configuration, plugins and cache. Projects can have multiple data sources, which are joined on dataset creation. Project configuration is available in project.config. To add a data source into a Project , use the import_source() method. The build tree of the current working directory can be converted to a Dataset with project.working_tree.make_dataset().

The Environment class is responsible for accessing built-in and project-specific plugins. For a Project object, there is an instance of related Environment in project.env.

Check the Data Model section of the User Manual: for more info about Project behavior and high-level details.

Library contents

Dataset Formats

The framework provides functions to read and write datasets in specific formats. It is supported by Extractor s, Importer s, and Converter s.

Dataset reading is supported by Extractor s and Importer s:

• An Extractor produces a list of DatasetItem s corresponding to the dataset. Annotations are available in the DatasetItem.annotations list. The SourceExtractor class is designed for loading simple, single-subset datasets. It should be used by default. The Extractor base class should be used when SourceExtractor ‘s functionality is not enough.

• An Importer detects dataset files and generates dataset loading parameters for the corresponding Extractor s. Importer s are optional, they only extend the Extractor functionality and make them more flexible and simple. They are mostly used to locate dataset subsets, but they also can do some data compatibility checks and have other required logic.

It is possible to add custom Extractor s and Importer s. To do this, you need to put an Extractor and Importer implementations to a plugin directory.

Dataset writing is supported by Converter s. A Converter produces a dataset of a specific format from dataset items. It is possible to add custom Converter s. To do this, you need to put a Converter implementation script to a plugin directory.

Dataset Conversions (“Transforms”)

A Transform is a function for altering a dataset and producing a new one. It can update dataset items, annotations, classes, and other properties. A list of available transforms for dataset conversions can be extended by adding a Transform implementation script into a plugin directory.

Model launchers

A list of available launchers for model execution can be extended by adding a Launcher implementation script into a plugin directory.

Plugins

Datumaro comes with a number of built-in formats and other tools, but it also can be extended by plugins. Plugins are optional components, which dependencies are not installed by default. In Datumaro there are several types of plugins, which include:

• Extractor - produces dataset items from data source

• Importer - recognizes dataset type and creates project

• Converter - exports dataset to a specific format

• transformation - modifies dataset items or other properties

• launcher - executes models

A plugin is a regular Python module. It must be present in a plugin directory:

• <project_dir>/.datumaro/plugins for project-specific plugins

• <datumaro_dir>/plugins for global plugins

A plugin can be used either via the Environment class instance, or by regular module importing:

import datumaro as dm
from datumaro.plugins.yolo_format.converter import YoloConverter

# Import a dataset
dataset = dm.Dataset.import_from(src_dir, 'voc')

# Load an existing project, save the dataset in some project-specific format
project = dm.project.Project('project/')
project.env.converters['custom_format'].convert(dataset, save_dir=dst_dir)

# Save the dataset in some built-in format
dm.Environment().converters['yolo'].convert(dataset, save_dir=dst_dir)
YoloConverter.convert(dataset, save_dir=dst_dir)

Using datumaro as a python module

Writing a plugin

A plugin is a Python module with any name, which exports some symbols. Symbols, starting with _ are not exported by default. To export a symbol, inherit it from one of the special classes:

from datumaro import Importer, Extractor, Transform, Launcher, Converter

The exports list of the module can be used to override default behavior:

class MyComponent1: ...
class MyComponent2: ...
exports = [MyComponent2] # exports only MyComponent2

There is also an additional class to modify plugin appearance in command line:

from datumaro import Converter

class MyPlugin(Converter):
    """
    Optional documentation text, which will appear in command-line help
    """

    NAME = 'optional_custom_plugin_name'

    def build_cmdline_parser(self, **kwargs):
        parser = super().build_cmdline_parser(**kwargs)
        # set up argparse.ArgumentParser instance
        # the parsed args are supposed to be used as invocation options
        return parser

Plugin example

datumaro/plugins/
- my_plugin1/file1.py
- my_plugin1/file2.py
- my_plugin2.py

my_plugin1/file2.py contents:

from datumaro import Transform
from .file1 import something, useful

class MyTransform(Transform):
    NAME = "custom_name" # could be generated automatically

    """
    Some description. The text will be displayed in the command line output.
    """

    @classmethod
    def build_cmdline_parser(cls, **kwargs):
        parser = super().build_cmdline_parser(**kwargs)
        parser.add_argument('-q', help="Very useful parameter")
        return parser

    def __init__(self, extractor, q):
        super().__init__(extractor)
        self.q = q

    def transform_item(self, item):
        return item

my_plugin2.py contents:

from datumaro import Converter, Extractor

class MyFormat: ...
class _MyFormatConverter(Converter): ...
class MyFormatExtractor(Extractor): ...

exports = [MyFormat] # explicit exports declaration
# MyFormatExtractor and _MyFormatConverter won't be exported

Command-line

Basically, the interface is divided on contexts and single commands. Contexts are semantically grouped commands, related to a single topic or target. Single commands are handy shorter alternatives for the most used commands and also special commands, which are hard to be put into any specific context. Docker is an example of similar approach.

List of plugins available through the CLI

Model-View-ViewModel (MVVM) UI pattern is used.