This the multi-page printable view of this section. Click here to print.

Return to the regular view of this page.

User Manual

Contents

1 - Installation

Dependencies

  • Python (3.7+)
  • Optional: OpenVINO, TensorFlow, PyTorch, MxNet, Caffe, Accuracy Checker, Git

Installation steps

Optionally, set up a virtual environment:

python -m pip install virtualenv
python -m virtualenv venv
. venv/bin/activate

Install:

# From PyPI:
pip install datumaro[default]

# From the GitHub repository:
pip install 'git+https://github.com/cvat-ai/datumaro[default]'

Read more about choosing between datumaro and datumaro[default] here.

Plugins

Datumaro has many plugins, which are responsible for dataset formats, model launchers and other optional components. If a plugin has dependencies, they can require additional installation. You can find the list of all the plugin dependencies in the plugins section.

Customizing installation

  • Datumaro has the following installation options:

    • pip install datumaro - for core library functionality
    • pip install datumaro[default] - for normal CLI experience

    In restricted installation environments, where some dependencies are not available, or if you need only the core library functionality, you can install Datumaro without extra plugins.

    The CLI variant (datumaro[default]) requires Git to be installed and available to work with Datumaro projects and dataset versioning features. You can find installation instructions for your platform here.

    In some cases, installing just the core library may be not enough, because there can be limited options of installing graphical libraries in the system (various Docker environments, servers etc). You can select between using opencv-python and opencv-python-headless by setting the DATUMARO_HEADLESS environment variable to 0 or 1 before installing the package. It requires installation from sources (using --no-binary):

    DATUMARO_HEADLESS=1 pip install datumaro --no-binary=datumaro

    This option can’t be covered by extras due to Python packaging system limitations.

  • When installing directly from the repository, you can change the installation branch with ...@<branch_name>. Also use --force-reinstall parameter in this case. It can be useful for testing of unreleased versions from GitHub pull requests.

2 - How to use Datumaro

As a standalone tool or a Python module:

datum --help

python -m datumaro --help
python datumaro/ --help
python datum.py --help

As a Python library:

import datumaro as dm
...
dataset = dm.Dataset.import_from(path, format)
...

Glossary

  • Basic concepts:

    • Dataset - A collection of dataset items, which consist of media and associated annotations.
    • Dataset item - A basic single element of the dataset. Also known as “sample”, “entry”. In different datasets it can be an image, a video frame, a whole video, a 3d point cloud etc. Typically, has corresponding annotations.
    • (Datumaro) Project - A combination of multiple datasets, plugins, models and metadata.

  • Project versioning concepts:

    • Data source - A link to a dataset or a copy of a dataset inside a project. Basically, a URL + dataset format name.
    • Project revision - A commit or a reference from Git (branch, tag, HEAD~3 etc.). A revision is referenced by data hash. The HEAD revision is the currently selected revision of the project.
    • Revision tree - A project build tree and plugins at a specified revision.
    • Working tree - The revision tree in the working directory of a project.
    • data source revision - a state of a data source at a specific stage. A revision is referenced by the data hash.
    • Object - The data of a revision tree or a data source revision. An object is referenced by the data hash.

  • Dataset path concepts:

    • Dataset revpath - A path to a dataset in a special format. They are supposed to specify paths to files, directories or data source revisions in a uniform way in the CLI.

      • dataset path - a path to a dataset in the following format: <dataset path>:<format>

        • format is optional. If not specified, will try to detect automatically

      • revision path - a path to a data source revision in a project. The syntax is: <project path>@<revision>:<target name>, any part can be omitted.

        • Default project is the current project (-p/--project CLI arg.) Local revpaths imply that the current project is used and this part should be omitted.
        • Default revision is the working tree of the project
        • Default build target is project

        If a path refers to project (i.e. target name is not set, or this target is exactly specified), the target dataset is the result of joining all the project data sources. Otherwise, if the path refers to a data source revision, the corresponding stage from the revision build tree will be used.

  • Dataset building concepts:

    • Stage - A revision of a dataset - the original dataset or its modification after transformation, filtration or something else. A build tree node. A stage is referred by a name.
    • Build tree - A directed graph (tree) with root nodes at data sources and a single top node called project, which represents a joined dataset. Each data source has a starting root node, which corresponds to the original dataset. The internal graph nodes are stages.
    • Build target - A data source or a stage name. Data source names correspond to the last stages of data sources.
    • Pipeline - A subgraph of a stage, which includes all the ancestors.

  • Other:

    • Transform - A transformation operation over dataset elements. Examples are image renaming, image flipping, image and subset renaming, label remapping etc. Corresponds to the transform command.

Command-line workflow

In Datumaro, most command-line commands operate on projects, but there are also few commands operating on datasets directly. There are 2 basic ways to use Datumaro from the command-line:

  • Use the convert, diff, merge commands directly on existing datasets

  • Create a Datumaro project and operate on it:

Basically, a project is a combination of datasets, models and environment.

A project can contain an arbitrary number of datasets (data sources). A project acts as a manager for them and allows to manipulate them separately or as a whole, in which case it combines dataset items from all the sources into one composite dataset. You can manage separate datasets in a project by commands in the datum source command line context.

Note that modifying operations (transform, filter, patch) are applied in-place to the datasets by default.

If you want to interact with models, you need to add them to the project first using the model add command.

A typical way to obtain Datumaro projects is to export tasks in CVAT UI.

Project data model

project model

Datumaro tries to combine a “Git for datasets” and a build system like make or CMake for datasets in a single solution. Currently, Project represents a Version Control System for datasets, which is based on Git and DVC projects. Each project Revision describes a build tree of a dataset with all the related metadata. A build tree consists of a number of data sources and transformation stages. Each data source has its own set of build steps (stages). Datumaro supposes copying of datasets and working in-place by default. Modifying operations are recorded in the project, so any of the dataset revisions can be reproduced when needed. Multiple dataset versions can be stored in different branches with the common data shared.

Let’s consider an example of a build tree: build tree There are 2 data sources in the example project. The resulting dataset is obtained by simple merging (joining) the results of the input datasets. “Source 1” and “Source 2” are the names of data sources in the project. Each source has several stages with their own names. The first stage (called “root”) represents the original contents of a data source - the data at the user-provided URL. The following stages represent operations, which needs to be done with the data source to prepare the resulting dataset.

Roughly, such build tree can be created by the following commands (arguments are omitted for simplicity):

datum create

# describe the first source
datum import <...> -n source1
datum filter <...> source1
datum transform <...> source1
datum transform <...> source1

# describe the second source
datum import <...> -n source2
datum model add <...>
datum transform <...> source2
datum transform <...> source2

Now, the resulting dataset can be built with:

datum export <...>

Project layout

project/
├── .dvc/
├── .dvcignore
├── .git/
├── .gitignore
├── .datumaro/
│   ├── cache/ # object cache
│   │   └── <2 leading symbols of obj hash>/
│   │       └── <remaining symbols of obj hash>/
│   │           └── <object data>
│   │
│   ├── models/ # project-specific models
│   │
│   ├── plugins/ # project-specific plugins
│   │   ├── plugin1/ # composite plugin, a directory
│   │   |   ├── __init__.py
│   │   |   └── file2.py
│   │   ├── plugin2.py # simple plugin, a file
│   │   └── ...
│   │
│   ├── tmp/ # temp files
│   └── tree/ # working tree metadata
│       ├── config.yml
│       └── sources/
│           ├── <source name 1>.dvc
│           ├── <source name 2>.dvc
│           └── ...
│
├── <source name 1>/ # working directory for the source 1
│   └── <source data>
└── <source name 2>/ # working directory for the source 2
    └── <source data>

Datasets and Data Sources

A project can contain an arbitrary number of Data Sources. Each Data Source describes a dataset in a specific format. A project acts as a manager for the data sources and allows to manipulate them separately or as a whole, in which case it combines dataset items from all the sources into one composite dataset. You can manage separate sources in a project by commands in the datum source command line context.

Datasets come in a wide variety of formats. Each dataset format defines its own data structure and rules on how to interpret the data. For example, the following data structure is used in COCO format:

/dataset/
- /images/<id>.jpg
- /annotations/

Datumaro supports complete datasets, having both image data and annotations, or incomplete ones, having annotations only. Incomplete datasets can be used to prepare images and annotations independently of each other, or to analyze or modify just the lightweight annotations without the need to download the whole dataset.

Check supported formats for more info about format specifications, supported import and export options and other details. The list of formats can be extended by custom plugins, check extending tips for information on this topic.

Use cases

Let’s consider few examples describing what Datumaro does for you behind the scene.

The first example explains how working trees, working directories and the cache interact. Suppose, there is a dataset which we want to modify and export in some other format. To do it with Datumaro, we need to create a project and register the dataset as a data source:

datum create
datum import <...> -n source1

The dataset will be copied to the working directory inside the project. It will be added to the project working tree.

After the dataset is added, we want to transform it and filter out some irrelevant samples, so we run the following commands:

datum transform <...> source1
datum filter <...> source1

The commands modify the data source inside the working directory, inplace. The operations done are recorded in the working tree.

Now, we want to make a new version of the dataset and make a snapshot in the project cache. So we commit the working tree:

datum commit <...>

cache interaction diagram 1

At this time, the data source is copied into the project cache and a new project revision is created. The dataset operation history is saved, so the dataset can be reproduced even if it is removed from the cache and the working directory. Note, however, that the original dataset hash was not computed, so Datumaro won’t be able to compare dataset hash on re-downloading. If it is desired, consider making a commit with an unmodified data source.

After this, we do some other modifications to the dataset and make a new commit. Note that the dataset is not cached, until a commit is done.

When the dataset is ready and all the required operations are done, we can export it to the required format. We can export the resulting dataset, or any previous stage.

datum export <...> source1
datum export <...> source1.stage3

Let’s extend the example. Imagine we have a project with 2 data sources. Roughly, it corresponds to the following set of commands:

datum create
datum import <...> -n source1
datum import <...> -n source2
datum transform <...> source1 # used 3 times
datum transform <...> source2 # used 5 times

Then, for some reasons, the project cache was cleaned from source1 revisions. We also don’t have anything in the project working directories - suppose, the user removed them to save disk space.

Let’s see what happens, if we call the diff command with 2 different revisions now.

cache interaction diagram 2

Datumaro needs to reproduce 2 dataset revisions requested so that they could be read and compared. Let’s see how the first dataset is reproduced step-by-step:

  1. source1.stage2 will be looked for in the project cache. It won’t be found, since the cache was cleaned.
  2. Then, Datumaro will look for previous source revisions in the cache and won’t find any.
  3. The project can be marked read-only, if we are not working with the “current” project (which is specified by the -p/--project command parameter). In the example, the command is datum diff rev1:... rev2:..., which means there is a project in the current directory, so the project we are working with is not read-only. If a command target was specified as datum diff <project>@<rev>:<source>, the project would be loaded as read-only. If a project is read-only, we can’t do anything more to reproduce the dataset and can only exit with an error (3a). The reason for such behavior is that the dataset downloading can be quite expensive (in terms of time, disk space etc.). It is supposed, that such side-effects should be controlled manually.
  4. If the project is not read-only (3b), Datumaro will try to download the original dataset and reproduce the resulting dataset. The data hash will be computed and hashes will be compared (if the data source had hash computed on addition). On success, the data will be put into the cache.
  5. The downloaded dataset will be read and the remaining operations from the source history will be re-applied.
  6. The resulting dataset might be cached in some cases.
  7. The resulting dataset is returned.

The source2 will be looked for the same way. In our case, it will be found in the cache and returned. Once both datasets are restored and read, they are compared.

Consider other situation. Let’s try to export the source1. Suppose we have a clear project cache and the source1 has a copy in the working directory.

cache interaction diagram 3

Again, Datumaro needs to reproduce a dataset revision (stage) requested.

  1. It looks for the dataset in the working directory and finds some data. If there is no source working directory, Datumaro will try to reproduce the source using the approach described above (1b).
  2. The data hash is computed and compared with the one saved in the history. If the hashes match, the dataset is read and returned (4). Note: we can’t use the cached hash stored in the working tree info - it can be outdated, so we need to compute it again.
  3. Otherwise, Datumaro tries to detect the stage by the data hash. If the current stage is not cached, the tree is the working tree and the working directory is not empty, the working copy is hashed and matched against the source stage list. If there is a matching stage, it will be read and the missing stages will be added. The result might be cached in some cases. If there is no matching stage in the source history, the situation can be contradictory. Currently, an error is raised (3b).
  4. The resulting dataset is returned.

After the requested dataset is obtained, it is exported in the requested format.

To sum up, Datumaro tries to restore a dataset from the project cache or reproduce it from sources. It can be done as long as the source operations are recorded and any step data is available. Note that cache objects share common files, so if there are only annotation differences between datasets, or data sources contain the same images, there will only be a single copy of the related media files. This helps to keep storage use reasonable and avoid unnecessary data copies.

Examples

Example: create a project, add dataset, modify, restore an old version

datum create
datum import <path/to/dataset> -f coco -n source1
datum commit -m "Added a dataset"
datum transform -t shapes_to_boxes
datum filter -e '/item/annotation[label="cat" or label="dog"]' -m i+a
datum commit -m "Transformed"
datum checkout HEAD~1 -- source1 # restore a previous revision
datum status # prints "modified source1"
datum checkout source1 # restore the last revision
datum export -f voc -- --save-images

3 - Supported Formats

List of supported formats:

Supported annotation types

  • Labels
  • Bounding boxes
  • Polygons
  • Polylines
  • (Segmentation) Masks
  • (Key-)Points
  • Captions
  • 3D cuboids
  • Super Resolution Annotation
  • Depth Annotation

Datumaro does not separate datasets by tasks like classification, detection etc. Instead, datasets can have any annotations. When a dataset is exported in a specific format, only relevant annotations are exported.

Dataset meta info file

It is possible to use classes that are not original to the format. To do this, use dataset_meta.json.

{
"label_map": {"0": "background", "1": "car", "2": "person"},
"segmentation_colors": [[0, 0, 0], [255, 0, 0], [0, 0, 255]],
"background_label": "0"
}
  • label_map is a dictionary where the class ID is the key and the class name is the value.
  • segmentation_colors is a list of channel-wise values for each class. This is only necessary for the segmentation task.
  • background_label is a background label ID in the dataset.

4 - Media formats

Datumaro supports the following media types:

  • 2D RGB(A) images
  • KITTI Point Clouds

To create an unlabelled dataset from an arbitrary directory with images use image_dir and image_zip formats:

datum create -o <project/dir>
datum import -p <project/dir> -f image_dir <directory/path/>

or, if you work with Datumaro API:

  • for using with a project:

    from datumaro.project import Project

project = Project.init()
project.import_source('source1', format='image_dir', url='directory/path/')
dataset = project.working_tree.make_dataset()

  • for using as a dataset:

    from datumaro import Dataset

dataset = Dataset.import_from('directory/path/', 'image_dir')

This will search for images in the directory recursively and add them as dataset entries with names like <subdir1>/<subsubdir1>/<image_name1>. The list of formats matches the list of supported image formats in OpenCV:

.jpg, .jpeg, .jpe, .jp2, .png, .bmp, .dib, .tif, .tiff, .tga, .webp, .pfm,
.sr, .ras, .exr, .hdr, .pic, .pbm, .pgm, .ppm, .pxm, .pnm

Once there is a Dataset instance, its items can be split into subsets, renamed, filtered, joined with annotations, exported in various formats etc.

To import frames from a video, you can split the video into frames with the split_video command and then use the image_dir format described above. In more complex cases, consider using FFmpeg and other tools for video processing.

Alternatively, you can use the video_frames format directly:

Note, however, that it can produce different results if the system environment changes. If you want to obtain reproducible results, consider splitting the video into frames by any method.

datum create -o <project/dir>
datum import -p <project/dir> -f video_frames <video/path.avi>

from datumaro import Dataset

dataset = Dataset.import_from('video.mp4', 'video_frames')

Datumaro supports the following video formats:

.3gp, .3g2, .asf, .wmv, .avi, .divx, .evo, .f4v, .flv, .mkv, .mk3d,
.mp4, .mpg, .mpeg, .m2p, .ps, .ts, .m2ts, .mxf, .ogg, .ogv, .ogx,
.mov, .qt, .rmvb, .vob, .webm

5 - Command reference

%%{init { 'theme':'neutral' }}%%
flowchart LR
  d(("#0009; datum #0009;")):::mainclass
  m(model):::nofillclass
  p(project):::nofillclass
  s(source):::nofillclass

  d===m
    m===m_add[add]:::hideclass
    m===m_info[info]:::hideclass
    m===m_remove[remove]:::hideclass
    m===m_run[run]:::hideclass
  d===p
    p===p_info[info]:::hideclass
    p===p_migrate[migrate]:::hideclass
  d===s
    s===s_add[add]:::hideclass
    s===s_info[info]:::hideclass
    s===s_remove[remove]:::hideclass
  d====_add[add]:::filloneclass
  d====_create[create]:::filloneclass
  d====_describe_downloads[describe-downloads]:::filloneclass
  d====_detect_format[detect-format]:::filloneclass
  d====_download[download]:::filloneclass
  d====_export[export]:::filloneclass
  d====_import[import]:::filloneclass
  d====_info[info]:::filloneclass
  d====_remove[remove]:::filloneclass
  d====_generate[generate]:::filloneclass
  d====_filter[filter]:::filltwoclass
  d====_transform[transform]:::filltwoclass
  d====_diff[diff]:::fillthreeclass
  d====_explain[explain]:::fillthreeclass
  d====_merge[merge]:::fillthreeclass
  d====_patch[patch]:::fillthreeclass
  d====_stats[stats]:::fillthreeclass
  d====_validate[validate]:::fillthreeclass
  d====_checkout[checkout]:::fillfourclass
  d====_commit[commit]:::fillfourclass
  d====_log[log]:::fillfourclass
  d====_status[status]:::fillfourclass

  classDef nofillclass fill-opacity:0;
  classDef hideclass fill-opacity:0,stroke-opacity:0;
  classDef filloneclass fill:#CCCCFF,stroke-opacity:0;
  classDef filltwoclass fill:#FFFF99,stroke-opacity:0;
  classDef fillthreeclass fill:#CCFFFF,stroke-opacity:0;
  classDef fillfourclass fill:#CCFFCC,stroke-opacity:0;

The command line is split into the separate commands and command contexts. Contexts group multiple commands related to a specific topic, e.g. project operations, data source operations etc. Almost all the commands operate on projects, so the project context and commands without a context are mostly the same. By default, commands look for a project in the current directory. If the project you’re working on is located somewhere else, you can pass the -p/--project <path> argument to the command.

Note: command behavior is subject to change, so this text might be outdated, always check the --help output of the specific command

Note: command parameters must be passed prior to the positional arguments.

Datumaro functionality is available with the datum command.

Usage:

datum [-h] [--version] [--loglevel LOGLEVEL] [command] [command args]

Parameters:

  • --loglevel (string) - Logging level, one of debug, info, warning, error, critical (default: info)
  • --version - Print the version number and exit.
  • -h, --help - Print the help message and exit.

5.1 - Checkout

This command allows to restore a specific project revision in the project tree or to restore separate revisions of sources. A revision can be a commit hash, branch, tag, or any relative reference in the Git format.

This command has multiple forms:

1) datum checkout <revision>
2) datum checkout [--] <source1> ...
3) datum checkout <revision> [--] <source1> <source2> ...

1 - Restores a revision and all the corresponding sources in the working directory. If there are conflicts between modified files in the working directory and the target revision, an error is raised, unless --force is used.

2, 3 - Restores only selected sources from the specified revision. The current revision is used, when not set.

“–” can be used to separate source names and revisions:

  • datum checkout name - will look for revision “name”
  • datum checkout -- name - will look for source “name” in the current revision

Usage:

datum checkout [-h] [-f] [-p PROJECT_DIR] [rev] [--] [sources [sources ...]]

Parameters:

  • --force - Allows to overwrite unsaved changes in case of conflicts
  • -p, --project (string) - Directory of the project to operate on (default: current directory).
  • -h, --help - Print the help message and exit.

Examples:

  • Restore the previous revision: datum checkout HEAD~1

  • Restore the saved version of a source in the working tree datum checkout -- source-1

  • Restore a previous version of a source datum checkout 33fbfbe my-source

5.2 - Commit

This command allows to fix the current state of a project and create a new revision from the working tree.

By default, this command checks sources in the working tree for changes. If there are unknown changes found, an error will be raised, unless --allow-foreign is used. If such changes are committed, the source will only be available for reproduction from the project cache, because Datumaro will not know how to repeat them.

The command will add the sources into the project cache. If you only need to record revision metadata, you can use the --no-cache parameter. This can be useful if you want to save disk space and/or have a backup copy of datasets used in the project.

If there are no changes found, the command will stop. To allow empty commits, use --allow-empty.

Usage:

datum commit [-h] -m MESSAGE [--allow-empty] [--allow-foreign]
  [--no-cache] [-p PROJECT_DIR]

Parameters:

  • --allow-empty - Allow commits with no changes
  • --allow-foreign - Allow commits with changes made not by Datumaro
  • --no-cache - Don’t put committed datasets into cache, save only metadata
  • -p, --project (string) - Directory of the project to operate on (default: current directory).
  • -h, --help - Print the help message and exit.

Example:

datum create
datum import -f coco <path/to/coco/>
datum commit -m "Added COCO"

5.3 - Convert datasets

This command allows to convert a dataset from one format to another. The command is a usability alias for create, add and export and just provides a simpler way to obtain the same results in simple cases. A list of supported formats can be found in the --help output of this command.

Usage:

datum convert [-h] [-i SOURCE] [-if INPUT_FORMAT] -f OUTPUT_FORMAT
  [-o DST_DIR] [--overwrite] [-e FILTER] [--filter-mode FILTER_MODE]
  [-- EXTRA_EXPORT_ARGS]

Parameters:

  • -i, --input-path (string) - Input dataset path. The current directory is used by default.
  • -if, --input-format (string) - Input dataset format. Will try to detect, if not specified.
  • -f, --output-format (string) - Output format
  • -o, --output-dir (string) - Output directory. By default, a subdirectory in the current directory is used.
  • --overwrite - Allows overwriting existing files in the output directory, when it is not empty.
  • -e, --filter (string) - XML XPath filter expression for dataset items
  • --filter-mode (string) - The filtering mode. Default is the i mode.
  • -p, --project (string) - Directory of the project to operate on (default: current directory).
  • -h, --help - Print the help message and exit.
  • -- <extra export args> - Additional arguments for the format writer (use -- -h for help). Must be specified after the main command arguments.

Example: convert a VOC-like dataset to a COCO-like one:

datum convert --input-format voc --input-path <path/to/voc/> \
              --output-format coco \
              -- --save-images

5.4 - Create project

The command creates an empty project. A project is required for the most of Datumaro functionality.

By default, the project is created in the current directory. To specify another output directory, pass the -o/--output-dir parameter. If output already directory contains a Datumaro project, an error is raised, unless --overwrite is used.

Usage:

datum create [-h] [-o DST_DIR] [--overwrite]

Parameters:

  • -o, --output-dir (string) - Allows to specify an output directory. The current directory is used by default.
  • --overwrite - Allows to overwrite existing project files in the output directory. Any other files are not touched.
  • -h, --help - Print the help message and exit.

Examples:

Example: create an empty project in the my_dataset directory

datum create -o my_dataset/

Example: create a new empty project in the current directory, remove the existing one

datum create
...
datum create --overwrite

5.5 - Describe downloadable datasets

This command reports reports various information about datasets that can be downloaded with the download command. The information is reported either as human-readable text (the default) or as a JSON object. The format can be selected with the --report-format option.

When the JSON output format is selected, the output document has the following schema:

{
    "<dataset name>": {
        "default_output_format": "<Datumaro format name>",
        "description": "<human-readable description>",
        "download_size": <total size of the downloaded files in bytes>,
        "home_url": "<URL of a web page describing the dataset>",
        "human_name": "<human-readable dataset name>",
        "num_classes": <number of classes in the dataset>,
        "subsets": {
            "<subset name>": {
                "num_items": <number of items in the subset>
            },
            ...
        },
        "version": "<version number>"
    },
    ...
}

home_url may be null if there is no suitable web page for the dataset.

num_classes may be null if the dataset does not involve classification.

version currently contains the version number supplied by TFDS. In future versions of Datumaro, datasets might come from other sources; the way version numbers will be set for those is to be determined.

New object members may be added in future versions of Datumaro.

Usage:

datum describe-downloads [-h] [--report-format {text,json}]
                         [--report-file REPORT_FILE]

Parameters:

  • -h, --help - Print the help message and exit.
  • --report-format (text or json) - Format in which to report the information. By default, text is used.
  • --report-file (string) - File to which to write the report. By default, the report is written to the standard output stream.

5.6 - Detect dataset format

This command attempts to detect the format of a dataset in a directory. Currently, only local directories are supported.

The detection result may be one of:

  • a single format being detected;
  • no formats being detected (if the dataset doesn’t match any known format);
  • multiple formats being detected (if the dataset is ambiguous).

The command outputs this result in a human-readable form and optionally as a machine-readable JSON report (see --json-report).

The format of the machine-readable report is as follows:

{
    "detected_formats": [
        "detected-format-name-1", "detected-format-name-2", ...
    ],
    "rejected_formats": {
        "rejected-format-name-1": {
            "reason": <reason-code>,
            "message": "line 1\nline 2\n...\nline N"
        },
        "rejected-format-name-2": ...,
        ...
    }
}

The <reason-code> can be one of:

  • "detection_unsupported": the corresponding format does not support detection.

  • "insufficient_confidence": the dataset matched the corresponding format, but it matched at least one other format better.

  • "unmet_requirements": the dataset didn’t meet at least one requirement of the corresponding format.

Other reason codes may be defined in the future.

Usage:

datum detect-format [-h] [-p PROJECT_DIR] [--show-rejections]
                    [--json-report JSON_REPORT]
                    url

Parameters:

  • <url> - Path to the dataset to analyse.
  • -h, --help - Print the help message and exit.
  • -p, --project (string) - Directory of the project to use as the context (default: current directory). The project might contain local plugins with custom formats, which will be used for detection.
  • --show-rejections - Describe why each supported format that wasn’t detected was rejected. This only affects the human-readable output; the machine-readable report always includes rejection information.
  • --json-report (string) - Path to which to save a JSON report describing detected and rejected formats. By default, no report is saved.

Example: detect the format of a dataset in a given directory, showing rejection information:

datum detect-format --show-rejections path/to/dataset

5.7 - Compare datasets

The command compares two datasets and saves the results in the specified directory. The current project is considered to be “ground truth”.

Datasets can be compared using different methods:

  • equality - Annotations are compared to be equal
  • distance - A distance metric is used

This command has multiple forms:

1) datum diff <revpath>
2) datum diff <revpath> <revpath>

1 - Compares the current project’s main target (project) in the working tree with the specified dataset.

2 - Compares two specified datasets.

<revpath> - a dataset path or a revision path.

Usage:

datum diff [-h] [-o DST_DIR] [-m METHOD] [--overwrite] [-p PROJECT_DIR]
  [--iou-thresh IOU_THRESH] [-f FORMAT]
  [-iia IGNORE_ITEM_ATTR] [-ia IGNORE_ATTR] [-if IGNORE_FIELD]
  [--match-images] [--all]
  first_target [second_target]

Parameters:

  • <target> (string) - Target dataset revpaths

  • -m, --method (string) - Comparison method.

  • -o, --output-dir (string) - Output directory. By default, a new directory is created in the current directory.

  • --overwrite - Allows to overwrite existing files in the output directory, when it is specified and is not empty.

  • -p, --project (string) - Directory of the project to operate on (default: current directory).

  • -h, --help - Print the help message and exit.

  • Distance comparison options:

    • --iou-thresh (number) - The IoU threshold for spatial annotations (default is 0.5).
    • -f, --format (string) - Output format, one of simple (text files and images) and tensorboard (a TB log directory)

  • Equality comparison options:

    • -iia, --ignore-item-attr (string) - Ignore an item attribute (repeatable)
    • -ia, --ignore-attr (string) - Ignore an annotation attribute (repeatable)
    • -if, --ignore-field (string) - Ignore an annotation field (repeatable) Default is id and group
    • --match-images - Match dataset items by image pixels instead of ids
    • --all - Include matches in the output. By default, only differences are printed.

Examples:

  • Compare two projects by distance, match boxes if IoU > 0.7, save results to TensorBoard: datum diff other/project -o diff/ -f tensorboard --iou-thresh 0.7

  • Compare two projects for equality, exclude annotation groups and the is_crowd attribute from comparison: datum diff other/project/ -if group -ia is_crowd

  • Compare two datasets, specify formats: datum diff path/to/dataset1:voc path/to/dataset2:coco

  • Compare the current working tree and a dataset: datum diff path/to/dataset2:coco

  • Compare a source from a previous revision and a dataset: datum diff HEAD~2:source-2 path/to/dataset2:yolo

  • Compare a dataset with model inference

datum create
datum import <...>
datum model add mymodel <...>
datum transform <...> -o inference
datum diff inference -o diff

5.8 - Download datasets

This command downloads a publicly available dataset and saves it to a local directory. In terms of syntax, this command is similar to convert, but instead of taking a local directory as the source, it takes a dataset ID. A list of supported datasets and output formats can be found in the --help output of this command.

Currently, the only source of datasets is the TensorFlow Datasets library. Therefore, to use this command you must install TensorFlow & TFDS, which you can do as follows:

pip install datumaro[tf,tfds]

To use a proxy for downloading, configure it with the conventional curl environment variables.

Usage:

datum download [-h] -i DATASET_ID [-f OUTPUT_FORMAT] [-o DST_DIR]
               [--overwrite] [-s SUBSET] [-- EXTRA_EXPORT_ARGS]

Parameters:

  • -h, --help - Print the help message and exit.
  • -i, --dataset-id (string) - ID of the dataset to download.
  • -f, --output-format (string) - Output format. By default, the format of the original dataset is used.
  • -o, --output-dir (string) - Output directory. By default, a subdirectory in the current directory is used.
  • --overwrite - Allows overwriting existing files in the output directory, when it is not empty.
  • --subset (string) - Which subset of the dataset to save. By default, all subsets are saved. Note that due to limitations of TFDS, all subsets are downloaded even if this option is specified.
  • -- <extra export args> - Additional arguments for the format writer (use -- -h for help). Must be specified after the main command arguments.

Example: download the MNIST dataset, saving it in the ImageNet text format:

datum download -i tfds:mnist -f imagenet_txt -- --save-images

5.9 - Run model inference explanation (explain)

Runs an explainable AI algorithm for a model.

This tool is supposed to help an AI developer to debug a model and a dataset. Basically, it executes model inference and tries to find relation between inputs and outputs of the trained model, i.e. determine decision boundaries and belief intervals for the classifier.

Currently, the only available algorithm is RISE (article), which runs model a single time and then re-runs a model multiple times on each image to produce a heatmap of activations for each output of the first inference. Each time a part of the input image is masked. As a result, we obtain a number heatmaps, which show, how specific image pixels affected the inference result. This algorithm doesn’t require any special information about the model, but it requires the model to return all the outputs and confidences. The original algorithm supports only classification scenario, but Datumaro extends it for detection models.

The following use cases available:

  • RISE for classification
  • RISE for object detection

Usage:

datum explain [-h] -m MODEL [-o SAVE_DIR] [-p PROJECT_DIR]
  [target] {rise} [RISE_ARGS]

Parameters:

  • <target> (string) - Target dataset revpath.By default, uses the whole current project. An image path can be specified instead. <image path> - a path to the file. <revpath> - a dataset path or a revision path.

  • <method> (string) - The algorithm to use. Currently, only rise is supported.

  • -m, --model (string) - The model to use for inference

  • -o, --output-dir (string) - Directory to save results to (default: display only)

  • -p, --project (string) - Directory of the project to operate on (default: current directory).

  • -h, --help - Print the help message and exit.

  • RISE options:

    • -s, --max-samples (number) - Number of algorithm model runs per image (default: mask size ^ 2).
    • --mw, --mask-width (number) - Mask width in pixels (default: 7)
    • --mh, --mask-height (number) - Mask height in pixels (default: 7)
    • --prob (number) - Mask pixel inclusion probability, controls mask density (default: 0.5)
    • --iou, --iou-thresh (number) - IoU match threshold for detections (default: 0.9)
    • --nms, --nms-iou-thresh (number) - IoU match threshold for detections for non-maxima suppression (default: no NMS)
    • --conf, --det-conf-thresh (number) - Confidence threshold for detections (default: include all)
    • -b, --batch-size (number) - Batch size for inference (default: 1)
    • --display - Visualize results during computations

Examples:

  • Run RISE on an image, display results: datum explain path/to/image.jpg -m mymodel rise --max-samples 50

  • Run RISE on a source revision: datum explain HEAD~1:source-1 -m model rise

  • Run inference explanation on a single image with online visualization

datum create <...>
datum model add mymodel <...>
datum explain -t image.png -m mymodel \
    rise --max-samples 1000 --display

Note: this algorithm requires the model to return all (or a reasonable amount) the outputs and confidences unfiltered, i.e. all the Label annotations for classification models and all the Bboxes for detection models. You can find examples of the expected model outputs in tests/test_RISE.py

For OpenVINO models the output processing script would look like this:

Classification scenario:

import datumaro as dm
from datumaro.util.annotation_util import softmax

def process_outputs(inputs, outputs):
    # inputs = model input, array or images, shape = (N, C, H, W)
    # outputs = model output, logits, shape = (N, n_classes)
    # results = conversion result, [ [ Annotation, ... ], ... ]
    results = []
    for output in outputs:
        confs = softmax(output[0])
        for label, conf in enumerate(confs):
            results.append(dm.Label(int(label)), attributes={'score': float(conf)})

    return results

Object Detection scenario:

import datumaro as dm

# return a significant number of output boxes to make multiple runs
# statistically correct and meaningful
max_det = 1000

def process_outputs(inputs, outputs):
    # inputs = model input, array or images, shape = (N, C, H, W)
    # outputs = model output, shape = (N, 1, K, 7)
    # results = conversion result, [ [ Annotation, ... ], ... ]
    results = []
    for input, output in zip(inputs, outputs):
        input_height, input_width = input.shape[:2]
        detections = output[0]
        image_results = []
        for det in detections:
            label = int(det[1])
            conf = float(det[2])
            x = max(int(det[3] * input_width), 0)
            y = max(int(det[4] * input_height), 0)
            w = min(int(det[5] * input_width - x), input_width)
            h = min(int(det[6] * input_height - y), input_height)
            image_results.append(dm.Bbox(x, y, w, h,
                label=label, attributes={'score': conf} ))

            results.append(image_results[:max_det])

    return results

5.10 - Export Datasets

This command exports a project or a source as a dataset in some format.

Check supported formats for more info about format specifications, supported options and other details. The list of formats can be extended by custom plugins, check extending tips for information on this topic.

Available formats are listed in the command help output.

Dataset format writers support additional export options. To pass such options, use the -- separator after the main command arguments. The usage information can be printed with datum import -f <format> -- --help.

Common export options:

  • Most formats (where applicable) support the --save-images option, which allows to export dataset images along with annotations. The option is disabled be default.
  • If --save-images is used, the image-ext option can be passed to specify the output image file extension (.jpg, .png etc.). By default, tries to Datumaro keep the original image extension. This option allows to convert all the images from one format into another.

This command allows to use the -f/--filter parameter to select dataset elements needed for exporting. Read the filter command description for more info about this functionality.

The command can only be applied to a project build target, a stage or the combined project target, in which case all the targets will be affected.

Usage:

datum export [-h] [-e FILTER] [--filter-mode FILTER_MODE] [-o DST_DIR]
  [--overwrite] [-p PROJECT_DIR] -f FORMAT [target] [-- EXTRA_FORMAT_ARGS]

Parameters:

  • <target> (string) - A project build target to be exported. By default, all project targets are affected.
  • -f, --format (string) - Output format.
  • -e, --filter (string) - XML XPath filter expression for dataset items
  • --filter-mode (string) - The filtering mode. Default is the i mode.
  • -o, --output-dir (string) - Output directory. By default, a subdirectory in the current directory is used.
  • --overwrite - Allows overwriting existing files in the output directory, when it is not empty.
  • -p, --project (string) - Directory of the project to operate on (default: current directory).
  • -h, --help - Print the help message and exit.
  • -- <extra format args> - Additional arguments for the format writer (use -- -h for help). Must be specified after the main command arguments.

Example: save a project as a VOC-like dataset, include images, convert images to PNG from other formats.

datum export \
  -p test_project \
  -o test_project-export \
  -f voc \
  -- --save-images --image-ext='.png'

5.11 - Filter datasets

This command allows to extract a sub-dataset from a dataset. The new dataset includes only items satisfying some condition. The XML XPath is used as a query format.

The command can be applied to a dataset or a project build target, a stage or the combined project target, in which case all the project targets will be affected. A build tree stage will be recorded if --stage is enabled, and the resulting dataset(-s) will be saved if --apply is enabled.

By default, datasets are updated in-place. The -o/--output-dir option can be used to specify another output directory. When updating in-place, use the --overwrite parameter (in-place updates fail by default to prevent data loss), unless a project target is modified.

The current project (-p/--project) is also used as a context for plugins, so it can be useful for dataset paths having custom formats. When not specified, the current project’s working tree is used.

There are several filtering modes available (the -m/--mode parameter). Supported modes:

  • i, items
  • a, annotations
  • i+a, a+i, items+annotations, annotations+items

When filtering annotations, use the items+annotations mode to point that annotation-less dataset items should be removed, otherwise they will be kept in the resulting dataset. To select an annotation, write an XPath that returns annotation elements (see examples).

Item representations can be printed with the --dry-run parameter:

<item>
  <id>290768</id>
  <subset>minival2014</subset>
  <image>
    <width>612</width>
    <height>612</height>
    <depth>3</depth>
  </image>
  <annotation>
    <id>80154</id>
    <type>bbox</type>
    <label_id>39</label_id>
    <x>264.59</x>
    <y>150.25</y>
    <w>11.19</w>
    <h>42.31</h>
    <area>473.87</area>
  </annotation>
  <annotation>
    <id>669839</id>
    <type>bbox</type>
    <label_id>41</label_id>
    <x>163.58</x>
    <y>191.75</y>
    <w>76.98</w>
    <h>73.63</h>
    <area>5668.77</area>
  </annotation>
  ...
</item>

The command can only be applied to a project build target, a stage or the combined project target, in which case all the targets will be affected. A build tree stage will be added if --stage is enabled, and the resulting dataset(-s) will be saved if --apply is enabled.

Usage:

datum filter [-h] [-e FILTER] [-m MODE] [--dry-run] [--stage STAGE]
  [--apply APPLY] [-o DST_DIR] [--overwrite] [-p PROJECT_DIR] [target]

Parameters:

  • <target> (string) - Target dataset revpath. By default, filters all targets of the current project.
  • -e, --filter (string) - XML XPath filter expression for dataset items
  • -m, --mode (string) - The filtering mode. Default is the i mode.
  • --dry-run - Print XML representations of the filtered dataset and exit.
  • --stage (bool) - Include this action as a project build step. If true, this operation will be saved in the project build tree, allowing to reproduce the resulting dataset later. Applicable only to main project targets (i.e. data sources and the project target, but not intermediate stages). Enabled by default.
  • --apply (bool) - Run this command immediately. If disabled, only the build tree stage will be written. Enabled by default.
  • -o, --output-dir (string) - Output directory. Can be omitted for main project targets (i.e. data sources and the project target, but not intermediate stages) and dataset targets. If not specified, the results will be saved inplace.
  • --overwrite - Allows to overwrite existing files in the output directory, when it is specified and is not empty.
  • -p, --project (string) - Directory of the project to operate on (default: current directory).
  • -h, --help - Print the help message and exit.

Example: extract a dataset with images with width < height

datum filter \
  -p test_project \
  -e '/item[image/width < image/height]'

Example: extract a dataset with images of the train subset

datum filter \
  -p test_project \
  -e '/item[subset="train"]'

Example: extract a dataset with only large annotations of the cat class and any non-persons

datum filter \
  -p test_project \
  --mode annotations \
  -e '/item/annotation[(label="cat" and area > 99.5) or label!="person"]'

Example: extract a dataset with non-occluded annotations, remove empty images. Use data only from the “s1” source of the project.

datum create
datum import --format voc -i <path/to/dataset1/> --name s1
datum import --format voc -i <path/to/dataset2/> --name s2
datum filter s1 \
  -m i+a -e '/item/annotation[occluded="False"]'

5.12 - Generate Datasets

Creates a synthetic dataset with elements of the specified type and shape, and saves it in the provided directory.

Currently, can only generate fractal images, useful for network compression. To create 3-channel images, you should provide the number of images, height and width. The images are colorized with a model, which will be downloaded automatically. Uses the algorithm from the article: https://arxiv.org/abs/2103.13023

Usage:

datum generate [-h] -o OUTPUT_DIR -k COUNT --shape SHAPE [SHAPE ...]
  [-t {image}] [--overwrite] [--model-dir MODEL_PATH]

Parameters:

  • -o, --output-dir (string) - Output directory
  • -k, --count (integer) - Number of images to be generated
  • --shape (integer, repeatable) - Dimensions of data to be generated (H, W)
  • -t, --type (one of: image) - Specify the type of data to generate (default: image)
  • --model-dir (path) - Path to load the colorization model from. If no model is found, the model will be downloaded (default: current dir)
  • --overwrite - Allows overwriting existing files in the output directory, when it is not empty.
  • -h, --help - Print the help message and exit.

Examples: Generate 300 3-channel fractal images with H=224, W=256 and store in the images/ dir:

datum generate -o images/ --count 300 --shape 224 256

5.13 - Print dataset info

This command outputs high level dataset information such as sample count, categories and subsets.

Usage:

datum info [-h] [--json] [-p PROJECT_DIR] [revpath]

Parameters:

  • <target> (string) - Target dataset revpath. By default, prints info about the joined project dataset.
  • --json - Print output data in JSON format
  • -p, --project (string) - Directory of the project to operate on (default: current directory).
  • -h, --help - Print the help message and exit.

Examples:

  • Print info about a project dataset: datum info -p test_project/

  • Print info about a COCO-like dataset: datum info path/to/dataset:coco

Sample output:

format: voc
media type: image
length: 5
categories:
    labels: background, aeroplane, bicycle, bird, boat, bottle, bus, car, cat, chair (and 12 more)
subsets:
    trainval:
        length: 5

JSON output format:

{
  "format": string,
  "media type": string,
  "length": integer,
  "categories": {
    "count": integer,
    "labels": [
      {
        "id": integer,
        "name": string,
        "parent": string,
        "attributes": [ string, ... ]
      },
      ...
    ]
  },
  "subsets": [
    {
      "name": string,
      "length": integer
    },
    ...
  ]
}

5.14 - Log

This command prints the history of the current project revision.

Prints lines in the following format: <short commit hash> <commit message>

Usage:

datum log [-h] [-n MAX_COUNT] [-p PROJECT_DIR]

Parameters:

  • -n, --max-count (number, default: 10) - The maximum number of previous revisions in the output
  • -p, --project (string) - Directory of the project to operate on (default: current directory).
  • -h, --help - Print the help message and exit.

Example output:

affbh33 Added COCO dataset
eeffa35 Added VOC dataset

5.15 - Merge Datasets

Consider the following task: there is a set of images (the original dataset) we want to annotate. Suppose we did this manually and/or automated it using models, and now we have few sets of annotations for the same images. We want to merge them and produce a single set of high-precision annotations.

Another use case: there are few datasets with different sets of images and labels, which we need to combine in a single dataset. If the labels were the same, we could just join the datasets. But in this case we need to merge labels and adjust the annotations in the resulting dataset.

In Datumaro, it can be done with the merge command. This command merges 2 or more datasets and checks annotations for errors.

In simple cases, when dataset images do not intersect and new labels are not added, the recommended way of merging is using the patch command. It will offer better performance and provide the same results.

Datasets are merged by items, and item annotations are merged by finding the unique ones across datasets. Annotations are matched between matching dataset items by distance. Spatial annotations are compared by the applicable distance measure (IoU, OKS, PDJ etc.), labels and annotation attributes are selected by voting. Each set of matching annotations produces a single annotation in the resulting dataset. The score (a number in the range [0; 1]) attribute indicates the agreement between different sources in the produced annotation. The working time of the function can be estimated as O( (summary dataset length) * (dataset count) ^ 2 * (item annotations) ^ 2 )

This command also allows to merge datasets with different, or partially overlapping sets of labels (which is impossible by simple joining).

During the process, some merge conflicts can appear. For example, it can be mismatching dataset images having the same ids, label voting can be unsuccessful if quorum is not reached (the --quorum parameter), bboxes may be too close (the -iou parameter) etc. Found merge conflicts, missing items or annotations, and other errors are saved into an output .json file.

In Datumaro, annotations can be grouped. It can be useful to represent different parts of a single object - for example, it can be different parts of a human body, parts of a vehicle etc. This command allows to check annotation groups for completeness with the -g/--groups option. If used, this parameter must specify a list of labels for annotations that must be in the same group. It can be particularly useful to check if separate keypoints are grouped and all the necessary object components in the same group.

This command has multiple forms:

1) datum merge <revpath>
2) datum merge <revpath> <revpath> ...

<revpath> - either a dataset path or a revision path.

1 - Merges the current project’s main target (“project”) in the working tree with the specified dataset.

2 - Merges the specified datasets. Note that the current project is not included in the list of merged sources automatically.

The command supports passing extra exporting options for the output dataset. The format can be specified with the -f/--format option. Extra options should be passed after the main arguments and after the -- separator. Particularly, this is useful to include images in the output dataset with --save-images.

Usage:

datum merge [-h] [-iou IOU_THRESH] [-oconf OUTPUT_CONF_THRESH]
  [--quorum QUORUM] [-g GROUPS] [-o DST_DIR] [--overwrite]
  [-p PROJECT_DIR] [-f FORMAT]
  target [target ...] [-- EXTRA_FORMAT_ARGS]

Parameters:

  • <target> (string) - Target dataset revpaths (repeatable)
  • -iou, --iou-thresh (number) - IoU matching threshold for spatial annotations (both maximum inter-cluster and pairwise). Default is 0.25.
  • --quorum (number) - Minimum count of votes for a label or attribute to be counted. Default is 0.
  • -g, --groups (string) - A comma-separated list of label names in annotation groups to check. The ? postfix can be added to a label to make it optional in the group (repeatable)
  • -oconf, --output-conf-thresh (number) - Confidence threshold for output annotations to be included in the resulting dataset. Default is 0.
  • -o, --output-dir (string) - Output directory. By default, a new directory is created in the current directory.
  • --overwrite - Allows to overwrite existing files in the output directory, when it is specified and is not empty.
  • -f, --format (string) - Output format. The default format is datumaro.
  • -p, --project (string) - Directory of the project to operate on (default: current directory).
  • -h, --help - Print the help message and exit.
  • -- <extra format args> - Additional arguments for the format writer (use -- -h for help). Must be specified after the main command arguments.

Examples:

Merge 4 (partially-)intersecting projects,

  • consider voting successful when there are no less than 3 same votes
  • consider shapes intersecting when IoU >= 0.6
  • check annotation groups to have person, hand, head and foot (? is used for optional parts)
datum merge project1/ project2/ project3/ project4/ \
  --quorum 3 \
  -iou 0.6 \
  --groups 'person,hand?,head,foot?'

Merge images and annotations from 2 datasets in COCO format: datum merge dataset1/:image_dir dataset2/:coco dataset3/:coco

Check groups of the merged dataset for consistency: look for groups consisting of person, hand head, foot datum merge project1/ project2/ -g 'person,hand?,head,foot?'

Merge two datasets, specify formats: datum merge path/to/dataset1:voc path/to/dataset2:coco

Merge the current working tree and a dataset: datum merge path/to/dataset2:coco

Merge a source from a previous revision and a dataset: datum merge HEAD~2:source-2 path/to/dataset2:yolo

Merge datasets and save in different format: datum merge -f voc dataset1/:yolo path2/:coco -- --save-images

5.16 - Models

Register model

Datumaro can execute deep learning models in various frameworks. Check the plugins section for more info.

Supported frameworks:

  • OpenVINO
  • Custom models via custom launchers

Models need to be added to the Datumaro project first. It can be done with the datum model add command.

Usage:

datum model add [-h] [-n NAME] -l LAUNCHER [--copy] [--no-check]
  [-p PROJECT_DIR] [-- EXTRA_ARGS]

Parameters:

  • -l, --launcher (string) - Model launcher name
  • --copy - Copy model data into project. By default, only the link is saved.
  • --no-check - Don’t check the model can be loaded
  • -n, --name (string) - Name of the new model (default: generate automatically)
  • -p, --project (string) - Directory of the project to operate on (default: current directory).
  • -h, --help - Print the help message and exit.
  • <extra args> - Additional arguments for the model launcher (use -- -h for help). Must be specified after the main command arguments.

Example: register an OpenVINO model

A model consists of a graph description and weights. There is also a script used to convert model outputs to internal data structures.

datum create
datum model add \
  -n <model_name> -l openvino -- \
  -d <path_to_xml> -w <path_to_bin> -i <path_to_interpretation_script>

Interpretation script for an OpenVINO detection model (convert.py): You can find OpenVINO model interpreter samples in datumaro/plugins/openvino/samples (instruction).

import datumaro as dm

max_det = 10
conf_thresh = 0.1

def process_outputs(inputs, outputs):
    # inputs = model input, array or images, shape = (N, C, H, W)
    # outputs = model output, shape = (N, 1, K, 7)
    # results = conversion result, [ [ Annotation, ... ], ... ]
    results = []
    for input, output in zip(inputs, outputs):
        input_height, input_width = input.shape[:2]
        detections = output[0]
        image_results = []
        for det in detections:
            label = int(det[1])
            conf = float(det[2])
            if conf <= conf_thresh:
                continue

            x = max(int(det[3] * input_width), 0)
            y = max(int(det[4] * input_height), 0)
            w = min(int(det[5] * input_width - x), input_width)
            h = min(int(det[6] * input_height - y), input_height)
            image_results.append(dm.Bbox(x, y, w, h,
                label=label, attributes={'score': conf} ))

            results.append(image_results[:max_det])

    return results

def get_categories():
    # Optionally, provide output categories - label map etc.
    # Example:
    label_categories = dm.LabelCategories()
    label_categories.add('person')
    label_categories.add('car')
    return { dm.AnnotationType.label: label_categories }

Remove Models

To remove a model from a project, use the datum model remove command.

Usage:

datum model remove [-h] [-p PROJECT_DIR] name

Parameters:

  • <name> (string) - The name of the model to be removed
  • -p, --project (string) - Directory of the project to operate on (default: current directory).
  • -h, --help - Print the help message and exit.

Example:

datum create
datum model add <...> -n model1
datum remove model1

Run Model

This command applies model to dataset images and produces a new dataset.

Usage:

datum model run

Parameters:

  • <target> (string) - A project build target to be used. By default, uses the combined project target.
  • -m, --model (string) - Model name
  • -o, --output-dir (string) - Output directory. By default, results will be stored in an auto-generated directory in the current directory.
  • --overwrite - Allows to overwrite existing files in the output directory, when it is specified and is not empty.
  • -p, --project (string) - Directory of the project to operate on (default: current directory).
  • -h, --help - Print the help message and exit.

Example: launch inference on a dataset

datum create
datum import <...>
datum model add mymodel <...>
datum model run -m mymodel -o inference

5.17 - Patch Datasets

Updates items of the first dataset with items from the second one.

By default, datasets are updated in-place. The -o/--output-dir option can be used to specify another output directory. When updating in-place, use the --overwrite parameter along with the --save-images export option (in-place updates fail by default to prevent data loss).

Unlike the regular project data source joining, the datasets are not required to have the same labels. The labels from the “patch” dataset are projected onto the labels of the patched dataset, so only the annotations with the matching labels are used, i.e. all the annotations having unknown labels are ignored. Currently, this command doesn’t allow to update the label information in the patched dataset.

The command supports passing extra exporting options for the output dataset. The extra options should be passed after the main arguments and after the -- separator. Particularly, this is useful to include images in the output dataset with --save-images.

This command can be applied to the current project targets or arbitrary datasets outside a project. Note that if the target dataset is read-only (e.g. if it is a project, stage or a cache entry), the output directory must be provided.

Usage:

datum patch [-h] [-o DST_DIR] [--overwrite] [-p PROJECT_DIR]
  target patch
  [-- EXPORT_ARGS]

<revpath> - either a dataset path or a revision path.

The current project (-p/--project) is also used as a context for plugins, so it can be useful for dataset paths having custom formats. When not specified, the current project’s working tree is used.

Parameters:

  • <target dataset> (string) - Target dataset revpath
  • <patch dataset> (string) - Patch dataset revpath
  • -o, --output-dir (string) - Output directory. By default, saves in-place
  • --overwrite - Allows to overwrite existing files in the output directory, when it is not empty.
  • -p, --project (string) - Directory of the project to operate on (default: current directory).
  • -h, --help - Print the help message and exit.
  • -- <export args> - Additional arguments for the format writer (use -- -h for help). Must be specified after the main command arguments.

Examples:

  • Update a VOC-like dataset with COCO-like annotations:
datum patch --overwrite dataset1/:voc dataset2/:coco -- --save-images
  • Generate a patched dataset, based on a project:
datum patch -o patched_proj1/ proj1/ proj2/
  • Update the “source1” source in the current project with a dataset:
datum patch -p proj/ --overwrite source1 path/to/dataset2:coco
  • Generate a patched source from a previous revision and a dataset:
datum patch -o new_src2/ HEAD~2:source-2 path/to/dataset2:yolo
  • Update a dataset in a custom format, described in a project plugin:
datum patch -p proj/ --overwrite dataset/:my_format dataset2/:coco

5.18 - Projects

Migrate project

Updates the project from an old version to the current one and saves the resulting project in the output directory. Projects cannot be updated inplace.

The command tries to map the old source configuration to the new one. This can fail in some cases, so the command will exit with an error, unless -f/--force is specified. With this flag, the command will skip these errors an continue its work.

Usage:

datum project migrate [-h] -o DST_DIR [-f] [-p PROJECT_DIR] [--overwrite]

Parameters:

  • -o, --output-dir (string) - Output directory for the updated project
  • -f, --force - Ignore source import errors (default: False)
  • --overwrite - Overwrite existing files in the save directory.
  • -p, --project (string) - Directory of the project to operate on (default: current directory).
  • -h, --help - Print the help message and exit.

Examples:

  • Migrate a project from v1 to v2, save the new project in other dir: datum project migrate -o <output/dir>

Prints project configuration info such as available plugins, registered models, imported sources and build tree.

Usage:

datum project info [-h] [-p PROJECT_DIR] [revision]

Parameters:

  • <revision> (string) - Target project revision. By default, uses the working tree.
  • -p, --project (string) - Directory of the project to operate on (default: current directory).
  • -h, --help - Print the help message and exit.

Examples:

  • Print project info for the current working tree: datum project info

  • Print project info for the previous revision: datum project info HEAD~1

Sample output:

Project:
  location: /test_proj

Plugins:
  extractors: ade20k2017, ade20k2020, camvid, cifar, cityscapes, coco, coco_captions, coco_image_info, coco_instances, coco_labels, coco_panoptic, coco_person_keypoints, coco_stuff, cvat, datumaro, icdar_text_localization, icdar_text_segmentation, icdar_word_recognition, image_dir, image_zip, imagenet, imagenet_txt, kitti, kitti_detection, kitti_raw, kitti_segmentation, label_me, lfw, market1501, mnist, mnist_csv, mot_seq, mots, mots_png, open_images, sly_pointcloud, tf_detection_api, vgg_face2, voc, voc_action, voc_classification, voc_detection, voc_layout, voc_segmentation, wider_face, yolo

  converters: camvid, mot_seq_gt, coco_captions, coco, coco_image_info, coco_instances, coco_labels, coco_panoptic, coco_person_keypoints, coco_stuff, kitti, kitti_detection, kitti_segmentation, icdar_text_localization, icdar_text_segmentation, icdar_word_recognition, lfw, datumaro, open_images, image_zip, cifar, yolo, voc_action, voc_classification, voc, voc_detection, voc_layout, voc_segmentation, tf_detection_api, label_me, mnist, cityscapes, mnist_csv, kitti_raw, wider_face, vgg_face2, sly_pointcloud, mots_png, image_dir, imagenet_txt, market1501, imagenet, cvat

  launchers:

Models:

Sources:
  'source-2':
    format: voc
    url: /datasets/pascal/VOC2012
    location: /test_proj/source-2/
    options: {}
    hash: 3eb282cdd7339d05b75bd932a1fd3201
    stages:
      'root':
        type: source
        hash: 3eb282cdd7339d05b75bd932a1fd3201
  'source-3':
    format: imagenet
    url: /datasets/imagenet/ILSVRC2012_img_val/train
    location: /test_proj/source-3/
    options: {}
    hash: e47804a3ec1a54c9b145e5f1007ec72f
    stages:
      'root':
        type: source
        hash: e47804a3ec1a54c9b145e5f1007ec72f

5.19 - Sources

These commands are specific for Data Sources. Read more about them here.

Import Dataset

Datasets can be added to a Datumaro project with the import command, which adds a dataset link into the project and downloads (or copies) the dataset. If you need to add a dataset already copied into the project, use the add command.

Dataset format readers can provide some additional import options. To pass such options, use the -- separator after the main command arguments. The usage information can be printed with datum import -f <format> -- --help.

The list of currently available formats is listed in the command help output.

A dataset is imported by its URL. Currently, only local filesystem paths are supported. The URL can be a file or a directory path to a dataset. When the dataset is read, it is read as a whole. However, many formats can have multiple subsets like train, val, test etc. If you want to limit reading only to a specific subset, use the -r/--path parameter. It can also be useful when subset files have non-standard placement or names.

When a dataset is imported, the following things are done:

  • URL is saved in the project config
  • data in copied into the project

Each data source has a name assigned, which can be used in other commands. To set a specific name, use the -n/--name parameter.

The dataset is added into the working tree of the project. A new commit is not done automatically.

Usage:

datum import [-h] [-n NAME] -f FORMAT [-r PATH] [--no-check]
  [-p PROJECT_DIR] url [-- EXTRA_FORMAT_ARGS]

Parameters:

  • <url> (string) - A file of directory path to the dataset.
  • -f, --format (string) - Dataset format
  • -r, --path (string) - A path relative to the source URL the data source. Useful to specify a path to a subset, subtask, or a specific file in URL.
  • --no-check - Don’t try to read the source after importing
  • -n, --name (string) - Name of the new source (default: generate automatically)
  • -p, --project (string) - Directory of the project to operate on (default: current directory).
  • -h, --help - Print the help message and exit.
  • -- <extra format args> - Additional arguments for the format reader (use -- -h for help). Must be specified after the main command arguments.

Example: create a project from images and annotations in different formats, export as TFrecord for TF Detection API for model training

# 'default' is the name of the subset below
datum create
datum import -f coco_instances -r annotations/instances_default.json path/to/coco
datum import -f cvat <path/to/cvat/default.xml>
datum import -f voc_detection -r custom_subset_dir/default.txt <path/to/voc>
datum import -f datumaro <path/to/datumaro/default.json>
datum import -f image_dir <path/to/images/dir>
datum export -f tf_detection_api -- --save-images

Add Dataset

Existing datasets can be added to a Datumaro project with the add command. The command adds a project-local directory as a data source in the project. Unlike the import command, it does not copy datasets and only works with local directories. The source name is defined by the directory name.

Dataset format readers can provide some additional import options. To pass such options, use the -- separator after the main command arguments. The usage information can be printed with datum add -f <format> -- --help.

The list of currently available formats is listed in the command help output.

A dataset is imported as a directory. When the dataset is read, it is read as a whole. However, many formats can have multiple subsets like train, val, test etc. If you want to limit reading only to a specific subset, use the -r/--path parameter. It can also be useful when subset files have non-standard placement or names.

The dataset is added into the working tree of the project. A new commit is not done automatically.

Usage:

datum add [-h] -f FORMAT [-r PATH] [--no-check]
  [-p PROJECT_DIR] path [-- EXTRA_FORMAT_ARGS]

Parameters:

  • <url> (string) - A file of directory path to the dataset.
  • -f, --format (string) - Dataset format
  • -r, --path (string) - A path relative to the source URL the data source. Useful to specify a path to a subset, subtask, or a specific file in URL.
  • --no-check - Don’t try to read the source after importing
  • -p, --project (string) - Directory of the project to operate on (default: current directory).
  • -h, --help - Print the help message and exit.
  • -- <extra format args> - Additional arguments for the format reader (use -- -h for help). Must be specified after the main command arguments.

Example: create a project from images and annotations in different formats, export in YOLO for model training

datum create
datum add -f coco -r annotations/instances_train.json dataset1/
datum add -f cvat dataset2/train.xml
datum export -f yolo -- --save-images

Example: add an existing dataset into a project, avoid data copying

To add a dataset, we need to have it inside the project directory:

proj/
├─ .datumaro/
├─ .dvc/
├─ my_coco/
│  └─ images/
│     ├─ image1.jpg
│     └─ ...
│  └─ annotations/
│     └─ coco_annotation.json
├─ .dvcignore
└─ .gitignore

datum create -o proj/
mv ~/my_coco/ proj/my_coco/ # move the dataset into the project directory
datum add -p proj/ -f coco proj/my_coco/

Remove Datasets

To remove a data source from a project, use the remove command.

Usage:

datum remove [-h] [--force] [--keep-data] [-p PROJECT_DIR] name [name ...]

Parameters:

  • <name> (string) - The name of the source to be removed (repeatable)
  • -f, --force - Do not fail and stop on errors during removal
  • --keep-data - Do not remove source data from the working directory, remove only project metainfo.
  • -p, --project (string) - Directory of the project to operate on (default: current directory).
  • -h, --help - Print the help message and exit.

Example:

datum create
datum import -f voc -n src1 <path/to/dataset/>
datum remove src1

5.20 - Get Project Statistics

This command computes various project statistics, such as:

  • image mean and std. dev.
  • class and attribute balance
  • mask pixel balance
  • segment area distribution

Usage:

datum stats [-h] [-p PROJECT_DIR] [target]

Parameters:

  • <target> (string) - Target source revpath. By default, computes statistics of the merged dataset.
  • -s, --subset (string) - Compute stats only for a specific subset
  • --image-stats (bool) - Compute image mean and std (default: True)
  • --ann-stats (bool) - Compute annotation statistics (default: True)
  • -p, --project (string) - Directory of the project to operate on (default: current directory).
  • -h, --help - Print the help message and exit.

Example:

datum stats -p test_project

Sample output:

{
    "annotations": {
        "labels": {
            "attributes": {
                "gender": {
                    "count": 358,
                    "distribution": {
                        "female": [
                            149,
                            0.41620111731843573
                        ],
                        "male": [
                            209,
                            0.5837988826815642
                        ]
                    },
                    "values count": 2,
                    "values present": [
                        "female",
                        "male"
                    ]
                },
                "view": {
                    "count": 340,
                    "distribution": {
                        "__undefined__": [
                            4,
                            0.011764705882352941
                        ],
                        "front": [
                            54,
                            0.1588235294117647
                        ],
                        "left": [
                            14,
                            0.041176470588235294
                        ],
                        "rear": [
                            235,
                            0.6911764705882353
                        ],
                        "right": [
                            33,
                            0.09705882352941177
                        ]
                    },
                    "values count": 5,
                    "values present": [
                        "__undefined__",
                        "front",
                        "left",
                        "rear",
                        "right"
                    ]
                }
            },
            "count": 2038,
            "distribution": {
                "car": [
                    340,
                    0.16683022571148184
                ],
                "cyclist": [
                    194,
                    0.09519136408243375
                ],
                "head": [
                    354,
                    0.17369970559371933
                ],
                "ignore": [
                    100,
                    0.04906771344455348
                ],
                "left_hand": [
                    238,
                    0.11678115799803729
                ],
                "person": [
                    358,
                    0.17566241413150147
                ],
                "right_hand": [
                    77,
                    0.037782139352306184
                ],
                "road_arrows": [
                    326,
                    0.15996074582924436
                ],
                "traffic_sign": [
                    51,
                    0.025024533856722278
                ]
            }
        },
        "segments": {
            "area distribution": [
                {
                    "count": 1318,
                    "max": 11425.1,
                    "min": 0.0,
                    "percent": 0.9627465303140978
                },
                {
                    "count": 1,
                    "max": 22850.2,
                    "min": 11425.1,
                    "percent": 0.0007304601899196494
                },
                {
                    "count": 0,
                    "max": 34275.3,
                    "min": 22850.2,
                    "percent": 0.0
                },
                {
                    "count": 0,
                    "max": 45700.4,
                    "min": 34275.3,
                    "percent": 0.0
                },
                {
                    "count": 0,
                    "max": 57125.5,
                    "min": 45700.4,
                    "percent": 0.0
                },
                {
                    "count": 0,
                    "max": 68550.6,
                    "min": 57125.5,
                    "percent": 0.0
                },
                {
                    "count": 0,
                    "max": 79975.7,
                    "min": 68550.6,
                    "percent": 0.0
                },
                {
                    "count": 0,
                    "max": 91400.8,
                    "min": 79975.7,
                    "percent": 0.0
                },
                {
                    "count": 0,
                    "max": 102825.90000000001,
                    "min": 91400.8,
                    "percent": 0.0
                },
                {
                    "count": 50,
                    "max": 114251.0,
                    "min": 102825.90000000001,
                    "percent": 0.036523009495982466
                }
            ],
            "avg. area": 5411.624543462382,
            "pixel distribution": {
                "car": [
                    13655,
                    0.0018431496518735067
                ],
                "cyclist": [
                    939005,
                    0.12674674030446592
                ],
                "head": [
                    0,
                    0.0
                ],
                "ignore": [
                    5501200,
                    0.7425510702956085
                ],
                "left_hand": [
                    0,
                    0.0
                ],
                "person": [
                    954654,
                    0.12885903974805205
                ],
                "right_hand": [
                    0,
                    0.0
                ],
                "road_arrows": [
                    0,
                    0.0
                ],
                "traffic_sign": [
                    0,
                    0.0
                ]
            }
        }
    },
    "annotations by type": {
        "bbox": {
            "count": 548
        },
        "caption": {
            "count": 0
        },
        "label": {
            "count": 0
        },
        "mask": {
            "count": 0
        },
        "points": {
            "count": 669
        },
        "polygon": {
            "count": 821
        },
        "polyline": {
            "count": 0
        }
    },
    "annotations count": 2038,
    "unannotated images": [
        "img00051",
        "img00052",
        "img00053",
        "img00054",
        "img00055",
    ],
    "unannotated images count": 5,

    "dataset": {
        "images count": 100,
        "unique images count": 97,
        "repeated images count": 3,
        "repeated images": [
            [["img00057", "default"], ["img00058", "default"]],
            [["img00059", "default"], ["img00060", "default"]],
            [["img00061", "default"], ["img00062", "default"]],
        ],
    },
    "subsets": {
        "default": {
            "images count": 100,
            "image mean": [
                107.06903686941979,
                79.12831698580979,
                52.95829558185416
            ],
            "image std": [
                49.40237673503467,
                43.29600731496902,
                35.47373007603151
            ],

        }
    },
}

5.21 - Status

This command prints the summary of the source changes between the working tree of a project and its HEAD revision.

Prints lines in the following format: <status> <source name>

The list of possible status values:

  • modified - the source data exists and it is changed
  • foreign_modified - the source data exists and it is changed, but Datumaro does not know about the way the differences were made. If changes are committed, they will only be available for reproduction from the project cache.
  • added - the source was added in the working tree
  • removed - the source was removed from the working tree. This status won’t be reported if just the source data is removed in the working tree. In such situation the status will be missing.
  • missing - the source data is removed from the working directory. The source still can be restored from the project cache or reproduced.

Usage:

datum status [-h] [-p PROJECT_DIR]

Parameters:

  • -p, --project (string) - Directory of the project to operate on (default: current directory).
  • -h, --help - Print the help message and exit.

Example output:

added source-1
modified source-2
foreign_modified source-3
removed source-4
missing source-5

5.22 - Transform Dataset

Often datasets need to be modified during preparation for model training and experimenting. In trivial cases it can be done manually - e.g. image renaming or label renaming. However, in more complex cases even simple modifications can require too much efforts, distracting the user from the real work. Datumaro provides the datum transform command to help in such cases.

This command allows to modify dataset images or annotations all at once.

This command is designed for batch dataset processing, so if you only need to modify few elements of a dataset, you might want to use other approaches for better performance. A possible solution can be a simple script, which uses Datumaro API.

The command can be applied to a dataset or a project build target, a stage or the combined project target, in which case all the project targets will be affected. A build tree stage will be recorded if --stage is enabled, and the resulting dataset(-s) will be saved if --apply is enabled.

By default, datasets are updated in-place. The -o/--output-dir option can be used to specify another output directory. When updating in-place, use the --overwrite parameter (in-place updates fail by default to prevent data loss), unless a project target is modified.

The current project (-p/--project) is also used as a context for plugins, so it can be useful for dataset paths having custom formats. When not specified, the current project’s working tree is used.

Usage:

datum transform [-h] -t TRANSFORM [-o DST_DIR] [--overwrite]
  [-p PROJECT_DIR] [--stage STAGE] [--apply APPLY] [target] [-- EXTRA_ARGS]

Parameters:

  • <target> (string) - Target dataset revpath. By default, transforms all targets of the current project.
  • -t, --transform (string) - Transform method name
  • --stage (bool) - Include this action as a project build step. If true, this operation will be saved in the project build tree, allowing to reproduce the resulting dataset later. Applicable only to main project targets (i.e. data sources and the project target, but not intermediate stages). Enabled by default.
  • --apply (bool) - Run this command immediately. If disabled, only the build tree stage will be written. Enabled by default.
  • -o, --output-dir (string) - Output directory. Can be omitted for main project targets (i.e. data sources and the project target, but not intermediate stages) and dataset targets. If not specified, the results will be saved inplace.
  • --overwrite - Allows to overwrite existing files in the output directory, when it is specified and is not empty.
  • -p, --project (string) - Directory of the project to operate on (default: current directory).
  • -h, --help - Print the help message and exit.
  • <extra args> - The list of extra transformation parameters. Should be passed after the -- separator after the main command arguments. See transform descriptions for info about extra parameters. Use the --help option to print parameter info.

Examples:

  • Split a VOC-like dataset randomly:
datum transform -t random_split --overwrite path/to/dataset:voc
  • Rename images in a project data source by a regex from frame_XXX to XXX:
datum create <...>
datum import <...> -n source-1
datum transform -t rename source-1 -- -e '|^frame_||'

Built-in transforms

Basic dataset item manipulations:

Subset manipulations:

  • random_split - Splits dataset into subsets randomly
  • split - Splits dataset into subsets for classification, detection, segmentation or re-identification
  • map_subsets - Renames and removes subsets

Annotation manipulations:

rename

Renames items in the dataset. Supports regular expressions. The first character in the expression is a delimiter for the pattern and replacement parts. Replacement part can also contain str.format replacement fields with the item (of type DatasetItem) object available.

Usage:

rename [-h] [-e REGEX]

Optional arguments:

  • -h, --help (flag) - Show this help message and exit
  • -e, --regex (string) - Regex for renaming in the form <sep><search><sep><replacement><sep>

Examples: Replace ‘pattern’ with ‘replacement’:

datum transform -t rename -- -e '|pattern|replacement|'

Remove the frame_ prefix from item ids:

datum transform -t rename -- -e '|^frame_|\1|'

Collect images from subdirectories into the base image directory using regex:

datum transform -t rename -- -e '|^((.+[/\\])*)?(.+)$|\2|'

Add subset prefix to images:

datum transform -t rename -- -e '|(.*)|{item.subset}_\1|'

id_from_image_name

Renames items in the dataset using image file name (without extension).

Usage:

id_from_image_name [-h]

Optional arguments:

  • -h, --help (flag) - Show this help message and exit

reindex

Replaces dataset item IDs with sequential indices.

Usage:

reindex [-h] [-s START]

Optional arguments:

  • -h, --help (flag) - Show this help message and exit
  • -s, --start (int) - Start value for item ids (default: 1)

ndr

Removes near-duplicated images in subset.

Remove duplicated images from a dataset. Keep at most -k/--num_cut resulting images.

Available oversampling policies (the -e parameter):

  • random - sample from removed data randomly
  • similarity - sample from removed data with ascending similarity score

Available undersampling policies (the -u parameter):

  • uniform - sample data with uniform distribution
  • inverse - sample data with reciprocal of the number of number of items with the same similarity

Usage:

ndr [-h] [-w WORKING_SUBSET] [-d DUPLICATED_SUBSET] [-a {gradient}]
  [-k NUM_CUT] [-e {random,similarity}] [-u {uniform,inverse}] [-s SEED]

Optional arguments:

  • -h, --help (flag) - Show this help message and exit
  • -w, --working_subset (str) - Name of the subset to operate (default: None)
  • -d, --duplicated_subset (str) - Name of the subset for the removed data after NDR runs (default: duplicated)
  • -a, --algorithm (one of: gradient) - Name of the algorithm to use (default: gradient)
  • -k, --num_cut (int) - Maximum output dataset size
  • -e, --over_sample (one of: random, similarity) - The policy to use when num_cut is bigger than result length (default: random)
  • -u, --under_sample (one of: uniform, inverse) - The policy to use when num_cut is smaller than result length (default: uniform)
  • -s, --seed (int) - Random seed

Example: apply NDR, return no more than 100 images

datum transform -t ndr -- \
  --working_subset train
  --algorithm gradient
  --num_cut 100
  --over_sample random
  --under_sample uniform

relevancy_sampler

Sampler that analyzes model inference results on the dataset and picks the most relevant samples for training.

Creates a dataset from the -k/--count hardest items for a model. The whole dataset or a single subset will be split into the sampled and unsampled subsets based on the model confidence. The dataset must contain model confidence values in the scores attributes of annotations.

There are five methods of sampling (the -m/--method option):

  • topk - Return the k items with the highest uncertainty data
  • lowk - Return the k items with the lowest uncertainty data
  • randk - Return random k items
  • mixk - Return a half using topk, and the other half using lowk method
  • randtopk - Select 3*k items randomly, and return the topk among them

Notes:

  • Each image’s inference result must contain the probability for all classes (in the scores attribute).
  • Requesting a sample larger than the number of all images will return all images.

Usage:

relevancy_sampler [-h] -k COUNT [-a {entropy}] [-i INPUT_SUBSET]
  [-o SAMPLED_SUBSET] [-u UNSAMPLED_SUBSET]
  [-m {topk,lowk,randk,mixk,randtopk}] [-d OUTPUT_FILE]

Optional arguments:

  • -h, --help (flag) - Show this help message and exit
  • -k, --count (int) - Number of items to sample
  • -a, --algorithm (one of: entropy) - Sampling algorithm (default: entropy)
  • -i, --input_subset (str) - Subset name to select sample from (default: None)
  • -o, --sampled_subset (str) - Subset name to put sampled data to (default: sample)
  • -u, --unsampled_subset (str) - Subset name to put the rest data to (default: unsampled)
  • -m, --sampling_method (one of: topk, lowk, randk, mixk, randtopk) - Sampling method (default: topk)
  • -d, --output_file (path) - A .csv file path to dump sampling results

Examples: Select the most relevant data subset of 20 images based on model certainty, put the result into sample subset and put all the rest into unsampled subset, use train subset as input. The dataset must contain model confidence values in the scores attributes of annotations.

datum transform -t relevancy_sampler -- \
  --algorithm entropy \
  --subset_name train \
  --sample_name sample \
  --unsampled_name unsampled \
  --sampling_method topk -k 20

random_sampler

Sampler that keeps no more than required number of items in the dataset.

Notes:

  • Items are selected uniformly (tries to keep original item distribution by subsets)
  • Requesting a sample larger than the number of all images will return all images

Usage:

random_sampler [-h] -k COUNT [-s SUBSET] [--seed SEED]

Optional arguments:

  • -h, --help (flag) - Show this help message and exit
  • -k, --count (int) - Maximum number of items to sample
  • -s, --subset (str) - Limit changes to this subset (default: affect all dataset)
  • --seed (int) - Initial value for random number generator

Examples: Select subset of 20 images randomly

datum transform -t random_sampler -- -k 20

Select subset of 20 images, modify only train subset

datum transform -t random_sampler -- -k 20 -s train

random_label_sampler

Sampler that keeps at least the required number of annotations of each class in the dataset for each subset separately.

Consider using the “stats” command to get class distribution in the dataset.

Notes:

  • Items can contain annotations of several selected classes (e.g. 3 bounding boxes per image). The number of annotations in the resulting dataset varies between max(class counts) and sum(class counts)
  • If the input dataset does not has enough class annotations, the result will contain only what is available
  • Items are selected uniformly
  • For reasons above, the resulting class distribution in the dataset may not be the same as requested
  • The resulting dataset will only keep annotations for classes with specified count > 0

Usage:

label_random_sampler [-h] -k COUNT [-l LABEL_COUNTS] [--seed SEED]

Optional arguments:

  • -h, --help (flag) - Show this help message and exit
  • -k, --count (int) - Minimum number of annotations of each class
  • -l, --label (str; repeatable) - Minimum number of annotations of a specific class. Overrides the -k/--count setting for the class. The format is <label_name>:<count>
  • --seed (int) - Initial value for random number generator

Examples: Select a dataset with at least 10 images of each class:

datum transform -t label_random_sampler -- -k 10

Select a dataset with at least 20 cat images, 5 dog, 0 car and 10 of each unmentioned class:

datum transform -t label_random_sampler -- \
  -l cat:20 \ # keep 20 images with cats
  -l dog:5 \ # keep 5 images with dogs
  -l car:0 \ # remove car annotations
  -k 10 # for remaining classes

resize

Resizes images and annotations in the dataset to the specified size. Supports upscaling, downscaling and mixed variants.

Usage:

resize [-h] [-dw WIDTH] [-dh HEIGHT]

Optional arguments:

  • -h, --help (flag) - Show this help message and exit
  • -dw, --width (int) - Destination image width
  • -dh, --height (int) - Destination image height

Examples: Resize all images to 256x256 size

datum transform -t resize -- -dw 256 -dh 256

remove_images

Removes specific dataset items by their ids.

Usage:

remove_images [-h] [--id IDs]

Optional arguments:

  • -h, --help (flag) - Show this help message and exit
  • --id (str) - Item id to remove. Id is ‘:’ pair (repeatable)

Examples:

Remove specific images from the dataset

datum transform -t remove_images -- --id 'image1:train' --id 'image2:test'

remove_annotations

Allows to remove annotations on specific dataset items.

Can be useful to clean the dataset from broken or unnecessary annotations.

Usage:

remove_annotations [-h] [--id IDs]

Optional arguments:

  • -h, --help (flag) - Show this help message and exit
  • --id (str) - Item id to clean from annotations. Id is ‘:’ pair. If not specified, removes all annotations (repeatable)

Examples: Remove annotations from specific items in the dataset

datum transform -t remove_annotations -- --id 'image1:train' --id 'image2:test'

remove_attributes

Allows to remove item and annotation attributes in a dataset.

Can be useful to clean the dataset from broken or unnecessary attributes.

Usage:

remove_attributes [-h] [--id IDs] [--attr ATTRIBUTE_NAME]

Optional arguments:

  • -h, --help (flag) - Show this help message and exit
  • --id (str) - Image id to clean from annotations. Id is ‘:’ pair. If not specified, affects all items and annotations (repeatable)
  • -a, --attr (flag) - Attribute name to be removed. If not specified, removes all attributes (repeatable)

Examples: Remove the is_crowd attribute from dataset

datum transform -t remove_attributes -- \
  --attr 'is_crowd'

Remove the occluded attribute from annotations of the 2010_001705 item in the train subset

datum transform -t remove_attributes -- \
  --id '2010_001705:train' --attr 'occluded'

random_split

Joins all subsets into one and splits the result into few parts. It is expected that item ids are unique and subset ratios sum up to 1.

Usage:

random_split [-h] [-s SPLITS] [--seed SEED]

Optional arguments:

  • -h, --help (flag) - Show this help message and exit
  • -s, --subset (str, repeatable) - Subsets in the form: ‘:’ (repeatable, default: {train: 0.67, test: 0.33})
  • --seed (int) - Random seed

Example: Split a dataset randomly to train and test subsets, ratio is 2:1

datum transform -t random_split -- --subset train:.67 --subset test:.33

split

Splits a dataset for model training, using task information:

  • classification splits Splits dataset into subsets (train/val/test) in class-wise manner. Splits dataset images in the specified ratio, keeping the initial class distribution.

  • detection & segmentation splits Each image can have multiple object annotations - bbox, mask, polygon. Since an image shouldn’t be included in multiple subsets at the same time, and image annotations shouldn’t be split, in general, dataset annotations are unlikely to be split exactly in the specified ratio. This split tries to split dataset images as close as possible to the specified ratio, keeping the initial class distribution.

  • reidentification splits In this task, the test set should consist of images of unseen people or objects during the training phase. This function splits a dataset in the following way:

  1. Splits the dataset into train + val and test sets based on person or object ID.
  2. Splits test set into test-gallery and test-query sets in class-wise manner.
  3. Splits the train + val set into train and val sets in the same way. The final subsets would be train, val, test-gallery and test-query.

Notes:

  • Each image is expected to have only one Annotation. Unlabeled or multi-labeled images will be split into subsets randomly.
  • If Labels also have attributes, also splits by attribute values.
  • If there is not enough images in some class or attributes group, the split ratio can’t be guaranteed.

In reidentification task,

  • Object ID can be described by Label, or by attribute (--attr parameter)
  • The splits of the test set are controlled by --query parameter Gallery ratio would be 1.0 - query.

Usage:

split [-h] [-t {classification,detection,segmentation,reid}]
  [-s SPLITS] [--query QUERY] [--attr ATTR_FOR_ID] [--seed SEED]

Optional arguments:

  • -h, --help (flag) - Show this help message and exit
  • -t, --task (one of: classification, detection, segmentation, reid) - Dataset task (default: classification)
  • -s, --subset (str; repeatable) - Subsets in the form: ‘:’ (default: {train: 0.5, val: 0.2, test: 0.3})
  • --query (float) - Query ratio in the test set (default: 0.5)
  • --attr (str) - Attribute name representing the ID (default: use label)
  • --seed(int) - Random seed

Example:

datum transform -t split -- -t classification \
  --subset train:.5 --subset val:.2 --subset test:.3

datum transform -t split -- -t detection \
  --subset train:.5 --subset val:.2 --subset test:.3

datum transform -t split -- -t segmentation \
  --subset train:.5 --subset val:.2 --subset test:.3

datum transform -t split -- -t reid \
  --subset train:.5 --subset val:.2 --subset test:.3 --query .5

Example: use person_id attribute for splitting

datum transform -t split -- -t detection --attr person_id

map_subsets

Renames subsets in the dataset.

Usage:

map_subsets [-h] [-s MAPPING]

Optional arguments:

  • -h, --help (flag) - Show this help message and exit
  • -s, --subset (str; repeatable) - Subset mapping of the form: src:dst

remap_labels

Changes labels in the dataset.

A label can be:

  • renamed (and joined with existing) - when --label <old_name>:<new_name> is specified
  • deleted - when --label <name>: is specified, or default action is delete and the label is not mentioned in the list. When a label is deleted, all the associated annotations are removed
  • kept unchanged - when --label <name>:<name> is specified, or default action is keep and the label is not mentioned in the list Annotations with no label are managed by the default action policy.

Usage:

remap_labels [-h] [-l MAPPING] [--default {keep,delete}]

Optional arguments:

  • -h, --help (flag) - Show this help message and exit
  • -l, --label (str; repeatable) - Label in the form of: <src>:<dst>
  • --default (one of: keep, delete) - Action for unspecified labels (default: keep)

Examples: Remove the person label (and corresponding annotations):

datum transform -t remap_labels -- -l person: --default keep

Rename person to pedestrian and human to pedestrian, join annotations that had different classes under the same class id for pedestrian, don’t touch other classes:

datum transform -t remap_labels -- \
  -l person:pedestrian -l human:pedestrian --default keep

Rename person to car and cat to dog, keep bus, remove others:

datum transform -t remap_labels -- \
  -l person:car -l bus:bus -l cat:dog --default delete

project_labels

Changes the order of labels in the dataset from the existing to the desired one, removes unknown labels and adds new labels. Updates or removes the corresponding annotations.

Labels are matched by names (case dependent). Parent labels are only kept if they are present in the resulting set of labels. If new labels are added, and the dataset has mask colors defined, new labels will obtain generated colors.

Useful for merging similar datasets, whose labels need to be aligned.

Usage:

project_labels [-h] [-l DST_LABELS]

Optional arguments:

  • -h, --help (flag) - Show this help message and exit
  • -l, --label (str; repeatable) - Label name (ordered)

Examples: Set dataset labels to [person, cat, dog], remove others, add missing. Original labels (for example): cat, dog, elephant, human. New labels: person (added), cat (kept), dog (kept).

datum transform -t project_labels -- -l person -l cat -l dog

shapes_to_boxes

Converts spatial annotations (masks, polygons, polylines, points) to enclosing bounding boxes.

Usage:

shapes_to_boxes [-h]

Optional arguments:

  • -h, --help (flag) - Show this help message and exit

Example: Convert spatial annotations between each other

datum transform -t boxes_to_masks
datum transform -t masks_to_polygons
datum transform -t polygons_to_masks
datum transform -t shapes_to_boxes

boxes_to_masks

Converts bounding boxes to masks.

Usage:

boxes_to_masks [-h]

Optional arguments:

  • -h, --help (flag) - Show this help message and exit

polygons_to_masks

Converts polygons to masks.

Usage:

polygons_to_masks [-h]

Optional arguments:

  • -h, --help (flag) - Show this help message and exit

masks_to_polygons

Converts masks to polygons.

Usage:

masks_to_polygons [-h]

Optional arguments:

  • -h, --help (flag) - Show this help message and exit

anns_to_labels

Collects all labels from annotations (of all types) and transforms them into a set of annotations of type Label

Usage:

anns_to_labels [-h]

Optional arguments:

  • -h, --help (flag) - Show this help message and exit

merge_instance_segments

Replaces instance masks and, optionally, polygons with a single mask. A group of annotations with the same group id is considered an “instance”. The largest annotation in the group is considered the group “head”, so the resulting mask takes properties from that annotation.

Usage:

merge_instance_segments [-h] [--include-polygons]

Optional arguments:

  • -h, --help (flag) - Show this help message and exit
  • --include-polygons (flag) - Include polygons

crop_covered_segments

Sorts polygons and masks (“segments”) according to z_order, crops covered areas of underlying segments. If a segment is split into several independent parts by the segments above, produces the corresponding number of separate annotations joined into a group.

Usage:

crop_covered_segments [-h]

Optional arguments:

  • -h, --help (flag) - Show this help message and exit

bbox_value_decrement

Subtracts one from the coordinates of bounding boxes

Usage:

bbox_values_decrement [-h]

Optional arguments:

  • -h, --help (flag) - Show this help message and exit

5.23 - Utilities

Split video into frames

Splits a video into separate frames and saves them in a directory. After the splitting, the images can be added into a project using the import command and the image_dir format.

This command is useful for making a dataset from a video file. Unlike direct video reading during model training, which can produce different results if the system environment changes, this command allows to split the video into frames and use them instead, making the dataset reproducible and stable.

This command provides different options like setting the frame step (the -s/--step option), file name pattern (-n/--name-pattern), starting (-b/--start-frame) and finishing (-e/--end-frame) frame etc.

Note that this command is equivalent to the following commands:

datum create -o proj
datum import -p proj -f video_frames video.mp4 -- <params>
datum export -p proj -f image_dir -- <params>

Usage:

datum util split_video [-h] -i SRC_PATH [-o DST_DIR] [--overwrite]
  [-n NAME_PATTERN] [-s STEP] [-b START_FRAME] [-e END_FRAME] [-x IMAGE_EXT]

Parameters:

  • -i, --input-path (string) - Path to the video file
  • -o, --output-dir (string) - Output directory. By default, a subdirectory in the current directory is used
  • --overwrite - Allows overwriting existing files in the output directory, when it is not empty
  • -n, --name-pattern (string) - Name pattern for the produced images (default: %06d)
  • -s, --step (integer) - Frame step (default: 1)
  • -b, --start-frame (integer) - Starting frame (default: 0)
  • -e, --end-frame (integer) - Finishing frame (default: none)
  • -x, --image-ext (string) Output image extension (default: .jpg)
  • -h, --help - Print the help message and exit

Example: split a video into frames, use each 30-rd frame:

datum util split_video -i video.mp4 -o video.mp4-frames --step 30

Example: split a video into frames, save as ‘frame_xxxxxx.png’ files:

datum util split_video -i video.mp4 --image-ext=.png --name-pattern='frame_%%06d'

Example: split a video, add frames and annotations into dataset, export as YOLO:

datum util split_video -i video.avi -o video-frames
datum create -o proj
datum import -p proj -f image_dir video-frames
datum import -p proj -f coco_instances annotations.json
datum export -p proj -f yolo -- --save-images

5.24 - Validate Dataset

This command inspects annotations with respect to the task type and stores the results in JSON file.

The task types supported are classification, detection, and segmentation (the -t/--task-type parameter).

The validation result contains

  • annotation statistics based on the task type
  • validation reports, such as
    • items not having annotations
    • items having undefined annotations
    • imbalanced distribution in class/attributes
    • too small or large values
  • summary

Usage:

datum validate [-h] -t TASK [-s SUBSET_NAME] [-p PROJECT_DIR]
  [target] [-- EXTRA_ARGS]

Parameters:

  • <target> (string) - Target dataset revpath. By default, validates the current project.
  • -t, --task-type (string) - Task type for validation
  • -s, --subset (string) - Dataset subset to be validated
  • -p, --project (string) - Directory of the project to operate on (default: current directory).
  • -h, --help - Print the help message and exit.
  • <extra args> - The list of extra validation parameters. Should be passed after the -- separator after the main command arguments:
    • -fs, --few-samples-thr (number) - The threshold for giving a warning for minimum number of samples per class
    • -ir, --imbalance-ratio-thr (number) - The threshold for giving imbalance data warning
    • -m, --far-from-mean-thr (number) - The threshold for giving a warning that data is far from mean
    • -dr, --dominance-ratio-thr (number) - The threshold for giving a warning bounding box imbalance
    • -k, --topk-bins (number) - The ratio of bins with the highest number of data to total bins in the histogram

Example : give warning when imbalance ratio of data with classification task over 40

datum validate -p prj/ -t classification -- -ir 40

Here is the list of validation items(a.k.a. anomaly types).

Anomaly Type Description Task Type
MissingLabelCategories Metadata (ex. LabelCategories) should be defined common
MissingAnnotation No annotation found for an Item common
MissingAttribute An attribute key is missing for an Item common
MultiLabelAnnotations Item needs a single label classification
UndefinedLabel A label not defined in the metadata is found for an item common
UndefinedAttribute An attribute not defined in the metadata is found for an item common
LabelDefinedButNotFound A label is defined, but not found actually common
AttributeDefinedButNotFound An attribute is defined, but not found actually common
OnlyOneLabel The dataset consists of only label common
OnlyOneAttributeValue The dataset consists of only attribute value common
FewSamplesInLabel The number of samples in a label might be too low common
FewSamplesInAttribute The number of samples in an attribute might be too low common
ImbalancedLabels There is an imbalance in the label distribution common
ImbalancedAttribute There is an imbalance in the attribute distribution common
ImbalancedDistInLabel Values (ex. bbox width) are not evenly distributed for a label detection, segmentation
ImbalancedDistInAttribute Values (ex. bbox width) are not evenly distributed for an attribute detection, segmentation
NegativeLength The width or height of bounding box is negative detection
InvalidValue There’s invalid (ex. inf, nan) value for bounding box info. detection
FarFromLabelMean An annotation has an too small or large value than average for a label detection, segmentation
FarFromAttrMean An annotation has an too small or large value than average for an attribute detection, segmentation

Validation Result Format:

{
    'statistics': {
        ## common statistics
        'label_distribution': {
            'defined_labels': <dict>,   # <label:str>: <count:int>
            'undefined_labels': <dict>
            # <label:str>: {
            #     'count': <int>,
            #     'items_with_undefined_label': [<item_key>, ]
            # }
        },
        'attribute_distribution': {
            'defined_attributes': <dict>,
            # <label:str>: {
            #     <attribute:str>: {
            #         'distribution': {<attr_value:str>: <count:int>, },
            #         'items_missing_attribute': [<item_key>, ]
            #     }
            # }
            'undefined_attributes': <dict>
            # <label:str>: {
            #     <attribute:str>: {
            #         'distribution': {<attr_value:str>: <count:int>, },
            #         'items_with_undefined_attr': [<item_key>, ]
            #     }
            # }
        },
        'total_ann_count': <int>,
        'items_missing_annotation': <list>, # [<item_key>, ]

        ## statistics for classification task
        'items_with_multiple_labels': <list>, # [<item_key>, ]

        ## statistics for detection task
        'items_with_invalid_value': <dict>,
        # '<item_key>': {<ann_id:int>: [ <property:str>, ], }
        # - properties: 'x', 'y', 'width', 'height',
        #               'area(wxh)', 'ratio(w/h)', 'short', 'long'
        # - 'short' is min(w,h) and 'long' is max(w,h).
        'items_with_negative_length': <dict>,
        # '<item_key>': { <ann_id:int>: { <'width'|'height'>: <value>, }, }
        'bbox_distribution_in_label': <dict>, # <label:str>: <bbox_template>
        'bbox_distribution_in_attribute': <dict>,
        # <label:str>: {<attribute:str>: { <attr_value>: <bbox_template>, }, }
        'bbox_distribution_in_dataset_item': <dict>,
        # '<item_key>': <bbox count:int>

        ## statistics for segmentation task
        'items_with_invalid_value': <dict>,
        # '<item_key>': {<ann_id:int>: [ <property:str>, ], }
        # - properties: 'area', 'width', 'height'
        'mask_distribution_in_label': <dict>, # <label:str>: <mask_template>
        'mask_distribution_in_attribute': <dict>,
        # <label:str>: {
        #     <attribute:str>: { <attr_value>: <mask_template>, }
        # }
        'mask_distribution_in_dataset_item': <dict>,
        # '<item_key>': <mask/polygon count: int>
    },
    'validation_reports': <list>, # [ <validation_error_format>, ]
    # validation_error_format = {
    #     'anomaly_type': <str>,
    #     'description': <str>,
    #     'severity': <str>, # 'warning' or 'error'
    #     'item_id': <str>,  # optional, when it is related to a DatasetItem
    #     'subset': <str>,   # optional, when it is related to a DatasetItem
    # }
    'summary': {
        'errors': <count: int>,
        'warnings': <count: int>
    }
}

item_key is defined as,

item_key = (<DatasetItem.id:str>, <DatasetItem.subset:str>)

bbox_template and mask_template are defined as,

bbox_template = {
    'width': <numerical_stat_template>,
    'height': <numerical_stat_template>,
    'area(wxh)': <numerical_stat_template>,
    'ratio(w/h)': <numerical_stat_template>,
    'short': <numerical_stat_template>, # short = min(w, h)
    'long': <numerical_stat_template>   # long = max(w, h)
}
mask_template = {
    'area': <numerical_stat_template>,
    'width': <numerical_stat_template>,
    'height': <numerical_stat_template>
}

numerical_stat_template is defined as,

numerical_stat_template = {
    'items_far_from_mean': <dict>,
    # {'<item_key>': {<ann_id:int>: <value:float>, }, }
    'mean': <float>,
    'stddev': <float>,
    'min': <float>,
    'max': <float>,
    'median': <float>,
    'histogram': {
        'bins': <list>,   # [<float>, ]
        'counts': <list>, # [<int>, ]
    }
}

6 - Extending

There are few ways to extend and customize Datumaro behavior, which is supported by plugins. Check our contribution guide for details on plugin implementation. In general, a plugin is a Python module. It must be put into a plugin directory:

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

Built-in plugins

Datumaro provides several builtin plugins. Plugins can have dependencies, which need to be installed separately.

TensorFlow

The plugin provides support of TensorFlow Detection API format, which includes boxes and masks.

Dependencies

The plugin depends on TensorFlow, which can be installed with pip:

pip install tensorflow

or

pip install tensorflow-gpu

or

pip install datumaro[tf]

or

pip install datumaro[tf-gpu]

Accuracy Checker

This plugin allows to use Accuracy Checker to launch deep learning models from various frameworks (Caffe, MxNet, PyTorch, OpenVINO, …) through Accuracy Checker’s API.

Dependencies

The plugin depends on Accuracy Checker, which can be installed with pip:

pip install 'git+https://github.com/openvinotoolkit/open_model_zoo.git#subdirectory=tools/accuracy_checker'

To execute models with deep learning frameworks, they need to be installed too.

OpenVINO™

This plugin provides support for model inference with OpenVINO™.

Dependencies

The plugin depends on the OpenVINO™ Toolkit, which can be installed by following these instructions

Dataset Formats

Dataset reading is supported by Extractors and Importers. An Extractor produces a list of dataset items corresponding to the dataset. An Importer creates a project from the data source location. It is possible to add custom Extractors and Importers. To do this, you need to put an Extractor and Importer implementation scripts to a plugin directory.

Dataset writing is supported by Converters. A Converter produces a dataset of a specific format from dataset items. It is possible to add custom Converters. 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.

7 - Links

8 - How to control telemetry data collection

The OpenVINO™ telemetry library is used to collect basic information about Datumaro usage.

A short description of the information collected:

Event Description
version Datumaro version
session start/end Accessory event, there is no additional info here
{cli_command}_result Datumaro command result with arguments passed*
error Stack trace in case of exception*

* All sensitive arguments, such as filesystem paths or names, are sanitized

To enable the collection of telemetry data, the ISIP consent file must exist and contain 1, otherwise telemetry will be disabled. The ISIP file can be created/modified by an OpenVINO installer or manually and used by other OpenVINO™ tools.

The location of the ISIP consent file depends on the OS:

  • Windows: %localappdata%\Intel Corporation\isip,
  • Linux, MacOS: $HOME/intel/isip.