Aggregate logs in evaluate

[LarsKue]

Fixes #481

[codecov]

Codecov Report

Attention: Patch coverage is 90.25974% with 15 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...ximators/backend_approximators/jax_approximator.py	91.52%	5 Missing ⚠️
...s/backend_approximators/tensorflow_approximator.py	89.58%	5 Missing ⚠️
...mators/backend_approximators/torch_approximator.py	89.36%	5 Missing ⚠️

Files with missing lines	Coverage Δ
...ximators/backend_approximators/jax_approximator.py	94.11% <91.52%> (-3.56%)	⬇️
...s/backend_approximators/tensorflow_approximator.py	90.90% <89.58%> (-3.54%)	⬇️
...mators/backend_approximators/torch_approximator.py	91.30% <89.36%> (-4.16%)	⬇️

[Copilot]

Pull Request Overview

Implements cumulative logging and averaging in the evaluate method for all backend approximators to fix metrics aggregation issues.

• Introduce _aggregate_logs and _mean_logs helpers to accumulate and normalize batch metrics.
• Override evaluate in Torch, TensorFlow, and JAX approximators using their respective *EpochIterator and callback flows.
• Wire up Keras CallbackList and ensure per-batch callbacks with aggregated logs.

Reviewed Changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.

File	Description
bayesflow/approximators/backend_approximators/torch_approximator.py	Added _aggregate_logs, _mean_logs, and updated evaluate with TorchEpochIterator and callbacks.
bayesflow/approximators/backend_approximators/tensorflow_approximator.py	Added _aggregate_logs, _mean_logs, and updated evaluate with TFEpochIterator and callbacks.
bayesflow/approximators/backend_approximators/jax_approximator.py	Added _aggregate_logs, _mean_logs, and updated evaluate with JAXEpochIterator, state sync, and callbacks.

Comments suppressed due to low confidence (2)

bayesflow/approximators/backend_approximators/jax_approximator.py:87

• [nitpick] The loop variable 'iterator' shadows the epoch_iterator and may be confusing; consider renaming it to 'batch_data' or similar to clarify its purpose.

for step, iterator in epoch_iterator:

bayesflow/approximators/backend_approximators/tensorflow_approximator.py:31

• Add unit tests for this new evaluate implementation to verify that log aggregation and averaging behave as expected across multiple batches.

def evaluate(

[Copilot]

Pull Request Overview

Adds an aggregate option to the evaluate method in all backend approximators, allowing batch-wise metrics to be summed and averaged rather than overwritten each step.

• Introduce _aggregate_fn and _reduce_fn in evaluate to accumulate and average metrics.
• Add aggregate and return_dict parameters to control output format.
• Ensure consistency in callback invocation and test function setup across Torch, TensorFlow, and JAX backends.

Reviewed Changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
bayesflow/approximators/backend_approximators/torch_approximator.py	Added evaluate override with optional aggregation logic
bayesflow/approximators/backend_approximators/tensorflow_approximator.py	Added evaluate override with optional aggregation logic
bayesflow/approximators/backend_approximators/jax_approximator.py	Added evaluate override with optional aggregation logic and JAX state handling

Comments suppressed due to low confidence (4)

bayesflow/approximators/backend_approximators/torch_approximator.py:26

• Add unit tests covering both aggregate=True and aggregate=False paths to verify that metrics are correctly summed and averaged.

        aggregate=False,

bayesflow/approximators/backend_approximators/torch_approximator.py:29

• Add a call to self._assert_compile_called("evaluate") at the start of evaluate to ensure the model has been compiled before evaluation.

# TODO: respect compiled trainable state

bayesflow/approximators/backend_approximators/jax_approximator.py:26

• The default aggregate=True in JAX differs from aggregate=False in the Torch and TensorFlow backends. Align the default value for consistency across backends.

        aggregate=True,

bayesflow/approximators/backend_approximators/torch_approximator.py:16

• This new evaluate method lacks docstrings for the aggregate and return_dict parameters; please add descriptions and expected behavior.

    def evaluate(

[LarsKue]

Superceded by #485