Skip to content

Changelog

2.0.1 (2026-07-20) — Fix pytorch extra, add meta-package scenarios

  • Fix: pytorch extra in pyproject.toml incorrectly depended on dqm-ml-images instead of dqm-ml-pytorch
  • Added: Meta-package test scenarios (10–13) exercising dqm-ml[job], [pytorch], [images], and [all] extras end-to-end
  • Removed: test_pypi_prerelease.sh — temporary script used while test.pypi.org configuration was not ready; now superseded by test_testpypi_prerelease.sh
  • Improved: test_pypi_release.sh now uses explicit --index-url for PyPI resolution
  • Updated: Packaging documentation to reflect 13 scenarios and script removal

2.0.0 (2026-07-17) Release

  • merged in main
  • tagged 2.0.0

2.0.0-rc4 (2026-07-17) — ProcessorRunner improvements, package isolation testing, configurable seeds

This release addresses review feedback from 2.0.0-rc3, enhancing ProcessorRunner with direct image-to-domain-gap support, adding package isolation smoke tests, and introducing configurable test seeds.

ProcessorRunner Enhancements

  • ProcessorRunner.run_gap() now accepts raw images directly via optional features parameter — compute domain gaps from image bytes without manual embedding extraction.
  • New _df_to_record_batch() helper for robust DataFrame-to-PyArrow conversion, handling FixedSizeListArray for embeddings.
  • Renamed metrics_processors parameter to processors in run() for clarity.
  • Separated select_columns intermediate results from final batch_features to avoid metric pollution.

Package Isolation Testing

  • New tests/packaging/ directory with smoke tests for all 5 packages: dqm-ml, dqm-ml-core, dqm-ml-images, dqm-ml-job, dqm-ml-pytorch.
  • Smoke scripts verify imports, basic functionality, and YAML-based execution in isolated virtual environments.
  • New tests/fixtures/packaging_fixtures.py with isolated_venv and wheels_dir fixtures.
  • New docs/packaging-tests.md and docs/testing.md documentation.

Configurable Test Seeds

  • New tests/utils/seeds.py with get_test_seed() function reading DQM_ML_TEST_SEED env var (default 42).
  • Session-scoped test_seed fixture in tests/conftest.py.
  • Updated ~35 test files and YAML templates to use configurable seeds instead of hardcoded values.
  • Seeds injected via tests/utils/jobs.py:generate_job() for YAML template generation.

CI Improvements

  • tqdm progress bars suppressed in CI via TQDM_DISABLE=1 in GitHub Actions and GitLab CI.
  • noxfile.py now sets DQM_ML_TEST_SEED=42 in all test sessions.
  • .mise.toml exports DQM_ML_TEST_SEED=42 for local development.

Documentation

  • New docs/testing.md with test organization, directory structure, and adding tests guide.
  • New docs/packaging-tests.md with package isolation testing documentation.
  • Improved READMEs for dqm-ml-core (+134 lines), dqm-ml-images (+100 lines), dqm-ml-pytorch (+252 lines).

Bug Fixes

  • S3 support fixes in CLI and processor (dqm_ml_job.cli, dqm_ml_job.utils.s3).
  • CI tag regex and PyPI publish workflow fixes.docs/index.md

2.0.0-rc3 (2026-06-30) — Pipeline DAG, Pydantic configs, example notebooks, and documentation

This release introduces a consistent configuration with Pydantic models and validation, adds a topological processor DAG, example notebooks and a user guide, improves domain-gap algorithms and image embedding, and refactors tests to generate synthetic data inline.

  • fix: #13 - Example Notebooks
  • fix: #10 - User guide

Pipeline Refactoring

  • Topological sort for processor DAG — generators before consumers — with configurable execution order via DatasetJob.execute() → topological_sort().
  • Accumulate-then-flush mode for single-path output patterns.
  • Memory-threshold parsing and configurable compute_max_memory.

Configuration Models (Pydantic)

  • New dqm_ml_core.models package with Pydantic models for processor, output, dataloader, and global configs.
  • Refactored parquet output writer to use Pydantic config with column exclude list, S3 path-prefix fallback, and FixedSizeListArray metric filtering.
  • Refactored S3 filesystem with StorageConfig model, retry strategies, role-based access, env-var fallback, and checksum validation.

Domain Gap Improvements

  • KLMVN variance-eps dampening and FID epsilon regularization.
  • PAD with CalibratedClassifierCV.
  • CMD multi-channel spatial moments.
  • Wildcard embedding column patterns and per-selection path-prefix injection.
  • New dqm_ml_core.utils.matching with include/exclude pattern resolution.

Image Embedding

  • Multi-input-column support, lazy model loading, auto device resolution.
  • Column prefix/suffix, S3 per-column path prefixes, configurable failure handling with rate thresholds.

Representativeness

  • Mean-std estimation, expected-counts method, per-column distribution parameters, improved model validation, path-prefix for sample paths.

Documentation & Examples

  • Restructured docs: flat files → docs/configuration/*.md, new metrics documentation, formal concepts page.
  • Added scenario example configs, debug pipeline script, and example notebooks for each metric.
  • Updated user guide in README with installation, metrics, configuration, and usage sections.
  • Schema generation utilities and docs/schema/config.json.
  • Google-format docstrings across source packages and test files.

Testing

  • New integration tests: batch invariance, pipeline ordering, pipeline scenarios, path prefix, output columns, pipeline data flow, full story, features embeddings.
  • New unit tests: matching, registry, representativeness, error policy, output configs, job, pandas loader, embedding edge cases, visual feature columns, domain gap processor, domain gap synthetic ordering.
  • New property-based tests: diversity, domain gap, representativeness, and visual features properties, plus visual features stress.
  • Stress image fixtures, expanded expected output YAML.
  • Updated config templates and benchmark stability improvements.

Code Quality

  • Tests refactored to reduce cognitive complexity
  • Synthetic data generated by tests at runtime, replacing 36 stale LFS-tracked test data files (parquet + JPGs).
  • Type checking improvments.
  • Removed LFS caching steps from CI workflows.
  • Cleaned up legacy CLI tests, stale example configs, and outdated documentation.

2.0.0-rc2 (2026-06-03) - S3 support, pipeline refactoring, new algorithms, code quality

This release adds S3 filesystem support, refactors the metric pipeline with dependency ordering, introduces diversity metrics and 4 new domain gap algorithms, and improves code quality.

  • fix: #18 - Integrate lexical and visual diversity metrics to v2
  • fix: #17 - Integrate categorical data diversity metric to v2
  • fix: #16 - Integrate Domain Gap metric MMD to v2
  • fix: #15 - Integrate Domain Gap metric PAD to v2
  • fix: #14 - Integrate Domain Gap metric CMD to v2
  • fix: #4 - Computed values differ between V1.1.2 and new implementation proposed for domain gap

S3 Filesystem Support

  • Shared S3 utility (dqm_ml_job.utils.s3.get_s3_filesystem) with env-var and dict-based configuration.
  • S3 support in Parquet data loading, output writing, image embedding, and visual features processors.
  • region parameter for providers like OVH S3.
  • Image loading improvements: S3-first fallback, Parquet-to-Pillow BytesIO fix.

Pipeline Refactoring

  • Topological sort for processor ordering (generators before consumers).
  • Delta metrics rewritten with PyArrow concat_tables instead of manual dict merging.
  • Accumulate-then-flush mode for single-path output patterns.
  • Dataloader column injection in feature outputs.

Diversity Metrics (new)

  • New DiversityProcessor with streaming batch-level computation.
  • Four indices: Simpson, Gini-Simpson, Shannon, and Richness.
  • Integration test fixtures and synthetic data generators.

Domain Gap Algorithms (new)

  • 4 new delta metrics: MMD-RBF, MMD-Poly, PAD, CMD with multi-layer support.
  • Image embedding refactoring and visual_features simplification for domain gap pipeline.
  • Updated representativeness processor for compatibility.
  • New domain gap benchmark suite and expanded integration fixtures.
  • Updated configuration templates and documentation.

.agents/ Refactoring

  • Replaced 364-line inline AGENTS.md with a shim pointing to .agents/AGENTS.md.
  • 7 standalone agent files: architecture, code style, git workflow, testing, project docs, and agent guidelines.

Code Quality & Bug Fixes

  • Renamed cryptic variables (self.fx → self.feature_extractor, vec → sum_fixed_to_vector, etc.).
  • Fixed h1/h2 variable-shadowing bug in Wasserstein loop (was using only first histogram dimension).
  • Fixed RepresentativenessProcessor double-count bug (exp_probs * total_count squaring total).
  • Reduced cognitive complexity below 15 in 5 modules by extracting focused helper methods (diversity, visual_features, domain_gap, image_embedding, job).
  • Replaced legacy np.random.* calls with seeded default_rng() Generator API for reproducible results.
  • Fixed type/security issues: missing gamma="auto" in SVC, logger.error → logger.exception, hardcoded /tmp → tempfile.gettempdir().
  • Fixed dataset_root_path being set to string "None" instead of Python None.
  • Consolidated duplicate string literals into module constants; fixed shellcheck [ → [[.
  • Added missing docstrings; corrected error message grammar; translated French comments.
  • Removed unused Dockerfile and container CI workflow.

Packaging & Versioning

  • Removed pinned version requirements for workspace-level setuptools-scm.
  • Fixed package discovery in pyproject.toml.
  • Configured PyTorch UV index for CPU torch downloads.

2.0.0-rc (2026-04-10) - Release Candidate for V2

This release is the V2.0.0 release candidate, featuring major refactoring, security fixes, and documentation improvements.

  • fix: #61 - Python 3.10/3.11 compatibility
  • fix: #60 - Fix parquet dataloader with image_path column
  • fix: #59 - Data selection tests for CLI
  • fix: #41 - Security warnings addressed
  • fix: #31 - Test package dependencies upgraded
  • fix: #29 - dqm-ml-pipeline renamed to dqm-ml-job
  • fix: #28 - Tests moved to workspace root
  • fix: #27 - Multi data selection configuration
  • fix: #26 - Domain gap computation for multiple selections
  • fix: #25 - Multiple data selection in same job
  • fix: #24 - Progress messages during post-processing
  • fix: #22 - Output folder auto-creation
  • fix: #11 - MkDocs documentation generation

Breaking Changes

  • Package rename: dqm-ml-v2 → dqm-ml (CLI entry point)
  • CLI command: dqm-ml-v2 → dqm-ml
  • Legacy removal: Removed dependency on legacy dqm-ml submodule

Security Fixes (10 vulnerabilities fixed)

Package Old Version New Version
strawberry-graphql 0.287.3 0.314.3
tornado 6.5.4 6.5.5
cryptography 46.0.4 46.0.7
requests - 2.33.0
pygments - 2.20.0

CVEs addressed: CVE-2026-35526, CVE-2026-35523, CVE-2026-31958, CVE-2025-47287, CVE-2026-26007, CVE-2026-34073, CVE-2026-37990, CVE-2026-25645

New Features

  • Test Strategy Documentation: Comprehensive testing docs in contributing.md with mermaid diagrams
  • Git LFS Guide: Added to quickstart.md for development setup
  • Test Fixtures: Session-scoped fixtures (coco_data, normal_dist, uniform_dist, etc.)
  • Test Utils: New tests/utils/ with helpers for files, jobs, plots
  • tqdm dependency added: Added to dqm-ml-job for progress display

Documentation Improvements

  • AGENTS.md: New contribution guidelines for AI agents
  • Test Strategy: Complete testing documentation with diagrams
  • Git LFS Section: Added to quickstart for large file handling
  • TODO/ROADMAP Sync: Aligned V1 metrics lists and version references
  • Mermaid Diagrams: Fixed dependency direction (core ← job, not →)
  • All docs updated: Replaced dqm-ml-v2 references with dqm-ml

Documentation Restructuring

  • New CLI Reference: docs/cli.md
  • New YAML Basics: docs/yaml_basics.md
  • New Data Loaders: docs/dataloaders.md
  • New Metrics Computation: docs/metrics_computation.md
  • New Python Compatibility: docs/python_compatibility.md
  • Mermaid diagrams: Config structure visualization
  • Navigation: Organized into separate pages in mkdocs.yml

Technical Changes

  • Pinned mkdocs-jupyter<0.26 to fix api-autonav compatibility
  • Refactored conftest.py - restructured fixtures
  • Added deprecation warning filters for strawberry and fiftyone

New CLI and Integration Tests

  • Test domain gap matrix between class pairs
  • Test quickstart examples with completeness config
  • New test fixtures and config files in tests/fixtures/getting_started/
  • Test data: COCO dataset with class information (102+ classes)
  • 18 tests added (unit + CLI)

Python Compatibility (Issue #61)

  • Python 3.10/3.11 support added
  • Fixed typing.override import: Changed from typing to typing_extensions in 9 package files
  • Fixed fiftyone lazy import in test fixtures (glob2 SyntaxError workaround)
  • Fixed domain gap delta metrics: Handle both pa.Array and scalar returns from compute_delta
  • Added _to_pa_array helper for robust type conversion in job.py
  • CI now tests: Python 3.10, 3.11, 3.12, 3.13
  • New documentation: docs/python_compatibility.md

Data Selection Features (Issues #25, #27, #59, #60)

  • split_by configuration: Create multiple selections from single dataloader based on column values
  • split_values: Specify values to create selections (e.g., split_by: class, split_values: [dog, cat])
  • filter configuration: Filter rows by column values
  • Fixed parquet dataloader with image_path column handling
  • New fixtures: coco_classes, bird, elephant for domain gap testing
  • Example configs: domain_gap_split_2classes.yaml, domain_gap_split_top10.yaml, domain_gap_matrix.yaml

Dependencies Updated

  • fiftyone: 1.14.1 → 1.13.0
  • torch: 2.10.0 → 2.11.0
  • torchvision: 0.25.0 → 0.26.0
  • Many other transitive dependencies upgraded

Testing

  • 72 tests passing (18 new tests)
  • Integration test fixtures for real data testing
  • CLI end-to-end tests
  • Test coverage reports available in docs/reports/
  • Python compatibility tests: 3.10, 3.11, 3.12, 3.13

1.1.6 (2026-02-10) - upgrade dependency and security check

This release is dedicated to rename dqm-ml-pipeline as sqm-ml-job, correct security issues detected by github, as well as initiated template for issues, feature and merge request

Other details on several changes

  • Security warning regarding dependency used in uv.lock as reference package version
  • urllib3 "2.6.2" => "2.6.3" : Decompression-bomb safeguards bypassed when following HTTP redirects (streaming API) High
  • torch 2.7.1 => 2.10.0 : PyTorch Improper Resource Shutdown or Release vulnerability Moderate
  • filelock "3.20.1" => "3.20.3" : filelock Time-of-Check-Time-of-Use (TOCTOU) Symlink Vulnerability in SoftFileLock Moderate
  • pynacl "1.6.1" => "1.6.2" : libsodium has Incomplete List of Disallowed Inputs Moderate

  • virtualenv "20.35.4" => "20.36.1" : virtualenv Has TOCTOU Vulnerabilities in Directory Creation Moderate

  • As fiftyone use a deprecated version of strawberry-graphql we pin the version to "strawberry-graphql==0.287.3

  • Other version change associate to uv sync --upgrade

  • default package version used for development and test upgraded, but without increasing min value dependencies

  • Remove dependency needed for test from default installation, only install when targetted nox command are executed

  • github-code-quality[bot] findings

  • SECURITY.md update

  • Unkow comand=> Unknown command miss spelling

  • add default issues template to github workflow

  • documentation presentation improvment

  • Add light/dark mode
  • Add dynamic horizontal size for content
  • Disable expand by default in navigation
  • Add navigation tab

  • Add navigation path

  • connection to sonarqube for code quality

1.1.5 (2026-02-09)

  • Quality correction / ci improvements Fixed
  • noxfile.py: Added a docs session to build the documentation. (Fixes #6)
  • noxfile.py: The format nox session now runs independently from the lint session. (Fixes #8)
  • dqm_ml_job/outputwriter/parquet.py: Ensured output directory is created if it does not exist, preventing crashes. (Fixes #22)

  • .github\ci.yml : adjust security permissions and other code quality checks

  • Repository organization changes :

  • Fixes : #28 : test moves to root of the workspace, and cover improved to 80%

  • fixes : #11 #12 : publication of mkdocs content document integrated

  • Documentation :

  • First documentation structure proposed, to be upgraded with feedback, it integrate fixes for #9 and #10

  • Conceptual change:

  • Notion of data selection introduced, and metrics are computed on a data selection, not a data loader.

  • delta metrics are compare between all dataselection available in the data selection.

  • Configuration changes

  • parquet data loader (fix : #27)
  • comparison of multiple data selection metrics fix : #25 and fix : #26

  • progression bar (optional) has been added to follow computation in the cli (fix : #24)

  • Limitations not delivered in v1.1.5

  • Change of dml-ml-pipeline to dqm-ml-job in future release for consistency #29

1.1.4 (2025-12-18)

  • github ci
  • add changelog
  • initiate release note and roadmap
  • match of dqm-ml version number

0.0.5 (2025-12-17)

Version use to check compatibility with legacy dqm-ml, after this we will match for this repository the versions with dqm-ml Version content :

  • dqm-ml-core : API proposal for future generic V2 api to allow generic usage of API
  • dqm-ml-images : beta version of feature computation for images in order to simplify computation on features computed
  • dqm-ml-pytorch : beta implementation of 3 of existing domain gap metrics relying on pytorch
  • dqm-ml-job : beta implementation of of metric computation cli from files, with grouping strategy (see roadmap)

This version allow use to generate metrics on welding uses case to demonstrate values of such an structural evolution