TensorFlow Model Analysis 0.52.0

Major Features and Improvements

• Adds official support for Python 3.12 and 3.13.
• Drops support for Python 3.9.
• Depends on tensorflow>=2.21.0,<2.22.0.
• Depends on protobuf>=6.31.1,<7.0.0.
• Depends on pyarrow>14.
• Updates the minimum Bazel version required to build TFMA to 7.4.1.

Bug fixes and other Changes

• Pickling Stability Architecture:
  • Refactored core test modules (rouge_test.py, stats_test.py, model_util_test.py, confidence_intervals_util_test.py, metrics_plots_and_validations_evaluator_test.py) to use a new class-based matcher architecture.
  • Replaced nested closures with module-level classes (e.g., CheckResult, CheckResultMean) to ensure full serializability for PrismRunner on Python 3.13.
  • Removed self (test instance) capture in Beam matchers to resolve RuntimeError: Unable to pickle fn during distributed execution.
  • Enabled --no_save_main_session for all Beam pipelines in the test suite to prevent unintentional serialization of the main session and shared resources.
• NumPy 2.0 & Python 3.13 Compatibility:
  • Standardized on safe scalar extraction by replacing float(ndarray) with .item() or metric_util.safe_to_scalar in aggregation, attributions, calibration, flip metrics, and NDCG modules to resolve TypeError in Beam pipelines.
  • Fixed a batching bug in flip_metrics.py to correctly process all examples in a Beam batch.
  • Implemented robust, warning-free division in AUC, PR AUC, and confusion matrix calculations using np.divide with where clauses to prevent RuntimeWarning.
  • Fixed TypeError in poisson_bootstrap.py by removing redundant size argument from poisson(1, 1) to return a scalar.
  • Implemented kind='stable' sort in metric_util.top_k_indices for deterministic tie-breaking.
• Bug Fixes and Functional Corrections:
  • Fixed a critical regression in metric_util.py where SubKey(k=k) incorrectly selected the first prediction instead of the requested k-th largest prediction.
  • Fixed UnparsedFlagAccessError in ModelSignaturesDoFn tests by removing direct absl.flags access in pickling-sensitive contexts.
  • Removed obsolete @unittest.expectedFailure decorators from tests that are now passing in the stabilized environment.
  • Fixed various indentation and syntax errors in utility tests.
  • Improved virtual environment relocation strategy to resolve Bazel sandbox access issues for numpy and other C-extension headers.
  • Fixed false_omission_rate in binary_confusion_matrices.py to return NaN when undefined, resolving proto mismatches in confusion_matrix_plot_test.py and score_distribution_plot_test.py.
  • Fixed NotFoundError in model_eval_lib_test.py by ensuring temporary directories exist before writing files using tf.io.gfile.makedirs.
  • Added missing numpy imports in Beam-based modules to fix NameError regressions.

Breaking Changes

• Python 3.9 is no longer supported. The minimum supported Python version is now 3.10.

Deprecations

• N/A