Changelog (archived)#

Note

This CHANGELOG refers to the time this project was maintained internally by XXII under the name “Libia”. Since the commit history has been removed for security reasons, the chaneglog is kept for informational purpose and should not be modified. The new CHANGELOG is here

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[2.1.1 - 2024-06-21]#

  • Pin numpy version to be <2

[2.1.0 - 2024-06-21]#

Added#

  • Add Dataset.remove_invalid_images and Dataset.remove_invalid_annotations methods.

  • Add mark_origin and overwrite_origin options to Dataset.merge method

  • Add from_pascalVOC_detection and from_pascalVOC_generic functions to load pascal datasets

  • Add dataset_regression fixture for pytest that will test that datasets are the same

  • Add more examples to documentation

Fixed#

  • Fix spelling errors

Changed#

  • Upgrade minimum version to 3.10, so long python 3.9!

  • Upgrade pre-commit template and run it

  • Change most dataset method return types to Self instead of simply "Dataset"

  • Change classmethod Dataset.from_template to be a simple method. Note that this change is not breaking, as Dataset.from_template(input_dataset, **kwargs) is equivalent to input_dataset.from_template(**kwargs)

  • from_coco and from_crowdhuman both try to parse intelligently the annotation file path to extract both the dataset name and the split name, thanks to a new function libia.dataset.io.common.parse_annotation_name

  • Dataset.merge now automatically convert images root of a dataset to absolute if the other is also absolute

  • to_fiftyone methods (for dataset and evaluator) now accept a existing option to handle existing dataset. You can now erase the existing dataset before uploading yours, or raise an error if it exists. Possibly breaking : default behaviour of to_fiftyone methods was “update” and is now “error”

  • Dataset.match_index now accepts a dataset as well as an image dataframe like before

  • Dataset.remap_from_other now accepts remove_not_mapped and remove_emptied_images options to remove classes that are not present in the other dataset.

  • Evaluator now accepts a prediction label map that is neither a subset nor a superset of ground truth label map, and will assume only false negative and false positive for the not mutual classes.

  • dummy_dataset now accepts options keypoints_share and add_confidence to make crowd datasets and predictions

  • Dataset.add_annotations and annotations_appender.append now accepts more flexible attributes shapes, and then broadcast them together.

[2.0.1 - 2024-05-29]#

Added#

  • Add the possibility to test dataset equality modulo columns that are all NaNs

  • Add warning message when label map is incomplete, and complete it with the simple id -> str(id) mapping for missing ids

  • Add check_exhaustive option to Dataset.check and assert_images_valid functions

Fixed#

  • Fix c2p CLI tool to effectively remove a detection when it is modified

  • Dataset.remove_empty_images now keeps the dataset name

  • add docs for darknet IO

  • Suppress some FutureWarning from pandas during tests

  • fix bug for caipy when split is pd.NA instead of None or np.NaN

  • fix bug when loading caipy with splits_to_read set to non existing splits

  • Code spelling

[2.0.0 - 2024-04-02]#

Added#

  • Add input format option for COCO loading, making it possible to load XY coordinates instead of just bounding boxes

  • Add from_coco_keypoints function for loading COCO data with points and only one class.

  • Add compatibility with caipyjson tags and attributes, and more generally any kind of nested dictionary

  • Add column boooleanizer (and debooleanizer) to go from a list objects to columns of boolean value for better queries

  • Add Crowd detection evaluator with Mean Average Error metric for count

  • Add reindex function

  • Add from_mot function for loading datasets in MOT format. See https://motchallenge.net/instructions/

  • Add a method to compute confusion matrix for DetectionEvaluator

  • Add reindex function

  • Add yolov7 compatibility with a Dataset.to_yolov7 method.

  • Add automatic compliance with schema when saving to caipy

  • Add compatibility with caipy splits independently indexed

  • Add iterator helper methods to Dataset like Dataset.iter_images and Dataset.iter_splits to make it easier to iterate by a specific attribute

  • When loaded with a schema, from_caipy automatically set missing arrays to the empty list and other fields to their default value specified in the schema when at least one sample in the caipy folder has the field set to a particular value in its caipyjson file, avoiding NaN values in the resulting dataframe.

  • Add to_parquet and from_parquet method to save and load dataset efficiently with pyarrow.

  • Add dataframe booleanized columns broadcasting functions, useful for merging datasets

  • Add better error messages when calling check functions from utils.testing

  • Add remap_from_other method to remap label map to match another dataset.

  • Add realign_label_map argument in Dataset.merge to avoid incompatible label maps error

  • Add assert_columns_properly_normalized for caipy json reading

  • Add Dataset.empty() method to create the same dataset object as before, but with an empty dataframe of annotations. This is useful when creating a prediction dataset.

  • Add AnnotationAppender.reset() and AnnotationAppender.finish() methods to be able to use the annotation appender outside a context window

  • Add category_ids_mapping optional argument to AnnotationAppender and related functions in order to remap the category ids from predictions

  • Add flatten_paths to cAIpy export function, which lets you save a dataset without subfolders.

  • Add c2f standalone script to quickly open a caipy dataset into fiftyone

  • Add from_files function, similar to ``from_folder` but when you already know what files or file patterns you want in the root folder.

  • Add difftools in libia.utils to compute difference between datasets. Useful when we want to update something related to it (like fiftyone)

  • Add libia.utils.doc_utils for examples in docstring, with a dummy dataset creator

  • Add Examples in all methods of Dataset object.

  • Add Dataset.reset_index_from_mapping method to remap index of images and annotationbs dataframes

  • BREAKING Remove Dataset.reindex method and rename it Dataset.match_index to avoid confusion with pandas.reindex

  • Add “See Also” admonitions in many methods to link methods together and to see the related tutorial each time

  • Add schemas tutorial

Changed#

  • Caipy save is much faster

  • Up-to-date dependencies

  • from_coco function now has label_map option in case the categories field is empty in the input json

  • from_coco assumes category_id to be 0 in case it is absent from annotations fields. It will error if it’s not absent from ALL annotations though.

  • BREAKING Evaluator.predictions renamed to Evaluator.predictions_dictionary for better clarity

  • BREAKING DetectionEvaluator.compute_matches and DetectionEvaluator.compute_precision_recall have changed their predictions option to predictions_names for better clarity.

  • Dataset.merge now tries to fuse dataframes with overlapping ids, as long as the common subset is the same

  • Dataset.reset_index now accepts a start_image_id.

  • BREAKING Dataset.dataset_path is deprecated in favor Dataset.images_root, similar to Evaluator.

  • Introduce the optional dataset_name attribute to be used when dataset name is not the folder name of images root but can be deduced from the loader function, e.g. in from_caipy

  • dataset merging now merge image indexes before concatenating the annotations. Useful when merging a dataset with annotations and the same dataset with pre-annotations.

  • refactor dataset merge logic in a dedicated module

  • dataset addition falls back to realing_label_map in merge when a IncompatibleLabelMapsError is raised.

  • add create_split_folder option in dataset_to_darknet function and related Dataset methods, allowing to save all images of a particular split in its dedicated folder.

  • Dataset.get_split now accept None value to get all images with a null split value if needed.

  • BREAKING Dataset.remap_from_DataFrame renamed to Dataset.remap_from_dataframe

  • Replace warning types from UserWarning to the right warning type (DeprecationWarning or RuntimeWarning)

  • Add pandas style Dataset.loc, Dataset.iloc, Dataset.loc_annot and Dataset.iloc_annot indexers, along with filter_images and filter_annotations method.

  • Add record_fo_ids options in Dataset.to_fiftyone and DetectionEvaluator.to_fiftyone methods to keep track of fiftyone’s UUID of each corresponding image and annotation.

  • Add markdownlint pre-commit hook (and make markdown documents compliant with it)

  • Add --watch argument in caipy_to_fiftyone script to perform live update of fiftyone datasets each time a file is modified in the caipy dataset. Useful when constructing a dataset progressively.

  • Add start_annotations_id option to Dataset.reset_index method.

  • Add supplementary checks and formatting to the Dataset basic constructor.

  • Add more explanation on crowd counting tutorial.

Fixed#

  • Get split does not rely on split being present in annotations anymore

  • crowdhuman head visibility is unknown

  • Class remapping is now compatible when label map is only a subset of remap dict

  • PNG to JPG conversion now works for RGBA images (note that the Alpha channel will be lost)

  • to_yolov5 now automatically convert split values like eval and valid to their yolov5 accepted equivalent (resp. test and val)

  • fix DetectionEvalutator.matches being tied to the class instead of the instance.

  • fix dependencies problem: sklearn is in core dependencies and matplotlib in optional “plot-utils” group

  • fix yolov7 problem, image path in txt files are also absolute. Please don’t use yolov7 export if you don’t need to, the dataset specs are terrible.

  • diverse pycharm warnings fixed

  • type hint of from_folder improved

  • from_folder method does not crash when folder is empty, but returns an empty dataset with a warning.

  • Warnings and pyright errors from last pandas version are suppressed

  • Use tight layout for confusion matrix plot result

  • Use json normalize when loading COCO so that it can be converted to fiftyone

  • Skip processing steps when converting an empty dataset to fiftyone or when appending empty annotations to the dataset with the annotation appender context manager

  • Prevent annotations index to be reset when using annotations appender

  • Prevent loss of dataset name when calling merge, reset_index, remap_classes

Removed#

  • libia.model subpackage (dead legacy code) got deleted

[1.4.0] - 2023-02-01#

Added#

  • Add CrowdHuman loading module See https://www.crowdhuman.org/

  • Add darknet_generic loading module

  • Add more test to improve coverage

  • introduce a BBOX_COLUMN_NAMES convention for bounding column names in dataset’s annotation dataframe

Fixed#

  • sum of datasets is now functional and tested (was not working before)

[1.3.1] - 2023-01-16#

Fixed#

  • Fix bug regarding confidence subsampling for PR curves

  • Proper extremal point for PR curves

  • Caipy split stays to None if no split is given when loading and data is in root

  • Caipy save keep added attributes during runtime when saving

[1.3.0] - 2023-01-10#

Added#

  • Add remove empty images method to dataset

  • Add remove emptied images option in remap classes

  • Add remove not mapped classes option in remap classes (not mapped were always removed before)

  • Add f_scores_betas to compute all wanted F-scores, F1, F0.5, F2, etc…

Changed#

  • PR curves are now indexed by recall with 101 evenly spaced values between 0 and 1 by default. The old behaviour can be retrieved by setting the option index_column to None.

  • Reworked evaluation demo

  • Improved documentation

[1.2.0] - 2023-01-06#

Added#

  • Add bounding box converter

  • Add image folder io, when input is simply a folder with images, but no annotation

  • load caipy generic does not have to specify an image folder anymore

  • conversion to fiftyone for datasets and evaluators

  • bugfix regarding annotation index when it’s duplicated

  • group continuous data with either interval labels (by default), mid-point, mean point or median point

Changed#

  • BREAKING evaluation predictions and matches are now dictionaries and can be used to evaluate multiple predictions sets at the same time

  • BREAKING group type alias is now either a column or a ContinuousGroup object (a dictionary that does the same thing but with better checking)

Fixed#

  • Fix several failing pyright tests because pandas stubs was updated

[1.1.0] - 2022-11-17#

Added#

  • Add caipy generic format

  • Add testing module in utils

  • More thorough tests for io

  • More complete notebook for demo_dataset

Fixed#

  • pre-commit’s flake8 repo url was moved from gitlab to gitHub

[1.0.0] - 2022-11-04#

Added#

  • dataset evaluation tool : see tutorials/demo_evaluation

  • dataset split tool : see tutorials/demo_split

  • new code checkers, including pyright and pandas stubs

[0.2.0] - 2022-07-18#

Added#

  • Features: Merge, Class remapping, etc.

On this page
Show Source