Skip to content

Latest commit

 

History

History
119 lines (97 loc) · 6.87 KB

CHANGELOG.md

File metadata and controls

119 lines (97 loc) · 6.87 KB
Raw

Changelog

All notable changes to this project will be documented in this file. This format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[0.1.2] - 2024-10-28

Added

  • CLI device specification
  • CLI compute arguments in reports dir
  • Preprocessing transform in image datasets when using the API

Fixed

  • Moved sklearn and gudhi dependencies to the main dependency tree

Changed

  • Default image feature extractor is now vit_b_32
  • Confusing synthetic and input metric goals where aggregated to quality, privacy, annotation and utility categories
  • Moved metrics to sepecific folders based on metric_group (feature-based, data-based)

[0.1.1] - 2024-10-24

Fixed

  • Fixed project configuration conflict between setup.py and pyproject.toml by reverting to poetry as main build engine
  • Updated dependencies in pyproject.toml and ignoring versions in requirements folder
  • Fixed error in validation type for the cli
  • Fixed example tabular and time-series scripts to use cli application

Removed

  • Removed setup.py file as it is deprecated in favor of pyproject.toml
  • Removed cookiecutter template files as they are not needed for the project

Changed

  • Changed default image extractor model from dino_vits8 to vit_b_32

[0.1.0] - 2024-10-16

  • Bumped version to 0.1.0 to indicate the first stable release of the package

[0.0.41] - 2024-08-31

Added

  • Documentation for image, tabular and time-series class metrics
  • Tabular and time-series unit tests for direct import from the package
  • Dummy notebook with examples of quality evaluation on the imagenette dataset
  • Mkdocs documentation for the package (only available locally and for metric classes)
    • Each modality has a seciton in the documentation site
    • Metrics are organized as input_val and synthesis_val groups
    • Groups may be divided into other subgroups depending on metrics' specificities
  • Added distance metric parameter for the PRDC metrics

Fixed

  • Entire code structure moved to package level under src/
    • Allows for easire import of metrics and models and correct package builds
  • Package submodules defined in upper init.py files to allow for direct import of metrics
  • Updated mkdocs from 1.5.3 to 1.6.1 to fix a bug in the previous version
  • Updated mkdocs-material from 9.4.14 to 9.5.34 to fix a bug in the previous version
  • Updated mkdocstrings from 0.24.0 to 0.26.0 to fix a bug in the previous version

[0.0.40] - 2024-07-31

Added

  • Standard version of main.py and documentation on how to run locally
  • Support for specific feature extractors in the main.py script
  • Added key -> value/array dtypes for metric results
  • API data access is now controlled by the API_DATA_ROOT environment variable in .env
  • Modality batch sizes are now controlled by the <MODALITY>_BATCH_SIZE environment variables in .env
  • Reorganized metrics into metric group categories to account for shared requirements. A diagram explaining the new metric structure can be found here
    • input metrics are grouped under quality, privacy or annotation categories
    • Synthesized metrics are grouped under feature or data categories
  • Older and confusing metric_subgoals were moved to the metric_goal field
    • metric_goal field is an optional subcategory of the metric_group field
  • Integrated image COCO annotation metrics into the API
    • Annotation files must be present under the input_val/annotations/ directory
    • Annotation files must be in COCO format and the filename given as a parameter in the API
  • API unit testing for every modality with FastAPI testing utilities
  • Documentation of shared metrics and image metrics
  • Added annotation_file field to the API for image datasets

Fixed

  • Using resize of 224 in DINO-ViT for image datasets
  • Updated pycaret from 3.1.0 to 3.3.1 due to a bug in the previous version
  • Removed redundant hooks in the Makefile
  • Text datasets are now kept under text/input_val/dataset/ for consistency with the other modalities
  • Input layers now extend the InputLayer abastract class for consistency
  • Renamed input variables in input layers from data_src1, data_src2 to reference and target
  • Refactored code struture from properties-based structure to a more modular class-based structure
    • Metrics are now classes that inherit from a base Metric class
    • Each metric has a compute method that returns the metric value
    • Older property attributes are now class attributes of the specific metric class
    • Introduced the MetricResult class to enforce and validate return types of the compute method
    • Input layers provide batched samples to the metrics via the batched_samples property
    • Embedder models are now passed to the get_embeddings method of the input layers (allows for more flexibility)
    • Moved common API and CLI code to the commonpackage to avoid code repetition
    • Added the ComputationManager class to manage the computation of metrics. Allows for the batch computation and merger of metric results and the computation of metrics in parallel. Also keeps track of a shared state of the metrics and their results.
  • Removed FhP git dependencies from the pyproject.toml file (not needed)
  • Dockerfile now uses only the requirements-prod.txt file for production builds
  • Dockerfile pip install with torch-cpu source for lighter builds
  • Removed unused environment variables from the .env file
  • Standard error messages being returned in the API for invalid requests

Changed

  • evaluation_level field in the API is now optional
  • metric_goal field in the API is now optional
  • metric_group field in the API is mandatory and indicates the group of the metric
  • annotation_file optional field for the API to specify the annotation file for image datasets
  • metric_subgoals field in the API was removed and replaced by the metric_goal field
  • metric_plots field in the returned JSON of the API was removed
    • Each metric now has a plot_params field that contains the plot parameters
  • Each metric now has an optional stats field that contains additional statistical information of the metric results
    • Previous iterations returned these statistics in separate metric results
  • The response body now contains an errors field to indicate any errors caught during the metric computation
    • This is only used in known errors/warnings that are known to occur in a specific metric (e.g. a metric that is supposed to be between 0 and 1 but is not due to numerical instability). Other API errors are still returned as HTTP status codes.
  • Renamed metric result type field to dtype to avoid conflicts with Python's type function
  • Renamed the dtype in the metric result to subtype