FiftyOne Release Notes

FiftyOne Teams 1.3.0

Released May 31, 2023

Includes all features from FiftyOne 0.21.0, plus:

General

• Added a Management SDK subpackage for programmatically configuring user roles, dataset permissions, plugins, and more

• Added support for authenticated API connections when using the Python SDK that respect user roles, dataset permissions, etc

• Logins now automatically redirect back to the page you were trying to access

• Improved non-persistent dataset cleanup behavior

• Fixed a bug that could cause the media cache to erroneously garbage collect large files while they are downloading

• Fixed a bug when cloning views into new datasets via the Teams UI

Admin

• Added support for uploading and managing plugins via the Teams UI

• Added support for cross account IAM roles when configuring cloud storage credentials

• Fixed a bug that prevented Azure/MinIO credentials uploaded via the Teams UI from being properly recognized by the App

FiftyOne 0.21.0

Released May 31, 2023

App

• Added support for viewing and executing operators in the App! #2679

• Added support for creating dynamic groups in the App #2934

• Added support for overlaying multiple point cloud slices in Looker3D #2912

• Added support for customizing the App color scheme via a new color scheme modal #2824

• Added support for configuring field visibility in the App’s sidebar #2924, #3024

• Added support for visualizing Label fields stored within top-level embedded document fields #2885

• Optimized App loading for datasets with large sample documents #3139

• Optimized App routes that involve synchronous computations #3066

• Fixed a URL filepath bug that could cause orthographic projections to fail to render #3122

• Fixed a layout bug when working with long brain keys in the Embeddings panel #3026

• Added a welcome message that displays when the App is launched for the first time with a new FiftyOne version #3092

Core

• Added support for creating dynamic grouped views #2475

• Added support for storing default color schemes for datasets #2824

• Added support for selecting/excluding fields via dynamically defined filters via select_fields() and exclude_fields() #2898

• Added support for evaluating keypoints #2776, #2928

• Added support for computing DICE score when evaluating segmentations #2777, #2901

• Added a new list_schema() aggregation for inferring the contents of nested list fields #2882

• Added support for declaring dynamic nested list fields #2882

• Handling missing label fields when deleting labels #2918

• Only match .txt files when reading YOLO labels #3127

• Improved behavior of transform_images() and transform_videos() utilities when processing media in-place #2931

• Added utils and helpful warnings that advise how to patch broken saved views and runs #2970, #2971

• Replaced pkg_resources with importlib.metadata #2930

Plugins

• Added Operators to the plugin framework #2679

• Added CLI methods for plugins and operators #3025, #3038

Annotation

• Added support for CVAT 2.4 #2959

• Added support for importing/exporting instances when using the Label Studio integration #2706, #2917

• Added support for importing multiclass classifications from Scale #3117

• Updated Scale integration to assume that imported line annotations are not closed shapes #3123

• Fixed broken Scale docs links and unlabeled annotation task support #2916

Zoo

• Added the Sama-COCO dataset to the zoo! #2904

Tutorials

• Updated detection mistakes tutorial to avoid unnecessarily resetting the App #3034

FiftyOne Teams 1.2.1

Released April 5, 2023

Includes all features from FiftyOne 0.20.1, plus:

General

• When your session expires, you are now automatically logged out rather than being presented with a cryptic server error

• Improved the accuracy of size estimates when exporting filepaths and/or tags from the Teams UI

Admin

• Added support for uploading Azure storage credentials for your deployment via the Settings > Cloud storage page

SDK

• Added support for working with media in Azure cloud storage. Refer to this section to see how to provide your storage credentials

Deployment

• Added support for deploying into Microsoft Azure environments

• Fixed a bug that prevented the dataset page from loading for deployments running MongoDB 4.4

FiftyOne 0.20.1

Released April 5, 2023

App

• Added support for storing datetimes as field metadata and viewing them in the App’s field tooltip #2861

• Fixed a bug when pulling color-by data for sample embeddings plots when viewing patches in the sample grid #2846

• Fixed a bug that prevented the sample grid from refreshing when composing multiple sidebar filters #2849

• Fixed a bug that prevented field-specific mask targets from being recognized when rendering segmentations in the App #2879

• Fixed a bug when rendering heatmaps stored as images on disk #2872, #2880

Core

• Added support for dynamically inferring fields on embedded lists and documents #2863, #2882

• Added support for listing datasets matching a glob pattern #2868

• Improved the robustness of merge_samples() when cleaning up after a failed merge #2844

• Using new libraries for ndjson and archive extraction #2864

• Fixed a bug that prevented text similarity searches searches from succeeding when GPU is available #2853

• Fixed a bug where stats() would report the wrong size for dataset views that select/exclude fields on MongoDB 5.2 or later #2840

• Fixed a bug with dynamic schema expansion of list fields #2855

• Fixed a bug when merging video samples into a grouped dataset that did not previously contain videos #2851

• Fixed a validation bug when importing COCO datasets whose description is not a string #2848

Documentation

• Updated the source URLs for the Caltech-101 and Caltech-256 datasets #2841

• Fixed a typo in the Caltech-256 dataset documentation #2842

FiftyOne Teams 1.2

Released March 22, 2023

Includes all features from FiftyOne 0.20.0, plus:

Admin settings

• Admins who use SSO to authorize new users to auto-join their FiftyOne Teams deployment can now configure the default role for those users

• Admins can now configure the default access level that Members receive on newly created datasets Dataset page

Dataset page

• Added support for viewing Segmentation and Heatmap data stored as images in the cloud in the App

• Added support for exporting one or more fields of a dataset in CSV format through the Teams UI

• Stack traces for unhandled errors are now presented directly in the App so that users can self-diagnose issues

Deployment

• Added support for sharded databases

FiftyOne 0.20.0

Released March 22, 2023

News

• Added support for querying by arbitrary text prompts in the App! #2633

• Released a Qdrant integration for native text and image searches on FiftyOne datasets!

• Released a Pinecone integration for native text and image searches on FiftyOne datasets!

App

• Switched the default sidebar mode to fast #2714

• Refactored sample/label tags in the App so that they are treated the same as any other list field #2557

• Added support for visualizing orthographic projection images for point cloud datasets/slices #2660

• Added a filter/selection indicator to the title of all Panels that can be clicked to clear the Panel’s current state #2652

• Any selection state associated with a Panel is now automatically cleared when the Panel is closed #2652

• Added a button to the saved view selector for clearing the current view #2661

• Added support for maximizing/hiding individual panels of the grouped modal #2688

• Added support for switching between multiple point cloud slices #2675

• Added keyboard shortcuts for opening Panels directly in split mode #2663

• Upgraded Looker3D controls #2753

• Upgraded the modal’s JSON viewer #2677

• Selected labels are not reset after applying a similarity search #2820

• Stack traces for unhandled errors are now presented directly in the App so that users can self-diagnose issues #2795, #2797

• Improved error handling when loading invalid/missing brain results in the Embeddings panel #2651, #2790

• More intuitive behavior when combining Embedding panel selections and sidebar filters #2741

• Ensure that URL is updated when loading saved views via a Python session #2740

• Switched to wildcard-based string matching in the sidebar #2736

• Plugins can now load components and utilities from runtime instead of compiling their own #2680

• Stability improvements when loading and handling errors in plugins #2758

• Informative error messages are now displayed when visualization results fail to load in the Embeddings panel #2751

• Resolved some edge cases when loading views with different schemas via Python sessions #2730

• Fixed a bug that would cause saving views to intermittently fail #2667

• Fixed a bug when using saved views with Python <3.9 in the App #2676, #2728

• Fixed a bug that could cause App crashes when loading SelectGroupSlices stages in the view bar #2669, #2743

• Fixed a bug that could cause App crashes when filtering keypoints #2774, #2779

• Fixed a bug when lassoing patch embeddings with the Map panel open #2754

• Fixed inconsistencies with selection, tagging, active slices, and sidebar stats in the modal for grouped datasets #2785, #2782, #2769, #2759, #2749, #2731

• Fixed a bug when pressing enter twice in a label tag popover #2757

• Fixed a bug where keyboard listeners in the modal would interfere with other input interactions #2786

• Fixed a bug where some users would see erroneous scrollbars #2794

• Fixed bugs when tagging labels in the grouped modal #2820

• Fixed a bug when retrieving values for filter dropdowns in the grouped modal #2817

• Fixed a bug that would raise an App error after deleting certain saved views #2801

• Fixed the formatting of the support field in the modal sidebar for clip views #2800

• Fixed a bug with URL rendering in the sidebar #2735

• Fixed a bug when streaming filtered frame labels #2682, #2733

• Fixed a bug when adding new tags to a selected sample or label #2703

• Fixed a bug when matching by tags that contain spaces #2658

Core

• Added support for querying by vectors and text prompts #2569

• Upgraded the similarity index interface, including Qdrant and Pinecone support, and the ability to add/remove embeddings to an existing index #2792

• Added support for storing and visualizing cuboids and rotated bounding boxes in the App #2296

• Added support for evaluating 3D object detections #2486

• Added a to_trajectories() view stage #1300

• Added support for generating orthographic projection images for point cloud datasets/slices #2656

• Added validation to set_values() #2770

• Frame collections are now lazily created only when necessary #2727

• Upgraded the document save implementaton to only use upsert operations when explicitly required #2727

• Added _dataset_id to all sample/frame documents in datasets #2711

• Added a save() and save_config() methods to RunResults #2696, #2772

• Added support for renaming existing runs via new rename_annotation_run(), rename_brain_run(), and rename_evaluation() methods #2696

• Added support for filtering by run type and config parameters when using list_annotation_runs(), list_brain_runs(), and list_evaluations() #2696, #2772

• Added an add_group_slice() method to declare new slices on grouped datasets #2727

• Added support for controlling whether saved views and runs are imported/exported in FiftyOneDataset format #2806

• Added support for negative integer mask targets #2686

• Downward migrations for future-but-compatible versions of FiftyOne are now skipped rather than raising an error #2683

• Fixed a bug when cloning datasets with run results #2772

• Fixed a bug with the dynamic=True syntax for declaring dynamic fields on list documents #2767

• Fixed a bug in deferred saves where filtered list updates were not being applied #2727

Annotation

• Added support for passing CVAT organization to annotation jobs #2716

Docs

• Added documentation for working with point cloud-only datasets #2724

• Added documentation for on-the-fly custom embedded document creation #2687

• Fixed broken torchvision dataset links in the docs #2771

Zoo

• Added a tensorflow-macos option when loading TF models from the Model Zoo #2685

Tutorials

• Added a Point-E tutorial showcasing the 3D Visualizer’s capabilities in the context of building a 3D self-driving dataset #2818

• Added a YOLOv8 tutorial #2755

• Updated the media in the Open Images tutorial #2665

FiftyOne Teams 1.1.1

Released February 14, 2023

Includes all features from FiftyOne 0.19.1, plus:

Plugins

• Resolved a bug that prevented Teams deployments from recognizing installed plugins

FiftyOne 0.19.1

Released February 14, 2023

App

• Fixed a bug when launching the App in Python 3.8 or earlier #2647

• Fixed a bug that prevented launching the App in Databricks notebooks #2647

Core

• Fixed a bug in certain environments that prevented progress bars from rendering correctly #2647

FiftyOne Teams 1.1

Released February 9, 2023

Includes all features from FiftyOne 0.19.0, plus:

User roles

• Renamed the existing Guest role to Collaborator

• Added a new Guest role. Note that Guest is a view-only role and does not contribute to your license count. You can add unlimited Guest users to your deployment!

Homepage

• Added a Recent views widget to the homepage that shows the most recent saved views that you have viewed in the Teams UI

Dataset page

• Added support for cloning the current view (including any filters, selections, etc) into a new dataset from the UI

• Added support for exporting the current view to local disk or a cloud bucket in various formats (filepaths only, filepaths and tags, media only, labels only, media and labels)

Deployment

• Added support for deploying Teams into environments with proxy networks

FiftyOne 0.19.0

Released February 9, 2023

News

• FiftyOne Teams documentation is now publicly available! #2388

App

• Added the Spaces framework #2524

• Added native support for visualizing embeddings #2524

• Refactored the map tab into a dedicated map panel #2524

• Refactored the histograms tab into a dedicated histograms panel #2524

• Added support for loading and saving views #2461

• Added support for visualizing Segmentation and Heatmap masks stored on disk #2358

• Added support for visualizing RGB segmentations #2483

• Added retries for all network requests to improve stability #2406

• Optimized the tagging menu #2368

• Optimized sample tagging on video datasets #2440

• Don’t refresh the background grid when applying tags in the modal #2594

• Only show supported keys in the evaluations dropdown #2427

• Fixed handling of None values when filtering numeric/list fields #2422, #2412, #2403

• Never show expanded filter list for ID fields #2408

• Ensure that the bookmark icon displays when extended selections exist #2366

• Automatically clear sample selection after sorting by similarity #2595

• Use consistent loading dots throughout the App #2321

• Fixed a bug when filtering by custom embedded list fields #2407

• Fixed bugs when screenshotting the App in notebook contexts #2398

• Fixed bugs when launching the App in Databricks notebooks #2397

• Show metadata for frame-level fields in the fields tooltip #2386

• Fixed bugs when configuring plugin settings and modal media fields #2383

• Fixed bugs with multiple media fields when loading views that exclude fields #2378, #2303

Core

• Added support for programmatically configuring space layouts #2524

• Added support for loading and saving views #2461

• Added support for storing Segmentation and Heatmap masks on disk #2301

• Added support for RGB segmentations in evaluate_segmentations() #2483

• Added a new transform_segmentations() utility #2483

• Added support for declaring dynamic fields on generated views via set_values() #2513

• Added support for importing and exporting datasets in CSV format #2616, #2450

• Added support for importing and exporting directories of arbitrary media files #2605

• Added a dedicated clear_cache() method for clearing a dataset’s run cache #2471

• Updated all plotting methods, eg scatterplot() to always rely on sample/label IDs when pulling data for plots #2614

• Updated compute_patch_embeddings() to store patch embeddings directly on Label objects when the embeddings_field argument is provided #2626

• Added support for passing frame-level fields directly to export() #2418

• Added an optional dynamic=True flag to set_values() #2372

• Added support for declaring custom Label attributes via set_values() #2372

• Adds a new set_label_values() utility for setting attributes on Label instances by their IDs #2372

• Always update dataset’s last_loaded_at property when they are loaded #2375

• Migrated runs to a separate database collection, for efficiency #2189

• Added an exact_frame_count() utility for computing extact video frame counts #2373

• Updated the 3D visualizer to use true centroid (not bottom-center) coordinates for 3D detections #2474

• Added support for loading specific group slice(s) when using iter_groups() and get_group() #2528

• Added an exclude_groups() view stage #2451

• Added support for importing annotations directly on grouped datasets #2349

• Added a group_collections() utility for merging multiple collections into a grouped dataset #2332

• Added support for converting an existing dataset into a grouped dataset via set_values() #2332

• Added support for deleting grouped fields when the dataset contains only one media type #2332

• Updated Dataset.stats() to include media from all slices of grouped datasets #2635

• Fixed a bug when calling to_frames() on a view that filters the frames of the input dataset #2361

• Fixed some bugs when passing multiple aggregations with the same field name and type to aggregate() #2617

• Fixed a bug when manually unwinding list fields in aggregations #2608

• Fixed a bug when loading datasets with CVAT attributes stored in VOC format #2359

• Fixed a bug in default sidebar group expansion #2441

Annotation

• Added support for CVAT 2.4 #2597

• Added support for providing custom task names for CVAT tasks #2353

• Fixed a bug when checking if CVAT projects exist #2491

• Fixed a bug when checking if CVAT tasks exist #2070

Zoo

• Added Open Images V7 to the zoo #2446

• Updated the KITTI multiview and quickstart-groups datasets to not use legacy 3D visualizer settings #2474

• Added support for filtering datasets when using list_zoo_datasets() #2448

Docs

• Added detailed plugin documentation #2442

• Added documentation for converting between common label formats #2498

• Added a pandas vs FiftyOne tutorial #2310

• Added a pandas vs FiftyOne cheat sheet #2329

• Added a FiftyOne terminology cheat sheet #2484

• Added a view stage cheat sheet #2452

• Added a filtering cheat sheet #2447

FiftyOne Teams 1.0

Released November 8, 2022

Includes all features from FiftyOne 0.18.0, plus:

News

• FiftyOne Teams is now generally available, read more here!

FiftyOne 0.18.0

Released November 8, 2022

App

• Significantly optimized the performance of the sidebar by lazily computing statistics only for currently visible fields #2191

• Added new sidebar modes with updated default behavior that further optimizes the performance of the App for large datasets #2191

• Added support for configuring the sidebar mode dynamically in the App and programmatically on a per-dataset basis #2191

• Added support for programmatically configuring sidebar groups and default expansion states on a per-dataset basis #2190

• Added support for viewing field-level descriptions via a new field tooltip #2216

• Added support for filtering by and viewing stats for custom embedded document attributes #1825

• Added a new light mode option! #2156

• Improved responsiveness of the sidebar when toggling fields on and off #2247

• Improved responsiveness and state management of the view bar #2230

• Restored the ability to shift-select multiple samples in the grid view #2110

• Fixed an issue that could cause unselected label fields to be inadvertently tagged when using the label tagging UI #2121

• Fixed an issue that would prevent label tags applied on patch views in the tagging UI from persisting to the underlying dataset #2113

• Fixed an issue that could arise when loading a group dataset with sparse alternate media fields #2164

• Fixed some issues with datetime rendering and timezone handling #2111, #2112

Core

• Added support for declaring custom dynamic attributes on datasets! #1825

• Added support for storing field-level metadata on datasets #2216

• Added native support for installing on Apple Silicon with MongoDB 6 #2165

• Dataset creation using default naming is now multiprocess-safe #2097

• Optimized the implementation of tagging samples and labels #2203, #2208

• Optimized the implementation of select(), select_by(), and select_groups() when performing ordered selections #2227

• Updated the logic of exists() to be more intuitive for frame fields #2209

• Upgraded server and MongoDB requirements to pymongo>=3.11, motor>=2.3 and newer pinned versions of mongoengine, starlette, and strawberry-graphql #2215

• Added support for modifying the filepaths of a frame view #2193

• Improved the implementation of merge_samples() and related methods to safely cleanup in case of failed merges #2135

• Fixed some bugs that could occur when creating frame views into grouped collections #2144

• Fixed a bug when using select_by() with ObjectId fields #2140

• Added an option to import annotation IDs when loading data stored in COCO format #2122

• Added support for including the export directory in the dataset.yaml file generated by YOLOv5 exports #2114

Annotation

• Updated the default CVAT endpoint to https://app.cvat.ai #2228

• Fixed a bug that would cause annotation runs involving unlabeled samples to crash when using the Label Studio backend #2145

Zoo

• Added support for using CUDA devices when running the CLIP model from the zoo #2201

FiftyOne 0.17.2

Released September 20, 2022

App

• Fixed a backward compatability bug when connecting to older database verions #2103

FiftyOne 0.17.1

Released September 20, 2022

Core

• Removed TypedDict usage introduced in v0.17.0 that is not supported in Python 3.7 #2100

FiftyOne 0.17.0

Released September 19, 2022

App

• Added support for visualizing grouped datasets in the App #1765

• Added support for visualizing point cloud samples in the modal #1765

• Added support for visualizing and interacting with GeoLocation data in a new Map panel #1976

• Added initial support for custom App plugins #1765

• Added support for configuring multiple media fields #1765

• Fixed Google Colab screenshotting and cell updates #2069

Core

• Added support for grouped datasets, e.g., multiple camera view scenes #1765

• Added support for point cloud samples in grouped datasets #1765

• Added an app_config property to datasets for configuring App behavior on a per-dataset basis #1765

• Added an optional rel_dir parameter to export() and draw_labels() #2060

• Added an optional abs_paths=True option to export() #2060

• Added an optional use_dirs=True parameter to export() that causes metadata to be exported in per-sample/frame JSON files #2028

• Updated the COCO importer to load all available label types by default #1869

• Fixed a bug when passing ordered=True to select_by() #2059

• Fixed an error that would occur when storing custom embedded documents on dynamic label attributes #2051

• Fixed a match_frames() bug that caused all frames to be included, even if the view filters the frames #2029

Docs

• Added a tutorial showing how to integrate FiftyOne into a Detectron2 model training pipeline #2054

Annotation

• Fixed a bug that occured when checking if tasks exist on CVAT v2 servers #2070

• Fixed an error that occured when deseriailizing Label Studio annotation results #2074

Zoo

• Added clip-vit-base32-torch to the model zoo! #2072

• Added the Quickstart Groups dataset to the dataset zoo! #1765

• Added the KITTI Multiview dataset to the dataset zoo! #1765

FiftyOne 0.16.6

Released August 25, 2022

App

• Fixed a bug that caused the App to break when sample tags contained . #1924

• Fixed search results alignment #1930

• Fixed App refreshes after view changes had occurred from the view bar #1931

• Fixed mask targets rendering in the tooltip #1943 #1949

• Fixed classification confusion matrix connections #1967

Core

• Added a save context that enables efficient batch edits of datasets and views #1727

• Added Plotly v5 support #1981

• Added a quantiles aggregation #1937

• Added support for writing transformed images/videos to new locations in the transform_images() and transform_videos() functions #2007

• Added support for configuring the package-wide logging level #2009

• Added more full-featured support for serializing and deserializing datasets, views, and samples via to_dict() and from_dict() #1922

• Added support for dynamic attributes when performing coerced exports #1993

• Introduced the notion of client compatability versions #2017

• Extended stats() to all sample collections #1940

• Added support for serializing aggregations #1911

• Added weighted_sample() and balanced_sample() utility methods #1925

• Added an optional new_ids=True option to Dataset.add_collection() that generates new sample/frame IDs when adding the samples #1927

• Added support for the path variable in dataset.yaml of YOLOv5 datasets #1903

• Fixed a bug that prevented using set_values() to set frame-level label fields #1922

• Fixed automatic declaration of frame fields when computing embeddings on a frame view #1922

• Fixed a regression that caused label ID fields to be returned as ObjectID #1922

• Fixed a bug that allowed default frame fields to be excluded #1922

• ClipsView instances will now report their metadata type as VideoMetadata #1922

• Fixed load_evaluation_view() when select_fields is True #1922

• Fixed boolean field parsing when declaring fields #1922

• Fixed a bug that caused nested embedded documents to corrupt datasets #1922

• Fixed a bug that prevented assignment of array-valued dynamic attributes to labels #1922

Annotation

• Added a new Label Studio integration! #1848

• Optimized loading CVAT annotations and performing operations on CVATAnnotationResults #1944

• Upgraded the AnnotationAPI interface #1997

• Fixed loading group IDs in CVAT video tasks #1917

• Fixed uploading to a CVAT project when no label schema is provided #1926

FiftyOne 0.16.5

Released June 24, 2022

App

• Fixed dataset selection searches #1907

• Fixed dataset results for long dataset names #1907

FiftyOne 0.16.4

Released June 21, 2022

App

• Fixed frame fields omission in the sidebar #1899

FiftyOne 0.16.3

Released June 20, 2022

App

• Added hotkey to hide overlays while pressed #1779

• Changed expanded view ESC sequence to reset zoom before frame scrubbing #1810

• Fixed the expanded view tooltip when a keypoint has nan point(s) #1828

• Fixed initial loading of keypoint skeletons #1828

• Fixed Classifications rendering in the grid #1828

• Fixed App loads for environments with old (<=v0.14.0) datasets that have yet to be migrated #1829

• Fixed spurious loading states from tagging in the expanded view #1834

• Fixed a bug that caused frame classifications to be incorrectly rendered in the grid #1877

• Fixed active (checked) field persistence in the grid when changing views #1878

• Fixed views and actions that contain BSON #1879

• Fixed JSON rendering in the expanded view for nested data #1880

• Fixed selection and expansion for bad media files #1882

• Fixed Other plot tab date and datetime fields with None values #1817

• Increased results from 10 to 200 for search selectors #1875

• Fixed App issues related to dataset deletion and dataset schema changes #1875

Core

• Added skeleton and skeleton_key to the OpenLABEL image and video importers #1812

• Fixed a database field issue in clone_frame_field() and clone_sample_field(), #1824

• Fixed using zoo models with the newest version of Torchvision #1838

• Added classifications_to_detections() for converting classifications to detections #1842

• Set forking as the default for macOS multiprocessing #1844

• Added dataset.tags for organizing datasets #1845

• Added functionality to explicitly define classes for evaluation methods #1858

• Fixed tfrecord shard enumeration, i.e. zero indexing #1859

• Added support for label field dicts when importing labeled datasets #1864

• Removed non-XML or non-TXT files from CVAT, KITTI, CVATVideo #1884

Annotation

• Updated CVAT task and project processing #1839

• Added the ability to upload and download group ids from CVAT #1876

FiftyOne 0.16.2

Released June 2, 2022

App

• Added explicit error handling when FFmpeg is installed so it is made clear to the user that it must be installed to use video datasets in the App #1801

• Fixed range requests for media files, e.g. mp4s, on the server #1786

• Fixed tag rendering in the grid #1808

• Fixed tagging selected labels in the expanded view #1808

• Fixed session.view = None #1808

• Fixed issues with patches views #1808

Core

• Fixed errors related to session-attached plots #1808

FiftyOne 0.16.1

Released May 26, 2022

App

• Fixed a bug that caused label rendering to be delayed until statistics were loaded #1776

• Fixed the v0.16.0 migration that prevents label lists, e.g. Detections from showing their label filters when expanded in the sidebar #1785

• Fixed expanded samples in clips views which appeared to be empty #1790

• Fixed “Sort by similarity” with a dist_field #1790

• Fixed “Color by” for simple values (classifications, tags, etc.) #1790

• Fixed changing datasets when sort by similarity is set #1790

• Fixed mask and map coloring #1790

• Fixed fortran array handling for masks and maps #1790

Core

• Fixed a formatting issue when raising an exception because unsupported plotting backend was requested #1794

FiftyOne 0.16.0

Released May 24, 2022

App

• Added routing, e.g. /datasets/:dataset-name #1713

• Redesigned the sidebar to support custom grouping and sorting of fields and tags #1713

• Added graceful handling of deleted datasets in the App #1713

• Fixed epoch rendering #1713

• Fixed empty heatmap rendering #1713

• Added stack traces to the new error page #1713

• Fixed ESC when viewing single frame clips #1713

• Fixed handling of unsupported videos #1713

• Added support for opening the expanded view while sample(s) are selected #1713

• Fixed keypoint skeleton rendering for named skeletons of frame fields #1713

Core

• Fixed edge cases in clone_frame_field(), merge_samples(), and rename_frame_field() #1749

• Fixed a bug that would cause non-persistent datasets to be prematurely deleted #1747

• Fixed loading relative paths in YOLOv5 format #1721

• Fixed image lists for the image_path parameter when importing GeoTIFF datasets #1728

• Added a find_duplicates() utility to automatically find duplicate objects based on IoU #1714

FiftyOne 0.15.1

Released April 26, 2022

App

• Added support for rendering keypoint skeletons #1601

• Added support for rendering per-point confidences and other custom per-point attributes on Keypoint objects #1601

• Added support for rendering Fortan-ordered arrays #1660

Core

• Added support for storing keypoint skeletons on datasets #1601

• Added a filter_keypoints() stage that applies per-point filters to Keypoint objects #1601

• Added support for rendering keypoints skeletons and missing keypoints to draw_labels() #1601

• Added support for per-point confidences and other custom per-point attributes on Keypoint objects. See this section for details #1601

• Added a concat() view stage that allows for concatenating one collection onto another #1662

• Non-persistent datasets are now automatically deleted when using a custom database_uri #1697

• Added a database_admin config setting that can control whether database migrations are allowed. See this page for details #1692

• Added a database_name config setting that allows for customizing the MongoDB database name #1692

• Classification attributes are now exported as tag attributes when exporting in CVATImageDataset format #1686

• The iscrowd attribute is now always populated when exporting in COCO format #1664

• Fixed a KeyError bug when loading dataset with relative paths on Windows #1675

Brain

• Added fiftyone-brain wheels for Python 3.10

• Added support for installing fiftyone-brain on Apple Silicon

Annotation

• Fixed a CSRF Failed error when connecting to some CVAT servers #1668

Integrations

• Updated the Lightning Flash integration to support Flash versions 0.7.0 or later #1671

Zoo

• Added the Families in the Wild dataset to the FiftyOne Dataset Zoo! #1663

FiftyOne 0.15.0

Released March 23, 2022

App

• Fixed Regression rendering in the visualizer #1604

Core

• Added a Dataset.delete_frames() method that allows for deleting frames by ID #1650

• Added a keep_fields() method to DatasetView and its subclasses #1616

• Added a lines() method that allows for plotting lines whose scatter points can be interactively selected via the typical interactive plotting workflows #1614

• Added an optional force_rgb=True syntax when importing/exporting/creating TF records using all relevant methods in fiftyone.utils.tf #1612

• Added support for passing additional kwargs to the fiftyone convert CLI command #1612

• Added support for annotating video-level labels when using draw_labeled_videos() #1619

• Added the ability to slice using a ViewField #1630

• Fixed bug in from_images_dir() where attempting to load 4-channel images errored even if force_rgb=True #1632

• Fixed a bug that prevented frames from being attached to video collections when aggregating expressions that involve both Sample-level and Frame-level fields #1644

• Added support for importing image and video datasets in OpenLABEL format #1609

Annotation

• Added support for CVATv2 servers when using the CVAT backend #1638

• Added an issue_tracker argument to annotate() when using the CVAT backend #1625

• Added a dest_field argument to load_annotations() which allows you to specify the name of the field to which to load annotations #1642

• Added a property to annotation backends that decides whether to allow annotation of video-level labels #1655

• Fixed a bug where views that dynamically modify label strings would result in labels not being uploaded to the annotation backend #1647

Docs

• Added documentation for defining custom EmbeddedDocument and DynamicEmbeddedDocument classes #1617

• Added documentation about boolean view indexing to the user guide #1617

• Added a recipe for creating views and view expressions #1641

FiftyOne 0.14.4

Released February 7, 2022

News

• With support from the ActivityNet team, FiftyOne is now a recommended tool for downloading, visualizing, and evaluating on the Activitynet dataset! Check out this guide for more details

App

• Fixed encoding of sample media URLs so image and video filepaths with special characters are supported

• Fixed an error that would occur when rendering empty Keypoint instances

Core

• Added an official Dockerfile

• Changed the default implementation of to_frames() to assume that the user has already sampled the frames offline and stored their locations in a filepath field of each Frame in their video dataset. See this section for more details

• Updated DatasetView.save() to save changes to (only) the samples in the view to the underlying dataset

• Added a new DatasetView.keep() method that deletes any samples that are not in the view from the underlying dataset

• Added InteractivePlot.save() and ViewPlot.save() methods that can be used to save plots as static images

• Added support for populating query distances on a dataset when using sort_by_similarity() to query by similarity

• Added a instances_to_polylines() utility that converts instance segmentations to Polylines format

• Added support for frame labels to all conversion methods in the fiftyone.utils.labels module

• Updated the implementation of Detection.to_polyline() so that all attributes are included rather than just ETA-supported ones

• Added support for including empty labels labels via an include_missing keyword argument in add_yolo_labels()

• Added a download_youtube_videos() utility for efficiently and robustly downloading videos or specific segments from YouTube

• Added a skip_failures flag to transform_images() and transform_videos()

• Added shuffle and seed parameters to FiftyOneImageLabelsDatasetImporter and FiftyOneVideoLabelsDatasetImporter

• Added an include_all_data parameter to YOLOv5DatasetImporter

• Resolved a bug that would previously cause an error when writing aggregations on video datasets that involve applying expressions directly to "frames"

Annotation

• Added support for importing and exporting sample-level tags in CVAT format

• Fixed a bug that prevented existing label fields such as Detections that can contain multiple annotation types (boxes or instances) from being specified in calls to annotate()

• CVAT login credentials are no longer included in exception messages

Zoo

• Added ActivityNet 100 to the dataset zoo!

• Added ActivityNet 200 to the dataset zoo!

• Added Kinetics 400 to the dataset zoo!

• Added Kinetics 600 to the dataset zoo!

• Added Kinetics 700 to the dataset zoo!

• Added Kinetics 700-2020 to the dataset zoo!

FiftyOne 0.14.3

Released January 13, 2022

Core

• Added hollow support for 32-bit systems (a database_uri must be used in such cases)

• Added support for indexing into datasets using boolean arrays or view expressions via new dataset[bool_array] and dataset[bool_expr] syntaxes

• Added support for registering custom EmbeddedDocument classes that can be used to populate fields and embedded fields of datasets

• Added support for importing and exporting confidence in YOLO formats

• Added support for directly passing a filename -> filepath mapping dict to the data_path parameter to dataset importers

• Added graceful casting of int-like and float-like values like np.float(1.0) to their respective Python primitives for storage in the database

• Changed the default to num_workers=0 when using methods like apply_model() to apply Torch models on Windows, which avoids multiprocessing issues

• Fixed a bug when calling evaluate_detections() with both the classes and compute_mAP=True arguments provided

• Fixed a bug that could arise when importing segmentation data from a COCO JSON that contains objects with [] segmentation data

• Fixed a bug in expressions containing near-epoch dates

• Added support for setting frame-level fields by passing frame number dicts to set_values()

• Fixes a bug that prevented set_values() from working as expected when key_field="id" argument is used

• Fixed a bug that occurred when computing patch embeddings defined by polylines

• Added decision thresholds to the tooltips of PR/ROC curves plotted via the following methods:

  • BinaryClassificationResults.plot_pr_curve()

  • BinaryClassificationResults.plot_roc_curve()

  • COCODetectionResults.plot_pr_curves()

  • OpenImagesDetectionResults.plot_pr_curves()

Brain

• Graceful handling of missing/uncomputable embeddings in compute_uniqueness()

• Graceful handling of edge cases like fraction <= 0 in find_duplicates(),

• Removed a spurious warning message that was previously logged when computing patch embeddings for a collection containing samples with no patches

Annotation

• Added a new Labelbox integration!

• Added an import_annotations() method for importing existing CVAT projects or task(s) into FiftyOne

• Added support for configuring the size of CVAT tasks via a new task_size parameter

• Added graceful handling of deleted tasks when importing annotations from CVAT via load_annotations()

• Added an unexpected parameter that provides a variety of options for handling unexpected annotations returned by the CVAT API

• Added support for passing request headers to the CVAT API

• Fixed a bug that occured when importing single frame track segments from CVAT

Zoo

• Fixed a regression in fiftyone==0.14.1 that prevented zoo datasets that use the Torch backend from being downloaded

• Added the following TF2 models to the Model Zoo!

  • centernet-hg104-1024-coco-tf2

  • centernet-resnet101-v1-fpn-512-coco-tf2

  • centernet-resnet50-v2-512-coco-tf2

  • centernet-mobilenet-v2-fpn-512-coco-tf2

  • efficientdet-d0-512-coco-tf2

  • efficientdet-d1-640-coco-tf2

  • efficientdet-d2-768-coco-tf2

  • efficientdet-d3-896-coco-tf2

  • efficientdet-d4-1024-coco-tf2

  • efficientdet-d5-1280-coco-tf2

  • efficientdet-d6-1280-coco-tf2

  • efficientdet-d7-1536-coco-tf2

  • ssd-mobilenet-v2-320-coco17

  • ssd-mobilenet-v1-fpn-640-coco17

FiftyOne 0.14.2

Released November 24, 2021

App

• Improved mask loading times for Segmentation, Heatmap, and Detection labels with instance masks

Core

• Optimized image metadata calculation to read only the bare minimum byte content of each image

• Improved handling of relative paths and user paths in config settings and environment variables

• Optimized database I/O and improved the helpfulness of warnings/errors that are generated when applying models via apply_model(), compute_embeddings(), and compute_patch_embeddings()

• Resolved a memory leak that could occur when computing predictions/embeddings for very large datasets with Torch models

Brain

• Added the points keyword argument to compute_visualization() for providing your own manually computed low-dimensional representation for use with interactive embeddings plots

• Graceful handling of missing/uncomputable embeddings in compute_visualization() and compute_similarity()

• Added checks that occur at the start of all methods to ensure that any required dependencies are installed prior to performing any expensive computations

Annotation

• Changed CVAT uploads to retain original filenames

• A helpful error is now raised when the "frames." prefix is omitted from label fields when requesting spatial annotations on video datasets

Zoo

• Patched an issue that prevented downloading the VOC-2007 and VOC-2012 datasets from the zoo

FiftyOne 0.14.1

Released November 15, 2021

App

• Optimized grid loading for collections that do not have metadata computed

• Fixed filtering by label for Colab notebooks

• Fixed a bug where the App would crash if an image or video MIME type could not be inferred from the filepath, e.g. without an extension

• Fixed first pixel coloring for segmentations

• Added graceful handling of nonfinites (-inf, inf, and nan)

Core

• Fixed clone() for views with a parent dataset that has brain runs

• Fixed sampling frames when using to_frames()

• Fixed importing of FiftyOneDataset with run results

• Added a Regression label type

• Added a random_split() method

• Added support for negating match_labels() queries

• Added a MaxResize transform

• Added image_max_size and image_max_dim parameters to TorchImageModelConfig

• Added support for non-sequential updates in set_values()

• Added a compute_max_ious() utility

• Added support for labels-only exports when working with YOLOv4Dataset and YOLOv5Dataset formats

• Added fiftyone.utils.beam for parallel import, merge, and export operations with Apache Beam

• Added an add_yolo_labels() utility that provides support for adding YOLO-formatted model predictions to an existing dataset

• Added support for importing/exporting multilabel classifications when using FiftyOneImageClassificationDataset format

• Fixed the force_reencode flag for reencode_videos()

• Converted COCO and Open Images dataset downloads to use multithreading rather than multiprocessing

• Updated evalution confusion matrices to always include rows and columns for missing/other

Annotation

• Added support for annotating multiple label fields in one CVAT task

• Added an allow_index_edits parameter to annotate() for disallowing video track index changes

• Improved label ID tracking in CVAT by leveraging CVAT’s server IDs in addition to label_id attributes

• Fixed a bug when annotating videos in CVAT with None label fields

• Fixed a bug when annotating new fields in CVAT

• Fixed a bug when annotating non-continuous tracks in CVAT

• Fixed a bug when annotating a track in CVAT that is present on the last frame of a video

• Fixed a bug when annotating with allow_additions=False

Docs

• Added a section on adding model predictions to existing datasets to the user guide

• Added explicit examples of labels-only imports and exports for all relevant datasets to the docs

• Documented how class lists are computed when exporting in formats like YOLO and COCO that require explicit class lists

• Documented the supported label types for all exporters

FiftyOne 0.14.0

Released October 15, 2021

App

• Added support for visualizing heatmaps using either transparency or a customizable colorscale

• Added a label opacity slider in both the sample grid and the expanded sample view

• Added support for visualizing clips views

• Added support for rendering and filtering DateField and DateTimeField data

• Improved error handling in the grid and when streaming frames

• Fixed a bug that caused incorrect label rendering for sparse frame labels in the video visualizer

• Added a default_app_address setting to the FiftyOne config for restricting sessions to a hostname. See this page for more details

Core

• Added a Heatmap label type

• Added support for adding date and datetime fields to FiftyOne datasets

• Added the to_clips() method for creating clips views into video datasets

• Added clip views sections to the App user guide page and dataset views user guide page

• Added support for exporting video clips in labeled video formats

• Added a trajectories=True flag to filter_labels() that allows for matching entire object trajectories for which a given filter matches the object in at least one frame of the video

• Added set operations is_subset(), set_equals(), unique(), union(), intersection(), difference(), and contains(all=True) to the view expression API

• Added date operations to_date(), millisecond(), second(), minute(), hour(), day_of_week(), day_of_month(), day_of_year(), month(), and year() to the view expression API

• Missing ground truth/predictions are now included by default when viewing confusion matrices for detection tasks

Annotation

• Added support for specifying per-class attributes when defining a label schema for an annotation task

• Added support for specifying whether labels can be added, deleted or moved and whether certain label attributes are read-only when configuring an annotation task

• Added support for respecting keyframe information when adding or editing video annotations

• Fixed a 0-based versus 1-based frame numbering bug when importing and exporting labels in CVAT video format

• Added support for adding/editing bounding box shapes (not tracks) if desired when annotating video frames using the CVAT backend

• Fixed a bug that prevented importing of video annotations from the CVAT backend that involved the splitting or merging of object tracks

• Added a project_name parameter that allows for creating annotation tasks within a new project when using the CVAT backend

• Added support for specifying a list of task assignees when creating video annotation tasks (which generate one task per video) using the CVAT backend

• Fixed a bug when adding/editing boolean attributes in an annotation task using the CVAT backend

• Added a new occluded attribute type option that links an attribute to the builtin occlusion icon when annotating label attributes using the CVAT backend

FiftyOne 0.13.3

Released September 22, 2021

App

• Improved the efficiency of loading label graphs for fields with many distinct values

• Fixed some audio-related bugs when viewing video samples with audio channels

• Fixed a bug that prevented boolean App filters from working properly

Core

• Added support for importing/exporting segmentation masks with greater than 256 classes when working with the ImageSegmentationDirectory format

• Added support for importing GeoTIFF images via a new GeoTIFFDataset format

• Added new split_labels() and merge_labels() methods that provide convenient syntaxes for moving labels between new and existing label fields of a dataset

• Added ensure_frames() and clear_frames() methods that can be used to conveniently initialize and clear the frames of video datasets, respectively

• Added support for using a MongoDB dataset whose version is not explicitly supported

• Removed the opencv-python-headless maximum version requirement

• Fixed a race condition that could prevent callbacks on interactive plots from working properly on sufficiently large datasets

Annotation

• Added support for annotating semantic segmentations and instance segmentations using the CVAT backend

• Added support for annotating polylines using the CVAT backend

• Added support for immutable attributes when annotating object tracks for video datasets using the CVAT backend

• Exposed the use_cache, use_zip_chunks, and chunk_size parameters when uploading annotations via the CVAT backend

• Fixed a bug that prevented multiple imports of the same annotation run from working as expected when a label is deleted but then later re-added

• Fixed a bug that prevented annotations for new label fields of video datasets from being imported properly

• Fixed a bug that would cause unsuppoted shapes such as polygons with less than 3 vertices to be deleted when editing existing labels with the CVAT backend

FiftyOne 0.13.2

Released September 3, 2021

App

• Improved aggregation queries resulting in ~10x faster statistics load times and time-to-interaction in the Fields Sidebar!

• Optimized in-App tagging actions

• Fixed count inconsistencies for large sets of StringField results in the Fields Sidebar

Core

• Added support for providing compound sort criteria when using the sort_by() stage

• Added support for configuring the wait time when using Session.wait() to block execution until the App is closed, including support for serving forever

• Fixed errors experienced by Windows users when running FiftyOne APIs that involved multiprocessing

• Resolved errors that could occur when importing CVAT XML files with empty-valued attributes in their schema and/or individual labels

• Added support for importing CVAT-style attributes when loading labels in COCO and VOC formats

FiftyOne 0.13.1

Released August 25, 2021

App

• Fixed id rendering in the grid when the id checkbox is active

Annotation

• Fixed a bug that could cause mismatches between media and their pre-existing labels when uploading data to CVAT for annotation whose source media lives in multiple directories

FiftyOne 0.13.0

Released August 24, 2021

App

• Added support for visualizing and filtering list fields

• Added support for visualizing segmentation masks of any integer type (uint8, uint16, etc.)

• Improved handling of falsey field values in the fields sidebar and image vizualizer

• Improved the JSON display format available from the expanded sample modal

• Resolved an issue that caused some users to see duplicate App screenshots when calling Session.freeze() in Jupyter notebooks

• Fixed a bug that prevented users from being able to click left/right arrows to navigate between samples in the expanded sample modal when working in Jupyter notebooks

• Fixed a bug where pressing the ESC key had no effect in the expanded sample modal when working with datasets with no label fields

• Fixed a bug that prevented the desktop App from launching when using source builds

Brain

• Added new find_unique(), unique_view(), and visualize_unique() methods to the SimilarityIndex object returned by compute_similarity() that enable you to identify a maximally unique set of images or objects in a dataset

• Added new find_duplicates(), duplicates_view(), and visualize_duplicates() methods to the SimilarityIndex object returned by compute_similarity() that enable you to identify near-duplicate images or objects in a dataset

• Added a new compute_exact_duplicates() method that can identify exactly duplicate media in a dataset

Core

• Added support for pip-installing FiftyOne on Apple Silicon Macs. Note that MongoDB must be self-installed in this case

• Added support for using a self-installed MongoDB database

• Added a group_by() view stage that allows for reorganizing the samples in a collection so that they are grouped by a specified field or expression

• Added a selection_mode property that enables users to change the behavior of App updates when selections are made in an interactive plot linked to labels. See this page for details

• Added support for YOLOv5 YAML files with multiple directories per dataset split

• Added support for importing/exporting confidences via the score field when working with BDD format

• Fixed some Windows-style path bugs

Annnotation

• Added a powerful annotation API that makes it easy to add or edit labels on your FiftyOne datasets or specific views into them

• Added a native CVAT integration that enables you to use the annotation API with CVAT

Docs

• Added a CVAT annotation tutorial

• Added a new example to the brain user guide that demonstrates unique and near-duplicate image workflows

• Added an object embeddings example to the embeddings visualization section of the brain user guide

• Added a new section to the plots user guide page explaining how to control the selection mode of interactive plots linked to labels

FiftyOne 0.12.0

Released August 10, 2021

App

• Resolved performance issues with scrolling via grid virtualization. Toggling fields or selecting samples is no longer impacted by the amount of samples that have been loaded

• Added the Show label option in the expanded sample view to toggle the label text above detections boxes

• Added support for zooming and panning in the expanded sample view

• Added support for cropping and zooming to content in the expanded sample view

• Added support for visualizing multiple segmentation frame fields simultaneously

• Added label streaming to the video visualizer

• Added volume and playback rate settings to the video visualizer

• Added the Crop to content option in patches or evaluation patches views which crops samples to only show the labels that make up the patch. Defaults to True

• Added count and filtered count values to categorical filters (BooleanField and StringField fields)

Core

• Added support for importing DICOM datasets

• Added better default behavior for the label_field parameter when importing datasets using methods like from_dir() and exporting datasets using export()

• When adding samples to datasets, None-valued sample fields are now gracefully ignored when expanding dataset schemas

Docs

• Added Using the image visualizer and Using the video visualizer sections to the App user guide

• Added sections covering merging datasets and batch updates to the dataset user guide page

Zoo

• Patched an Open Images issue where classes or attrs requirements were being ignored when loading a dataset with no max_samples requirement

FiftyOne 0.11.2.1

Released July 31, 2021

Zoo

• Patched an Open Images issue where label files were not being downloaded when running a load_zoo_dataset() call that does not include classes or attrs options in an environment where Open Images has never been downloaded

• Patched loading of Cityscapes datasets

• Patched loading of COCO datasets

FiftyOne 0.11.2

Released July 27, 2021

App

• Added support for calling Session.open_tab() from remote Jupyter notebooks

• Fixed a bug that could cause Session.wait() to exit when the App’s tab is refreshed in the browser

Core

• Added a plotly<5 requirement, which prevents an issue that may cause callbacks for selection events in interactive plots to not trigger as expected when using Plotly V5

• Added support for evaluating polygons and instance segmentations to evaluate_detections(). See this page for usage details

• Added support for creating patch views and evaluation patch views into the frames of video datasets

• Greatly improved the efficiency of creating evaluation patch views

• Added support for recursively listing data directories when loading datasets from disk

• Added support for controlling whether/which object attributes are imported/exported in formats like COCO that support arbitrary object attributes

• Updated all dataset import/export routines to support/prefer custom object attributes stored directly on Label instances as dynamic fields rather than in the attributes dict

• The ImageSegmentationDirectory format now supports exporting segmentations defined by Detections with instance masks and Polylines

• Added an objects_to_segmentations() utility for converting Detections with instance fields and Polylines to Segmentation format

• Added graceful handling of edges cases like empty views and missing labels to all evaluation routines

• Added improved support for creating, viewing, and dropping dropping sample- and frame-level indexes on datasets

• Added additional indexes on patch and frames views to enable efficient ID-based queries

• Added support for gracefully loading and deleting evaluations and brain methods executed in future versions of FiftyOne (e.g., after downgrading your FiftyOne package version)

• Added an optional progress flag to iter_samples() that renders a progress bar tracking the progress of the iteration

• Added support for installing FiftyOne on RHEL7 (Red Hat Enterprise Linux)

• A helpful error message is now raised when a user tries to load a dataset from a future version of FiftyOne without following the downgrade instructions

• Fixed a bug that prevented FiftyOne from being imported on read-only filesystems

• Fixed a bug that prevented the proper loading of the Open Images V6 dataset after partial downloads involving only a subset of the available label types

Zoo

• Added support for importing license data when loading the COCO-2014 and COCO-2017 datasets from the zoo

• The inapplicable classes flag will now be ignored when loading the unlabeled test split of COCO-2014 and COCO-2017

• Improved the partial download behavior of the Open Images V6 dataset when the optional classes and attrs parameters are provided

• Fixed a bug that prevented Windows users from downloading the Open Images V6 dataset

FiftyOne 0.11.1

Released June 29, 2021

App

• Updated the expired Slack community link in the App menu bar

FiftyOne 0.11.0

Released June 29, 2021

News

• With support from the COCO team, FiftyOne is now a recommended tool for downloading, visualizing, and evaluating on the COCO dataset! Check out this guide for more details

App

• Fixed a bug that prevented sample_id fields from appearing in the App when working with frames and patches views

Core

• Added various new parameters to methods like Dataset.from_dir() and SampleCollection.export(), including data_path, labels_path, and export_media that allow for customizing the import and export of datasets. For example, you can now perform labels-only imports and exports

• Added new Dataset.merge_dir() and Dataset.merge_importer() methods for merging datasets from disk into existing FiftyOne datasets

• Added support for importing and exporting datasets in YOLOv5 format

• Updated the GeoJSONDataset dataset type to support both image and video datasets

• Added support for importing and exporting extra attributes in COCO format via a new extra_attrs parameter

Zoo

• Added support for partial downloads and loading segmentations to the COCO-2014 and COCO-2017 datasets

• When performing partial downloads on the Open Images v6 Dataset involving a subset of the available classes, all labels for matching samples will now be loaded by default

Docs

• Added a new page demonstrating how to use FiftyOne to download, visualize, and evaluate on the COCO dataset

• Added a new page demonstrating how to use FiftyOne to download, visualize, and evaluate on the Open Images dataset

FiftyOne 0.10.0

Released June 21, 2021

News

• We’ve collaborated with the PyTorch Lightning team to make it easy to train Lightning Flash tasks on your FiftyOne datasets. Check out this guide for more details

Core

• Updated the apply_model() and compute_embeddings() methods to natively support applying Lightning Flash models to FiftyOne datasets!

Docs

• Added a new page demonstrating how to use the Lightning Flash integration

FiftyOne 0.9.4

Released June 15, 2021

App

• Added support for matching samples by ID in the Filters Sidebar

• Fixed a bug that caused the App to crash when selecting samples with the Color by value setting active

• Fixed a bug that caused the App to crash on some Windows machines by ensuring the correct MIME type is set for JavaScript files

Core

• Improved the performance of importing data into FiftyOne by 2x or more!

• Added a to_frames() view stage that enables on-the-fly conversion of video datasets into frames views

• Added last(), head(), and tail() methods to the Frames class

• Added new exclude_fields(), select_frames(), and match_frames() view stages that enable selecting specific frames of video collections via IDs or filter expressions, respectively

• Added a new match_labels() view stage that enables matching samples that have specific labels without actually filtering the non-matching labels

• Added support for exporting image patches using export() by specifying an image classification dataset type and including a spatial label_field that defines the image patches to extract

• Added support for automatically coercing single label fields like Detection into the corresponding multiple label field type Detections when using export() to export in dataset formats that expect list-type fields

• Added support for executing an aggregation on multiple fields via the abbreviated syntax ids, filepaths = dataset.values(["id", "filepath"])

• Exposed the id field of all samples and frames in dataset schemas

• Added support for merging the elements of list fields via Dataset.merge_samples() and Document.merge()

• Added a number of useful options to Dataset.merge_samples(), including fields, omit_fields, and merge_lists

• Improved the efficiency of Dataset.merge_samples() when the overwrite=False option is provided

• Added an optional bool flag to the match_tags() view stage that allows for optionally matching samples without the specified tags

• Added support for computing filehashes via the hashlib module to compute_filehash()

• Updated the import_from_labelbox() method to use the correct label ID (“DataRow ID”, not “ID”)

• Added an optional edges argument to scatterplot() and location_scatterplot() that enables drawing undirected edges between scatterpoints

• Fixed a bug in limit_labels() that would cause views to contain empty label lists if the source dataset contains None-valued fields

• Fixed a bug that prevented ViewExpression.contains() from accepting ViewExpression instances as arguments

Zoo

• Fixed a string encoding issue that prevented some Windows users from loading the Open Images V6 dataset

• Updated the vgg16-imagenet-tf1 model (formerly named vgg16-imagenet-tf) to reflect the fact that it only supports TensorFlow 1.X

Docs

• Added example usages of to_frames() to the user guide

FiftyOne 0.9.3

Released May 18, 2021

App

• Fixed an issue that prevented some datasets and views that contain vector or array data (e.g., logits) from properly loading in the App

• Fixed a bug that prevented loading video datasets in the App in Google Colab environments

FiftyOne 0.9.2

Released May 16, 2021

Zoo

• Fixed a multiprocessing bug that prevented Mac users running Python 3.8 or later from loading the Open Images V6 dataset

FiftyOne 0.9.1

Released May 12, 2021

App

• Fixed a bug that caused the App to crash when choosing to Color by value

FiftyOne 0.9.0

Released May 12, 2021

News

• We’ve collaborated with the Open Images Team at Google to make FiftyOne a recommended tool for downloading, visualizing, and evaluating on the Open Images Dataset! Check out this guide for more details

App

• Added a Patches action for easy switching to object/evaluation patches views. See this page for usage details

• Added a Sort by similarity action that enables sorting by similarity to the selected samples/patches. See this page for usage details

• Added a zoom slider in the top right of the sample grid that adjusts the tile size of the sample grid

• Added the ability to clear filters for entire field groups, e.g. Labels and Scalars, in the Filters Sidebar

• Added filepath to the Scalars group in the Filters Sidebar

• Added a Label tags graphs tab

• Refreshed numeric, string, and boolean filter styles with improved functionality and interaction

• Added support for Session.wait() in browser contexts

Brain

• Added a compute_similarity() method for indexing samples and object patches by similarity. See this page for usage details

Core

• Added support for Open Images-style detection evaluation when using evaluate_detections(). See this page for usage details

• Added the to_patches() and to_evaluation_patches() view stages for transforming collections into flattened views with respect to labels and evaluations, respectively. See this page for usage details

• Added support for applying image models to the frames of video datasets when using apply_model(), compute_embeddings(), and compute_patch_embeddings()

• Added full support for embedded documents (e.g. labels) in values() and set_values()

• Added support for passing expressions directly to aggregations

• Added an optional omit_empty flag to select_labels() and exclude_labels() that controls whether samples with no labels are omitted when filtering

• Added a Dataset.delete_labels() method for efficiently deleting labels via a variety of natural syntaxes

• Deprecated Dataset.remove_sample() and Dataset.remove_samples() in favor of a single Dataset.delete_samples() method

• Brain results and evaluation results that are loaded via load_evaluation_results() load_brain_results() are now cached on the Dataset object in-memory so that subsequent retrievals of the results in the same session will be instant

Zoo

• Added Open Images V6 to the dataset zoo!

Docs

• Added a new Open Images tutorial

• Added object patches and evaluation patches sections to the App guide

• Added a similarity section to the Brain guide

• Added Open Images evaluation and evaluation patches sections to the evaluation guide

• Added object patches and evaluation patches sections to the views guide

• Added example uses of to_patches() and to_evaluation_patches() to the object detection tutorial

• Added example use of to_patches() to the detection mistakes tutorial

• Added example use of to_patches() to the adding detections recipe

FiftyOne 0.8.0

Released April 5, 2021

App

• Added the ability to tag samples and labels directly from the App in both the sample grid (macro) and expanded sample view (micro) with respect to and filters or currently selected samples/labels

• Added a LABEL TAGS section to the Filters Sidebar to coincide with the introduction of label tags

• Added label tooltips that display on hover in the expanded sample view

• Expanded actions to list of button groups in the sample grid and expanded sample view

• Added support for rendering semantic labels in the new tooltip in the expanded sample view for Segmentation mask values (pixel values) using the new Dataset.mask_targets and Dataset.default_mask_targets fields

• Fixed hiding, clearing, and only showing selected samples in the samples grid

Brain

• Added a compute_visualization() method that uses embeddings and dimensionality reduction methods to generate interactive visualizations of the samples and/or labels in a dataset. Check out this page for details. Features include:

  • Provide your own embeddings, or choose a model from the Model Zoo, or use the provided default model

  • Supported dimensionality reduction methods include UMAP, t-SNE, and PCA

  • Use this capability in a Jupyter notebook and you can interact with the plots to select samples/labels of interest in a connected Session

• Added support for saving brain method results on datasets. Previous brain results can now be loaded at any time via Dataset.load_brain_results()

• Added support for providing a custom Model or model from the Model Zoo to compute_uniqueness()

Core

• Added a fiftyone.core.plots module that provides a powerful API for visualizing datasets, including interactive plots when used in Jupyter notebooks. See this page for more information. Highlights include:

  • plot_confusion_matrix(): an interactive confusion matrix that can be attached to a Session object to visually explore model predictions

  • scatterplot(): an interacive scatterplot of 2D or 3D points that can be attached to a Session to explore the samples/labels in a dataset based on their locations in a low-dimensional embedding space

  • location_scatterplot(): an interacive scatterplot of a dataset via its GeoLocation coordinates

  • Added GeoLocation and GeoLocations label types that can be used to store arbitrary GeoJSON location data on samples

  • Added the GeoJSONDataset dataset type for importing and exporting datasets in GeoJSON format

  • Added SampleCollection.geo_near() and SampleCollection.geo_within() view stages for querying datasets with location data

• Upgraded the implementation of the FiftyOneDataset format, which is now 10-100x faster at importing/exporting datasets

• Added support for generating zip/tar/etc archives to SampleCollection.export() by passing an archive path rather than a directory path

• Added Dataset.from_archive() and Dataset.add_archive() factory methods for importing datasets stored in archives

• Added support for saving evaluation results on a dataset. Results can now be loaded at any time via Dataset.load_evaluation_results()

• Added a tags attribute to all Label types that can store a list of string tags for the labels (analogous to the tags attribute of Sample)

• Added a number of methods for working with sample and label tags:

  • SampleCollection.tag_samples()

  • SampleCollection.untag_samples()

  • SampleCollection.count_sample_tags()

  • SampleCollection.tag_labels()

  • SampleCollection.untag_labels()

  • SampleCollection.count_label_tags()

• BREAKING CHANGE: Renamed all applicable API components that previously referenced “objects” to use the more widely applicable term “labels”. Affected attributes, classes, and methods are:

  • Session.selected_labels (previously selected_objects)

  • SampleCollection.select_labels() (previously select_labels())

  • SampleCollection.select_labels() (previously exclude_labels())

  • SelectLabels (previously SelectObjects)

  • ExcludeLabels (previously ExcludeObjects)

• Added new keyword arguments ids, tags, and fields to SampleCollection.select_labels() and SampleCollection.select_labels() and their corresponding view stages that enable easier-to-use selection of labels by their IDs or tags

• Added Session.select_labels() for programmatically selecting labels as well a setters for Session.selected and Session.selected_labels

• Added Dataset.classes and Dataset.default_classes fields that enable storing class label lists at the dataset-level that can be automatically used by methods like Dataset.evaluate_classifications() when knowledge of the full schema of a model is required

• Added Dataset.mask_targets and Dataset.default_mask_targets fields for providing semantic labels for Segmentation mask values to be used in the App’s expanded sample view

• Improved the runtime of Dataset.merge_samples() by ~100x for image datasets and ~10x for video datasets

• Added an Dataset.add_collection() method for adding the contents of a SampleCollection to another Dataset

• Added the trigonometric view expresssions cos, sin, tan, cosh sinh, tanh, arccos, arcsin, arcan arccosh, arcsinh, and arctanh

• Added a randn expression that can generate Gaussian random numbers

• Fixed a bug that prevented evaluate_detections() from being able to process video datasets

• Added support for applying intensive view stages such as sorting to datasets whose database representation exceeds 100MB

• Fixed schema errors in DatasetView instances that contain selected or excluded fields

• Fixed copying of DatasetView instances where ViewField is used

Zoo

• Added the quickstart-geo dataset to enable quick exploration of location-based datasets

CLI

• Removed the --desktop flag from the fiftyone app connect command

Docs

• Added a tutorial demonstrating how to use compute_visualization() on image datasets

• Added an interactive plots page to the user guide

• Added a Tags and tagging section to the App user guide

• Added a visualizing embedding section to the Brain user guide

FiftyOne 0.7.4

Released March 2, 2021

App

• Fixed a bug that prevented Session updates from triggering App updates

• Fixed hiding labels in the expanded sample view

Brain

• Added support for tracking and cleaning up brain runs similar to how evaluations are tracked. See get_brain_info(), list_brain_runs(), load_brain_view(), and delete_brain_run() for details

• Updated compute_mistakenness() to use FiftyOne’s evaluation framework

Core

• Decoupled loading video Sample and SampleView and their frames so the samples are loaded efficiently and frames are only loaded when requested

• Add a 90 character limit to progress bars in notebook contexts to prevent overflow issues

• Added low-level utility methods list_datasets() and delete_dataset() for managing a corrupted database

• Added automatic field generation for labelbox_id_field when using import_from_labelbox()

CLI

• Added a dataset stats command

FiftyOne 0.7.3

Released February 18, 2021

App

• Added filtering widgets to the Filters Sidebar for StringFields and BooleanFields

• Added histogram plots for StringFields and BooleanFields in the Scalars tab

• Moved None selection for StringFields to the input format in the Filters Sidebar

• Changed None options to only be present when None values exist for a supported Field in the Filters Sidebar

• Added Color by label support for Classification, Classifications, BooleanField, and StringField

• Added support excluding selected values for a StringField in the Fields Sidebar

• Various style and interaction improvements in the Filters Sidebar

• The App will no longer crash when samples whose source media is unsupported or missing are loaded

Core

• Added evaluate_classifications(), evaluate_detections(), and evaluate_segmentations() methods that provide support for evaluating various types of labels. See the new evaluation page of the user guide for more details

• Added one() for retrieving one matched Sample from a Dataset or DatasetView

• Added support for cloning and saving views into video datasets via clone() and save()

• Added support for extracting batches of frame-level and/or array fields via the values() aggregation

• Added support for setting batches of frame-level and/or array fields via set_values()

• Added support for accessing samples from a Dataset or DatasetView via the dataset[filepath] syntax

• Added support for passing Sample and any Sample iterable, e.g. DatasetView, to methods like remove_samples(), exclude(), and select()

• Changed the default value for only_matches for filter_classifications(), filter_detections(), filter_field(), filter_labels(), filter_keypoints(), and filter_polylines() from False to True

• compute_metadata() will now gracefully skip samples for which media metadata cannot be computed

• Added a stats() method for listing helpful info about the size of various entities of a dataset

Zoo

• Added support for storing logits for many zoo models when using apply_model()

• Default confidence thresholds for zoo models are now stored on a per-model basis rather than as a global default value in apply_model(). All detection models still have a default confidence threshold of 0.3, and all other model types have no default confidence threshold

CLI

• Added a migration API to provide better support for downgrading the version of your fiftyone package

Docs

• Added a new evaluation page to the user guide that explains how to evaluate various types of models with FiftyOne

• Removed legacy --index flags from the install instructions from the troubleshooting page which prevented a valid installation

FiftyOne 0.7.2

Released January 28, 2021

App

• Changed the Filters Sidebar label filters to only return matched samples, i.e., samples with at least one matching label with respect to a filter

• Fixed a bug in Colab notebooks that allowed for the .ipynb file to grow unnecessarily large

• Improved plotting of numeric fields in the Scalars tab, including [min, max) ranges for tooltips and integer binning when appropriate

• Fixed a bug that prevented select_fields() and exclude_fields() from being properly respected by the Filters Sidebar

• Fixed a bug that prevented selected samples from being cleared when modifying your view or choosing an option from the select samples dropdown

• Added an AppConfig for configuring options like the color pool to use when drawing Label fields. See this page for more info

Core

• Added the MapLabels and SetField view stages

• Added the HistogramValues and Sum aggregations

• Added over a dozen new ViewExpression methods including powerful transformations like map_values(), reduce(), and sort()

• Exposed all Aggregtaions as single execution methods on the SampleCollection interface, e.g., distinct()

• Added support for all Label types in filter_labels()

• Added support for generalized field paths (embedded fields, lists, etc) to the Bounds, Count, CountValues, and Distinct aggregations

• Removed the deprecated ConfidenceBounds, CountLabels, and DistinctLabels aggregations

• Removed the redundant match_tag() stage in favor of match_tags()

• Removed AggregationResult classes in favor of returning Aggregation results directly as builtin types

• Added the optional config keyword argument to launch_app() and Session for overriding the default AppConfig.

Zoo

• Added a default confidence threshold of 0.3 when applying models from the Model Zoo via apply_model(), which omits spurious low quality predictions from many models

CLI

• Added a fiftyone app config command for inspecting the default App config

• Improved ctrl + c exit handling for CLI commands

Docs

• Added a new section to the Configuring FiftyOne guide explaining how to programmatically configure the App’s behavior

• Updated the Dataset views guide to provide a thorough overview of new functionality provided by stages like SetField

• Updated the Aggregations guide to provide a thorough overview and examples of various aggregation functionality, including advanced usage tips

• Added an FAQ section providing instructions for working with remote Jupyter notebooks

• Added code examples to all ViewStage class docstrings and their corresponding sample collection methods, e.g., map_labels()

• Added code examples to all Aggregation class docs and their corresponding sample collection methods, e.g., bounds()

FiftyOne 0.7.1

Released January 8, 2021

App

• Added automatic screenshotting for notebook environments

• Fixed a bug where the Filters Sidebar statistics would not load for empty views

• Fixed style inconsistencies in Firefox

Core

• Added Session.freeze() for manually screenshotting the active App in a notebook environment

• Renamed Dataset.clone_field() to Dataset.clone_sample_field() for consistency with other operations

• Added a Dataset.clone_frame_field() method for cloning frame-level fields of video datasets

• Added DatasetView.clone_sample_field() and DatasetView.clone_frame_field() methods that allow cloning views into sample fields (e.g., after filtering the labels in a list field)

• Added a DatasetView.clone() method for cloning a view as a new dataset

• Added a DatasetView.save() method for saving a view, overwriting the contents of the underlying dataset

• Re-implemented Dataset.clone_sample_field() and Dataset.merge_samples() via efficient DB-only operations

• Added the overwrite keyword argument to the Dataset() constructor

• Added a database_dir option to the FiftyOne Config

• Added a default_app_port option to the FiftyOne Config

Zoo

• Added a CenterNet model to the model zoo

• Upgraded the Model Zoo so that many detection models that previously required TensorFlow 1.X can now be used with TensorFlow 2.X

• Added Caltech-256 to the dataset zoo

• Added ImageNet Sample to the dataset zoo

• Caltech-101 is now available natively in the dataset zoo without the TF backend

• KITTI is now available natively in the dataset zoo without the TF backend

• Fixed a bug that prevented ImageNet 2012 from loading properly when using the TF backend

CLI

• Added support for controlling the error level when applying zoo models

Docs

• Added a Dataset Zoo listing that describes all datasets in the zoo

• Added a Model Zoo listing that describes all models in the zoo

FiftyOne 0.7.0

Released December 21, 2020

App

• Added web browser support, which is now the default settting

• Added IPython notebook support, e.g. Jupyter and Google Colab

• The desktop App can now be installed as an optional dependency

• Fixed an issue where the App would freeze after filtering labels in the Filters Sidebar

Core

• Added a Model Zoo containing over 70 pretrained detection, classification, and segmentation models that you can use to generate predictions and embeddings

• Moved project hosting to pypi.org

• Added the Session.show() method for displaying the App in IPython notebook cells

• Added an in-App feedback form. We would love to hear from you!

• Added Python 3.9 support

• Removed Python 3.5 support

CLI

• Added a fiftyone zoo models command that provides access to the model zoo

• Moved the dataset zoo commands to fiftyone zoo datasets (previously they were at fiftyone zoo)

• Added a --desktop flag to commands that launch the App that enables launching the App as a desktop App (rather than a web browser)

FiftyOne 0.6.6

Released November 25, 2020

App

• Added a dropdown in the header to change datasets from the App

• Added the ability to refresh the App by clicking the FiftyOne logo in the header

• Fixed a bug the caused numeric (scalar field) range sliders to disappear after changing the default value

• Fixed a bug that prevented the App state from updating appropriately after applying label filters

Brain

• Added support for computing mistakenness for detections when using compute_mistakenness()

Core

• Fixed a bug that prevented COCO datasets from being loaded from the Dataset Zoo

CLI

• Added support for customizing the local port when connecting to the App via the CLI

• Added an --ssh-key option to the app connect command

Docs

• Added a tutorial demonstrating how to use compute_mistakenness() to detect label mistakes for detection datasets

• Added questions to the FAQ page:

  • Can I launch multiple App instances on a machine?

  • Can I connect multiple App instances to the same dataset?

  • Can I connect to multiple remote sessions?

  • Can I serve multiple remote sessions from a machine?

FiftyOne 0.6.5

Released November 16, 2020

App

• Added concurrency to the server wich greatly improves loading speeds and time-to-interaction in the Grid, View Bar, and Filters Sidebar for larger datasets and views

• Renamed the Display Options Sidebar to the Filters Sidebar

• Added support for coloring by label value in the Filters Sidebar

• Added support for filtering keypoint, keypoints, polyline, polylines fields by label value in the Filters Sidebar

• Moved plot tabs into an expandable window that can be resized and maximized. This allows for viewing distributions and the sample grid at the same time

• Fixed video loading in the grid and modal for video samples with metadata

• Fixed showing and hiding samples in the select sample menu

• Fixed a memory usage bug in the sample grid

Core

• Added Cityscapes and LFW to the Dataset Zoo

• Added support for renaming and deleting embedded document fields of samples via Dataset.rename_sample_field() and Dataset.delete_sample_field()

• Added support for renaming and deleting embedded document fields of frames of video samples via Dataset.rename_frame_field() and Dataset.delete_frame_field()

• Added support for deleting fields and embedded fields of individual samples via Sample.clear_field() and del sample[field_name]

• Added support for deleting fields and embedded fields of frame labels via Frame.clear_field() and del frame[field_name]

• Added support for reading/writing video datasets in JSON format via Dataset.from_json() and SampleCollection.write_json(), respectively

• Added a module for importing and exporting annotations from Scale AI

• Added a module for importing and exporting annotations from Labelbox

• Fixed a bug that prevented Dataset.add_sample() and Dataset.add_samples() from working properly when provided samples that belong to other sample collections

• Fixed a bug that prevented frame labels from being properly cloned when calling Dataset.clone() on video datasets

Docs

• Added an Environments page that outlines how to work with local, remote, and cloud data. Includes instructions for AWS, Google Cloud, and Azure

• Add an FAQ page

FiftyOne 0.6.4

Released October 29, 2020

App

• Improved page load times for video datasets

• Improved support for frame- and sample-level labels in display options for video datasets

• Added support for all label types in the labels distributions tab

• Added support for selecting and hiding labels in the sample modal

Brain

• Added support for computing uniqueness within regions-of-interest specified by a set of detections/polylines when using compute_uniqueness()

Core

• Added the filter_labels() view stage, which supercedes the old dedicated per-label-type filtering stages

• Added select_labels() and exclude_labels() to select or exclude labels from a dataset or view

• Added an aggregations framework for computing aggregate values via aggregate()

• Added the selected_labels session attribute, which holds the currently selected labels in the App

• Added support for adding, renaming, and deleting frame-level fields of video datasets

• Added the TorchImagePatchesDataset that emits tensors of patches extracted from images defined by sets of Detections associated with the images

FiftyOne 0.6.3

Released October 20, 2020

App

• Added sample-level display options stats, filtering, and toggling for video datasets

Core

• Added support for importing and exporting video classification datasets organized as directory trees on disk

• Added BDD100K, HMDB51, and UCF101 to the Dataset Zoo

• Added new versions of COCO that contain instance segmentations to the Dataset Zoo

• Added utilities for selecting labels from datasets via the Python library

• Added a boolean only_matches parameter to all filter stages that enables the user to specify that a view should only contain samples that match the given filter

• Improved performance when ingesting video datasets with frame-level labels

• Added a reencode_videos() utility to re-encode the videos in a sample collection so that they are visualizable in the FiftyOne App

FiftyOne 0.6.2

Released October 15, 2020

App

• Improved page and grid load times for video datasets by around 10x

• Added filtering, toggling, and statistics for labels with respect to the frame schema in the display options sidebars for video datasets

• Added margins to the grid view for both image and video datasets

• Fixed list parameter input submission in the view bar

• Fixed an issue causing some label counts to be incorrect after filters are applied

• Added support for using the keyboard to select labels when filtering

Brain

• compute_uniqueness() and compute_hardness() now support multilabel classification tasks

Core

• Polyline instances can now represent labels composed of multiple shapes

• Segmentations can now be imported and exported when using COCO Object Detection Format.

• Polylines and keypoints can now be imported and exported when using CVAT image format

• Polylines and keypoints can now be imported and exported when using CVAT video format

• Added support for rendering annotated versions of video samples with their frame labels overlaid via draw_labels()

• Added support for launching quickstarts as remote sessions

• Added Frames.update() and Frames.merge() methods to replace and merge video frames, respectively

• Fixed Dataset.merge_samples() to properly merge the frame-by-frame contents of video samples

• Fixed a bug where sample.copy() would not create a copy of the frames of a video sample

FiftyOne 0.6.1

Released October 7, 2020

App

• Added support for visualizing keypoints, polylines, and segmentation masks

• Added autocompletion when selecting SortBy fields in the view bar

• Added support for viewing index fields of Detection labels in the media viewer, if present

• Fixed counting of Classifications fields in the expanded sample view

• Fixed a bug that prevented label filters from fully resetting when a reset or clear button is pressed

Core

• Added support for storing keypoints, polylines, and segmentation masks on samples

• Added support for setting an index attribute on Detection instances that defines a unique identifier for an object (e.g., across frames of a video)

• Added support for importing and exporting datasets in YOLOv4 format

• Added support for importing and exporting datasets in CVAT video format

• Added support for importing and exporting video datasets in FiftyOneDataset format

• Added frame field schemas to string representations for video datasets/views

CLI

• Added options to fiftyone datasets delete to delete all datasets matching a pattern and all non-persistent datasets

Docs

• Added a recipe for merging datasets

• Fixed some table widths and other display issues

FiftyOne 0.6.0

Released October 1, 2020

App

• Added support for visualizing video datasets in the App

Core

• Added support for storing frame labels on video samples

• Added support for importing and exporting datasets of unlabeled videos

• Added support for importing and exporting labeled video datasets in ETA VideoLabels format.

• Added support for importing and exporting video datasets in custom formats

• Improved the performance of Dataset.rename_sample_field()

• Added support for using disk space when running aggregation pipelines on large datasets

• Added support for automatically creating database indexes when sorting by sample fields, for efficiency

• Fixed issues with serializing vector fields and numpy arrays

FiftyOne 0.5.6

Released September 23, 2020

App

• Added autocompletion to view bar stage fields that accept field names (for example, Exists)

• Fixed an issue that would prevent datasets with no numeric labels or scalars from loading in the App

• Fixed an error that could occur when a view included no samples

• Added notifications in the App that are displayed if errors occur on the backend

• Improved keyboard navigation between view bar stages

Core

• Added support for loading (possibly-randomized) subsets of datasets when importing them via DatasetImporter instances, or via factory methods such as Dataset.from_dir()

• Added support for optionally skipping unlabeled images when importing image datasets via LabeledImageDatasetImporter instances

• Added a Dataset.merge_samples() method for merging samples in datasets via joining by filepath

• Added a Dataset.rename_sample_field() method for renaming sample fields of datasets

FiftyOne 0.5.5

Released September 15, 2020

App

• Added support for filtering samples by numeric fields in the sidebar

• Confidence bounds are now computed for the confidence slider in the label filter - a [0, 1] range is no longer assumed

• Fixed an issue that would cause certain stages to be reevaluated when the view bar was edited

• Improved responsiveness when adding stages in the view bar, filtering, and selecting samples

• Simplified placeholders in the view bar

• Added support for filtering sample JSON in the expanded sample view to match the labels displayed in the media viewer

• Updated the instructions that appear when starting the App before connecting to a session

Core

• Added support for Session.wait() for remote sessions, to make starting a remote session from a script easier

FiftyOne 0.5.4

Released September 9, 2020

App

• Added support for selecting/excluding samples from the current view in the App by selecting them and then choosing the appropriate option from a sample selection menu

• Added autocomplete when creating new stages in the view bar

• Updated the look-and-feel of the view bar to clarify when a stage and/or the entire view bar are active, and to make the bar more visually consistent with the rest of the App

• Media viewer options are maintained while browsing between samples in expanded sample view

• Improved the look-and-feel of confidence sliders when filtering labels

• Limited floating point numbers to three decimals when rendering them in the media viewer

• Fixed some bugs related to content overflow in the view bar

Core

• Added support for exporting Classification labels in dataset formats that expect Detections labels

• Added support for importing/exporting supercategories for datasets in COCO format

FiftyOne 0.5.3

Released September 1, 2020

App

• Added support for filtering labels in the expanded sample view

• Added support for displaying detection attributes in the expanded sample view

• Added an option to display confidence when viewing labels in the expanded sample view

• Updated look-and-feel of display options in the expanded sample view to match the main image grid view

• Adopting a default color palette for sample fields in the App that ensures that they are visually distinct

• Fixed a bug that prevented the App from loading empty views

• Fixed a bug that prevented the view bar from using default values for some stage parameters

Core

• Added support for checking that a field does not exist via a new boolean parameter of the exists() view stage

• Fixed a bug that prevented FiftyOne from starting for some Windows users

• Fixed a bug that caused take() and shuffle() view stages with random seeds to be regenerated when handing off between the App and Python shell

FiftyOne 0.5.2

Released August 26, 2020

App

• Added a label filter to the App that allows you to interactively explore your labels, investigating things like confidence thresholds, individual classes, and more, directly from the App

• Added an App error page with support for refreshing the App if something goes wrong

• The App can now be closed and reopened within the same session

Core

• Added a fiftyone quickstart command that downloads a small dataset, launches the App, and prints some suggestions for exploring the dataset

• Added support for multiple simultaneous FiftyOne processes. You can now operate multiple App instances (using different ports), Python shells, and/or CLI processes.

• Added support for automatically expanding labels from multitask formats such as BDDDataset and FiftyOneImageLabelsDataset into separate label fields when importing datasets

• Added support for exporting multiple label fields in supported formats such as BDDDataset and FiftyOneImageLabelsDataset via the export() method

• Added support for filtering fields via the filter_field() method

• Provided a more helpful error message when using the Dataset Zoo with no backend ML framework installed

• Made pycocotools an optional dependency to make installation on Windows easier

Docs

• Updated the evaluate object detections tutorial to make it more friendly for execution on CPU-only machines

• Refreshed all App-related media in the docs to reflect the new App design introduced in v0.5.0

FiftyOne 0.5.1

Released August 18, 2020

App

• Statistics in the display options sidebar now reflect the current view, not the entire dataset

• Improved image tiling algorithm that prevents single images from filling an entire grid row

• Added support for toggling label visibility within the expanded sample modal

• Improved display of long label and tag names throughout the app

• Enhanced view bar functionality, including keyword arguments, type annotations, error messages, help links, and overall stability improvements

• Added keyboard shortcuts for interacting with the view bar:

  • DEL and BACKSPACE delete the raised (active) stage

  • ESC toggles focus on the ViewBar, which activates shortcuts

  • TAB, ENTER, and ESC submits stage input fields

  • LEFT and RIGHT ARROW traverses view stages and add-stage buttons

  • SHIFT + LEFT ARROW and SHIFT + RIGHT ARROW traverse stages

Core

• Greatly improved the performance of loading dataset samples from the database

• Added support for renaming and cloning datasets

• Added more string matching operations when querying samples, including starts_with(), ends_with(), contains_str() and matches_str()

Docs

• Added a tutorial demonstrating performing error analysis on the Open Images Dataset powered by FiftyOne

FiftyOne 0.5.0

Released August 11, 2020

News

• FiftyOne is now open source! Read more about this exciting development in this press release

App

• Major design refresh, including a new look-and-feel for the App

• Added view bar that supports constructing dataset views directly in the App

• Redesigned expanded sample view:

  • Improved look-and-feel, with modal-style form factor

  • Added support for maximizing the media player

  • Added support for maximizing the raw sample view

  • Added arrow controls to navigate between samples

Core

• Added support for importing and exporting FiftyOne datasets via the FiftyOneDataset type

• Added a Dataset.info field that can be used to store dataset-level info in FiftyOne datasets

• Added a shuffle() view stage for randomly shuffling the samples in a dataset

• Upgraded the take() view stage so that each instance of a view maintains a deterministic set of samples

FiftyOne 0.4.1

Released August 4, 2020

Core

• Added a powerful fiftyone.core.expressions module for constructing complex DatasetView match(), sort_by(), etc. stages

• Added an evaluate_detections() utility for evaluating object detections in FiftyOne datasets

• Adding support for rendering annotated versions of sample data with their labels overlaid via a draw_labels() method

Docs

• Added a tutorial demonstrating object detection evaluation workflows powered by FiftyOne

• Added full documentation for constructing DatasetViews with powerful matching, filtering, and sorting operations

• Added a recipe showing how to render annotated versions of samples with label field(s) overlaid

• Upgraded dataset creation docs that simplify the material and make it easier to find the creation strategy of interest

• Improved layout of tutorials, recipes, and user guide landing pages

FiftyOne 0.4.0

Released July 21, 2020

App

• Fixed an issue that could cause launching the App to fail on Windows under Python 3.6 and older

Core

• Added support for importing datasets in custom formats via the DatasetImporter interface

• Added support for exporting datasets to disk in custom formats via the DatasetExporter interface

• Added support for parsing individual elements of samples in the SampleParser interface

• Added an option to image loaders in fiftyone.utils.torch to convert images to RGB

• Fixed an issue where Dataset.delete_sample_field() would not permanently delete fields if they were modified after deletion

• Improved the string representation of ViewStage instances

Docs

• Added a recipe demonstrating how to convert datasets on disk between common formats

• Added recipes demonstratings how to write your own custom dataset importers, custom dataset exporters, and custom sample parsers

• Added a Configuring FiftyOne page to the User Guide that explains how to customize your FiftyOne Config

FiftyOne 0.3.0

Released June 24, 2020

App

• Fixed an issue that could prevent the App from connecting to the FiftyOne backend

Core

• Added support for importing and exporting datasets in several common formats:

  • COCO: COCODetectionDataset

  • VOC: VOCDetectionDataset

  • KITTI: KITTIDetectionDataset

  • Image classification TFRecords: TFImageClassificationDataset

  • TF Object Detection API TFRecords: TFObjectDetectionDataset

  • CVAT image: CVATImageDataset

  • Berkeley DeepDrive: BDDDataset

• Added Dataset.add_dir() and Dataset.from_dir() to allow for importing datasets on disk in any supported format

• Added a convert_dataset() method to convert between supported dataset formats

• Added support for downloading COCO 2014/2017 through the FiftyOne Dataset Zoo via the Torch backend

CLI

• Added fiftyone convert to convert datasets on disk between any supported formats

• Added fiftyone datasets head and fiftyone datasets tail to print the head/tail of datasets

• Added fiftyone datasets stream to stream the samples in a dataset to the terminal with a less-like interface

• Added fiftyone datasets export to export datasets in any available format

FiftyOne 0.2.1

Released June 19, 2020

Core

• Added preliminary Windows support

• Dataset.add_images_dir() now skips non-images

• Improved performance of adding samples to datasets

CLI

• Fixed an issue that could cause port forwarding to hang when initializing a remote session

FiftyOne 0.2.0

Released June 12, 2020

App

• Added distribution graphs for label fields

• Fixed an issue causing cached images from previously-loaded datasets to be displayed after loading a new dataset

Core

• Added support for persistent datasets

• Added a class-based view stage approach via the ViewStage interface

• Added support for serializing collections as JSON and reading datasets from JSON

• Added support for storing numpy arrays in samples

• Added a config option to control visibility of progress bars

• Added progress reporting to Dataset.add_samples()

• Added a SampleCollection.compute_metadata() method to enable population of the metadata fields of samples

• Improved reliability of shutting down the App and database services

• Improved string representations of Dataset and Sample objects

CLI

• Added support for creating datasets and launching the App