Dev

[adku1173]

No description provided.

[github-actions]

Output of diff-cover 🔍

-------------
Diff Coverage
Diff: origin/main...HEAD, staged and unstaged changes
-------------
src/acoupipe/datasets/base.py (62.9%): Missing lines 53,79-80,145,382,557-558,565,567-568,571-578,580,582-584,586
src/acoupipe/datasets/experimental.py (83.1%): Missing lines 362,364-366,368-373,632-633,644
src/acoupipe/datasets/features.py (98.5%): Missing lines 802
src/acoupipe/datasets/ir.py (47.9%): Missing lines 28,32-33,38,64-66,73-76,88-96,109,111-114
src/acoupipe/datasets/micgeom.py (100%)
src/acoupipe/datasets/spectra_analytic.py (88.9%): Missing lines 168-170
src/acoupipe/datasets/synthetic.py (67.3%): Missing lines 419,441-442,447-448,562-563,565,793,833,917-919,927,931-934,938-939,941-942,944-946,948,950-951,953-954,956,958-965,969-971,973-974,978-981,985-993,1003,1011-1014,1016-1025,1027-1028,1035,1039-1040,1051,1057,1089-1091,1103
src/acoupipe/datasets/utils.py (65.0%): Missing lines 69-71,78-79,86,90
src/acoupipe/loader.py (50.0%): Missing lines 79-80
src/acoupipe/pipeline.py (60.0%): Missing lines 317-318
src/acoupipe/sampler.py (83.3%): Missing lines 480
src/acoupipe/writer.py (57.8%): Missing lines 155-156,158,161-162,165-167,172,179,190,230-236,268-274,314,317
-------------
Total:   639 lines
Missing: 187 lines
Coverage: 70%
-------------

[Copilot]

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

This PR refactors dataset generation and testing infrastructure, adds optional impulse-response (IR) support and a new SRIRACHA dataset path, and modernizes TFRecord/HDF5 serialization and project tooling (uv, org CI, docs theme).

Changes:

• Centralizes pytest fixtures into tests/conftest.py, migrates dataset validation from .npy files to snapshot-style checks, and adds IR optional-dependency tests.
• Refactors feature construction to use BaseFeatureCatalog + BaseFeatureCollectionBuilder, adds TFRecord encoding inference + dynamic-shape metadata, and introduces remote_args for Ray actor resource configuration.
• Adds IR helper module (acoupipe.datasets.ir), new SRIRACHA dataset wiring/docs, and updates packaging/docs/CI to a uv + org-workflow setup.

Reviewed changes

Copilot reviewed 47 out of 335 changed files in this pull request and generated 16 comments.

Show a summary per file

File	Description
tests/test_spectra.py	Removes local fixtures/imports in favor of shared conftest.py fixtures.
tests/test_sampler.py	Switches to shared fixtures; trims imports now provided via conftest.py.
tests/test_loader.py	Moves HDF5 temp-file fixture to conftest.py.
tests/test_ir.py	Adds tests for IR backend fallback and optional dependency error reporting.
tests/test_datasets.py	Migrates dataset value checks to snapshot-based assertions; adjusts multiprocessing skips.
tests/test_dataset_tf.py	Removes duplicated temp-dir/dataset fixtures to rely on conftest.py.
tests/create_validation_data.py	Deletes legacy script for creating .npy validation data.
tests/conftest.py	Introduces shared fixtures for datasets, samplers, pipelines, and HDF5 temp data.
src/acoupipe/writer.py	Refactors TFRecord encoders, adds dtype/shape inference and optional shape sidecar features.
src/acoupipe/sampler.py	Changes LocationSampler.sample_loc and how locations are applied to targets.
src/acoupipe/pipeline.py	Adds Ray remote_args plumbing and debug logging in SamplerActor.
src/acoupipe/loader.py	Wraps metadata loading with a catch-all exception handler and prints errors.
src/acoupipe/datasets/utils.py	Adds new calc_transfer and removes blockwise_transfer.
src/acoupipe/datasets/synthetic.py	Major feature-builder refactor; adds IR-backed ISM dataset/config; passes remote_args.
src/acoupipe/datasets/spectra_analytic.py	Adds effective DoF computation (df_eq) and changes window handling.
src/acoupipe/datasets/micgeom.py	Replaces hardcoded arrays with loading mic geometry from Acoular XML.
src/acoupipe/datasets/ir.py	Adds optional IR backends (gpuRIR/pyroomacoustics) with guided ImportErrors.
src/acoupipe/datasets/features.py	Moves to data-dependent feature execution; adds TF encoder inference integration.
src/acoupipe/datasets/experimental.py	Extends scenarios (SRIRACHA), refactors IR handling, adds remote_args passthrough.
src/acoupipe/datasets/base.py	Refactors default-feature building and changes public dataset method signatures; adds remote_args.
pyproject.toml	Updates Python/acoular constraints, replaces extras with dependency groups, adds pytest config and uv setup.
examples/point_source_example.py	Updates grid extent usage for plotting.
docs/source/index.rst	Updates datasets overview and citations; adjusts figure sizing and content.
docs/source/contents/script/quickstart.py	Updates plotting extent usage.
docs/source/contents/script/features.py	Updates plotting extent usage.
docs/source/contents/script/experimental.py	Updates example frequency and formatting; switches to grid extent property.
docs/source/contents/script/all_datasets.py	Adds a new script generating comparative plots across datasets/scenarios.
docs/source/contents/jupyter/modify.ipynb	Clears execution counts and updates extent usage.
docs/source/contents/jupyter/load_tfrecord.ipynb	Clears execution counts and updates extent usage.
docs/source/contents/jupyter/generate.ipynb	Updates extent usage.
docs/source/contents/install.rst	Rewrites install docs around uv/pip and dependency groups; updates Docker snippet formatting.
docs/source/contents/datasets/synthetic.rst	Adds dedicated DatasetSynthetic documentation page.
docs/source/contents/datasets/store.rst	Adds documentation page for HDF5/TFRecord storage.
docs/source/contents/datasets/sriracha.rst	Adds dedicated DatasetSRIRACHA documentation page.
docs/source/contents/datasets/miracle.rst	Adds dedicated DatasetMIRACLE documentation page.
docs/source/contents/datasets.rst	Replaces autoapi dataset pages with curated dataset pages + store page.
docs/source/conf.py	Switches theme to pydata-sphinx-theme, adds sphinx-design, configures AutoAPI skips.
docs/source/bib/refs.bib	Adds SRIRACHA and MIRACLE citations; expands AcouPipe reference fields.
docs/source/_static/css/custom_pydata_sphinx_theme.css	Adds custom layout styling for pydata-sphinx-theme.
README.rst	Major rewrite: installation, datasets overview (adds SRIRACHA), quickstart, citations.
.python-version	Pins dev Python to 3.13.
.github/workflows/tests.yml	Removes local test workflow (replaced by organizational CI).
.github/workflows/orga-ci.yml	Adds organizational CI workflow calls (tests, coverage, docs, lint, links).
.github/workflows/docs.yml	Migrates docs build to uv and Python 3.13; removes pull_request trigger.
.github/workflows/checks.yml	Removes local lint/link/cff workflows (replaced by organizational CI).
.github/actions/setup-hatch/action.yml	Removes hatch setup action (uv-based workflows replace it).

Comments suppressed due to low confidence (1)

src/acoupipe/sampler.py:1

• This change is very likely to break LocationSampler: target.loc is commonly a tuple (e.g. (x, y, z)), so target.loc.copy() will raise AttributeError. Even if target.loc is an array, loc[loc_axs, 0] assumes an extra dimension (shape (3, 1)), which won’t hold for a 1D (3,) location vector. Finally, assigning target.loc = new_loc changes the type from tuple to ndarray, which can break downstream Acoular expectations. Suggested fix: keep the previous conversion pattern (np.array(target.loc)) and update elements with 1D indexing (e.g. loc[loc_axs] = ...), and assign back using the expected type (likely tuple(loc)).

"""Random processes to sample values according to a specified random distribution (random variable).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

calc_transfer() currently ignores time_axis (padding always uses -1 and the FFT always uses axis=-1), assumes fftfreq is always a numeric array (will crash if None despite the docstring), and always returns H_sel.T which will reorder axes incorrectly for inputs with ndim>2. Additionally, the docstring claims integer fftfreq should be treated as bin indices, but the implementation always treats fftfreq as Hz. Suggested fix: honor time_axis via np.moveaxis, pad along the actual time_axis, handle fftfreq is None by returning all bins, and implement the integer-bin-index path separately (no Hz-to-bin conversion). Also remove the unused np.fft.rfftfreq(...) line or store it for validation.

[Copilot]

calc_transfer() currently ignores time_axis (padding always uses -1 and the FFT always uses axis=-1), assumes fftfreq is always a numeric array (will crash if None despite the docstring), and always returns H_sel.T which will reorder axes incorrectly for inputs with ndim>2. Additionally, the docstring claims integer fftfreq should be treated as bin indices, but the implementation always treats fftfreq as Hz. Suggested fix: honor time_axis via np.moveaxis, pad along the actual time_axis, handle fftfreq is None by returning all bins, and implement the integer-bin-index path separately (no Hz-to-bin conversion). Also remove the unused np.fft.rfftfreq(...) line or store it for validation.

Suggested change

	L = ir.shape[time_axis]
	# Select FFT length: power-of-two >= max(L, blocksize)
	nfft = 1 << (max(L, blocksize) - 1).bit_length()
	# Zero-pad IRs to nfft (never truncate)
	if nfft > L:
	pad_width = [(0, 0)] * ir.ndim
	pad_width[-1] = (0, nfft - L)
	ir = np.pad(ir, pad_width, mode='constant')
	# One-sided FFT with window matching the (zero-padded) length (i.e., use full nfft)
	H = np.fft.rfft(ir, n=nfft, axis=-1) # shape: (..., nfft//2+1)
	np.fft.rfftfreq(nfft, d=1.0 / fs)
	if np.any(fftfreq < 0.0) or np.any(fftfreq > fs / 2.0):
	msg = 'Requested frequencies must satisfy 0 <= f <= fs/2.'
	raise ValueError(msg)
	# Map to nearest bin; require exact (within tolerance) grid match
	idx_float = fftfreq * nfft / fs
	idx = np.rint(idx_float).astype(int)
	tol = 1e-9
	if np.max(np.abs(idx_float - idx)) > tol:
	msg = (
	'Some requested frequencies are not on the FFT grid. '
	'Provide integer bin indices (recommended), or choose f that matches k*fs/nfft.'
	)
	raise ValueError(msg)
	H_sel = np.take(H, idx, axis=-1)
	return H_sel.T
	ir = np.asarray(ir)
	ir_fft = np.moveaxis(ir, time_axis, -1)
	L = ir_fft.shape[-1]
	# Select FFT length: power-of-two >= max(L, blocksize)
	nfft = 1 << (max(L, blocksize) - 1).bit_length()
	# Zero-pad IRs to nfft (never truncate)
	if nfft > L:
	pad_width = [(0, 0)] * ir_fft.ndim
	pad_width[-1] = (0, nfft - L)
	ir_fft = np.pad(ir_fft, pad_width, mode='constant')
	# One-sided FFT with window matching the (zero-padded) length (i.e., use full nfft)
	H = np.fft.rfft(ir_fft, n=nfft, axis=-1) # shape: (..., nfft//2+1)
	freq_grid = np.fft.rfftfreq(nfft, d=1.0 / fs)
	if fftfreq is None:
	H_sel = H
	else:
	fftfreq = np.asarray(fftfreq)
	if np.issubdtype(fftfreq.dtype, np.integer):
	if np.any(fftfreq < 0) or np.any(fftfreq > blocksize // 2):
	msg = 'Requested bin indices must satisfy 0 <= k <= blocksize//2.'
	raise ValueError(msg)
	ratio = nfft // blocksize
	idx = fftfreq.astype(int) * ratio
	else:
	if np.any(fftfreq < 0.0) or np.any(fftfreq > fs / 2.0):
	msg = 'Requested frequencies must satisfy 0 <= f <= fs/2.'
	raise ValueError(msg)
	# Map to nearest bin; require exact (within tolerance) grid match
	idx_float = fftfreq * nfft / fs
	idx = np.rint(idx_float).astype(int)
	tol = 1e-9
	if np.max(np.abs(freq_grid[idx] - fftfreq)) > tol:
	msg = (
	'Some requested frequencies are not on the FFT grid. '
	'Provide integer bin indices (recommended), or choose f that matches k*fs/nfft.'
	)
	raise ValueError(msg)
	H_sel = np.take(H, idx, axis=-1)
	return np.moveaxis(H_sel, -1, time_axis)

[Copilot]

calc_transfer() currently ignores time_axis (padding always uses -1 and the FFT always uses axis=-1), assumes fftfreq is always a numeric array (will crash if None despite the docstring), and always returns H_sel.T which will reorder axes incorrectly for inputs with ndim>2. Additionally, the docstring claims integer fftfreq should be treated as bin indices, but the implementation always treats fftfreq as Hz. Suggested fix: honor time_axis via np.moveaxis, pad along the actual time_axis, handle fftfreq is None by returning all bins, and implement the integer-bin-index path separately (no Hz-to-bin conversion). Also remove the unused np.fft.rfftfreq(...) line or store it for validation.

Suggested change

	L = ir.shape[time_axis]
	# Select FFT length: power-of-two >= max(L, blocksize)
	nfft = 1 << (max(L, blocksize) - 1).bit_length()
	# Zero-pad IRs to nfft (never truncate)
	if nfft > L:
	pad_width = [(0, 0)] * ir.ndim
	pad_width[-1] = (0, nfft - L)
	ir = np.pad(ir, pad_width, mode='constant')
	# One-sided FFT with window matching the (zero-padded) length (i.e., use full nfft)
	H = np.fft.rfft(ir, n=nfft, axis=-1) # shape: (..., nfft//2+1)
	np.fft.rfftfreq(nfft, d=1.0 / fs)
	if np.any(fftfreq < 0.0) or np.any(fftfreq > fs / 2.0):
	msg = 'Requested frequencies must satisfy 0 <= f <= fs/2.'
	raise ValueError(msg)
	# Map to nearest bin; require exact (within tolerance) grid match
	idx_float = fftfreq * nfft / fs
	idx = np.rint(idx_float).astype(int)
	tol = 1e-9
	if np.max(np.abs(idx_float - idx)) > tol:
	msg = (
	'Some requested frequencies are not on the FFT grid. '
	'Provide integer bin indices (recommended), or choose f that matches k*fs/nfft.'
	)
	raise ValueError(msg)
	H_sel = np.take(H, idx, axis=-1)
	return H_sel.T
	ir = np.asarray(ir)
	ir_work = np.moveaxis(ir, time_axis, -1)
	L = ir_work.shape[-1]
	# Select FFT length: power-of-two >= max(L, blocksize)
	nfft = 1 << (max(L, blocksize) - 1).bit_length()
	# Zero-pad IRs to nfft (never truncate)
	if nfft > L:
	pad_width = [(0, 0)] * ir_work.ndim
	pad_width[-1] = (0, nfft - L)
	ir_work = np.pad(ir_work, pad_width, mode='constant')
	# One-sided FFT with window matching the (zero-padded) length (i.e., use full nfft)
	H = np.fft.rfft(ir_work, n=nfft, axis=-1) # shape: (..., nfft//2+1)
	if fftfreq is None:
	H_sel = H
	else:
	fftfreq = np.asarray(fftfreq)
	if np.issubdtype(fftfreq.dtype, np.integer):
	if np.any(fftfreq < 0) or np.any(fftfreq > blocksize // 2):
	msg = 'Requested bin indices must satisfy 0 <= k <= blocksize//2.'
	raise ValueError(msg)
	ratio = nfft // blocksize
	idx = fftfreq.astype(int) * ratio
	else:
	if np.any(fftfreq < 0.0) or np.any(fftfreq > fs / 2.0):
	msg = 'Requested frequencies must satisfy 0 <= f <= fs/2.'
	raise ValueError(msg)
	# Map to nearest bin; require exact (within tolerance) grid match
	idx_float = fftfreq * nfft / fs
	idx = np.rint(idx_float).astype(int)
	tol = 1e-9
	if np.max(np.abs(idx_float - idx)) > tol:
	msg = (
	'Some requested frequencies are not on the FFT grid. '
	'Provide integer bin indices (recommended), or choose f that matches k*fs/nfft.'
	)
	raise ValueError(msg)
	H_sel = np.take(H, idx, axis=-1)
	return np.moveaxis(H_sel, -1, time_axis)

[Copilot]

calc_transfer() currently ignores time_axis (padding always uses -1 and the FFT always uses axis=-1), assumes fftfreq is always a numeric array (will crash if None despite the docstring), and always returns H_sel.T which will reorder axes incorrectly for inputs with ndim>2. Additionally, the docstring claims integer fftfreq should be treated as bin indices, but the implementation always treats fftfreq as Hz. Suggested fix: honor time_axis via np.moveaxis, pad along the actual time_axis, handle fftfreq is None by returning all bins, and implement the integer-bin-index path separately (no Hz-to-bin conversion). Also remove the unused np.fft.rfftfreq(...) line or store it for validation.

Suggested change

	L = ir.shape[time_axis]
	# Select FFT length: power-of-two >= max(L, blocksize)
	nfft = 1 << (max(L, blocksize) - 1).bit_length()
	# Zero-pad IRs to nfft (never truncate)
	if nfft > L:
	pad_width = [(0, 0)] * ir.ndim
	pad_width[-1] = (0, nfft - L)
	ir = np.pad(ir, pad_width, mode='constant')
	# One-sided FFT with window matching the (zero-padded) length (i.e., use full nfft)
	H = np.fft.rfft(ir, n=nfft, axis=-1) # shape: (..., nfft//2+1)
	np.fft.rfftfreq(nfft, d=1.0 / fs)
	if np.any(fftfreq < 0.0) or np.any(fftfreq > fs / 2.0):
	msg = 'Requested frequencies must satisfy 0 <= f <= fs/2.'
	raise ValueError(msg)
	# Map to nearest bin; require exact (within tolerance) grid match
	idx_float = fftfreq * nfft / fs
	idx = np.rint(idx_float).astype(int)
	tol = 1e-9
	if np.max(np.abs(idx_float - idx)) > tol:
	msg = (
	'Some requested frequencies are not on the FFT grid. '
	'Provide integer bin indices (recommended), or choose f that matches k*fs/nfft.'
	)
	raise ValueError(msg)
	H_sel = np.take(H, idx, axis=-1)
	return H_sel.T
	ir = np.asarray(ir)
	time_axis = np.core.numeric.normalize_axis_index(time_axis, ir.ndim)
	L = ir.shape[time_axis]
	# Select FFT length: power-of-two >= max(L, blocksize)
	nfft = 1 << (max(L, blocksize) - 1).bit_length()
	# Work on a view with time axis last so padding/FFT/selection are applied consistently
	ir_moved = np.moveaxis(ir, time_axis, -1)
	# Zero-pad IRs to nfft (never truncate)
	if nfft > L:
	pad_width = [(0, 0)] * ir_moved.ndim
	pad_width[-1] = (0, nfft - L)
	ir_moved = np.pad(ir_moved, pad_width, mode='constant')
	# One-sided FFT with window matching the (zero-padded) length (i.e., use full nfft)
	H = np.fft.rfft(ir_moved, n=nfft, axis=-1) # shape: (..., nfft//2+1)
	if fftfreq is None:
	H_sel = H
	else:
	fftfreq = np.asarray(fftfreq)
	if np.issubdtype(fftfreq.dtype, np.integer):
	if np.any(fftfreq < 0) or np.any(fftfreq > blocksize // 2):
	msg = 'Requested bin indices must satisfy 0 <= k <= blocksize//2.'
	raise ValueError(msg)
	scale = nfft // blocksize
	idx = fftfreq.astype(int) * scale
	else:
	if np.any(fftfreq < 0.0) or np.any(fftfreq > fs / 2.0):
	msg = 'Requested frequencies must satisfy 0 <= f <= fs/2.'
	raise ValueError(msg)
	# Map to nearest bin; require exact (within tolerance) grid match
	idx_float = fftfreq * nfft / fs
	idx = np.rint(idx_float).astype(int)
	tol = 1e-9
	if np.max(np.abs(idx_float - idx)) > tol:
	msg = (
	'Some requested frequencies are not on the FFT grid. '
	'Provide integer bin indices (recommended), or choose f that matches k*fs/nfft.'
	)
	raise ValueError(msg)
	H_sel = np.take(H, idx, axis=-1)
	return np.moveaxis(H_sel, -1, time_axis)

[Copilot]

calc_transfer() currently ignores time_axis (padding always uses -1 and the FFT always uses axis=-1), assumes fftfreq is always a numeric array (will crash if None despite the docstring), and always returns H_sel.T which will reorder axes incorrectly for inputs with ndim>2. Additionally, the docstring claims integer fftfreq should be treated as bin indices, but the implementation always treats fftfreq as Hz. Suggested fix: honor time_axis via np.moveaxis, pad along the actual time_axis, handle fftfreq is None by returning all bins, and implement the integer-bin-index path separately (no Hz-to-bin conversion). Also remove the unused np.fft.rfftfreq(...) line or store it for validation.

Suggested change

	# Zero-pad IRs to nfft (never truncate)
	if nfft > L:
	pad_width = [(0, 0)] * ir.ndim
	pad_width[-1] = (0, nfft - L)
	ir = np.pad(ir, pad_width, mode='constant')
	# One-sided FFT with window matching the (zero-padded) length (i.e., use full nfft)
	H = np.fft.rfft(ir, n=nfft, axis=-1) # shape: (..., nfft//2+1)
	np.fft.rfftfreq(nfft, d=1.0 / fs)
	if np.any(fftfreq < 0.0) or np.any(fftfreq > fs / 2.0):
	msg = 'Requested frequencies must satisfy 0 <= f <= fs/2.'
	raise ValueError(msg)
	# Map to nearest bin; require exact (within tolerance) grid match
	idx_float = fftfreq * nfft / fs
	idx = np.rint(idx_float).astype(int)
	tol = 1e-9
	if np.max(np.abs(idx_float - idx)) > tol:
	msg = (
	'Some requested frequencies are not on the FFT grid. '
	'Provide integer bin indices (recommended), or choose f that matches k*fs/nfft.'
	)
	raise ValueError(msg)
	H_sel = np.take(H, idx, axis=-1)
	return H_sel.T
	# Move the requested time axis to the end so padding/FFT operate on the
	# actual time dimension regardless of the original layout.
	ir = np.moveaxis(ir, time_axis, -1)
	# Zero-pad IRs to nfft (never truncate)
	if nfft > L:
	pad_width = [(0, 0)] * ir.ndim
	pad_width[-1] = (0, nfft - L)
	ir = np.pad(ir, pad_width, mode='constant')
	# One-sided FFT along the time axis (now the last axis).
	H = np.fft.rfft(ir, n=nfft, axis=-1) # shape: (..., nfft//2+1)
	if fftfreq is None:
	idx = np.arange(H.shape[-1])
	else:
	fftfreq_arr = np.asarray(fftfreq)
	# Integer inputs are interpreted as direct FFT bin indices.
	if np.issubdtype(fftfreq_arr.dtype, np.integer):
	idx = fftfreq_arr.astype(int, copy=False)
	if np.any(idx < 0) or np.any(idx > nfft // 2):
	msg = 'Requested FFT bin indices must satisfy 0 <= k <= nfft//2.'
	raise ValueError(msg)
	else:
	# Non-integer inputs are interpreted as frequencies in Hz.
	if np.any(fftfreq_arr < 0.0) or np.any(fftfreq_arr > fs / 2.0):
	msg = 'Requested frequencies must satisfy 0 <= f <= fs/2.'
	raise ValueError(msg)
	# Map to nearest bin; require exact (within tolerance) grid match.
	idx_float = fftfreq_arr * nfft / fs
	idx = np.rint(idx_float).astype(int)
	tol = 1e-9
	if np.max(np.abs(idx_float - idx)) > tol:
	msg = (
	'Some requested frequencies are not on the FFT grid. '
	'Provide integer bin indices (recommended), or choose f that matches k*fs/nfft.'
	)
	raise ValueError(msg)
	H_sel = np.take(H, idx, axis=-1)
	return np.moveaxis(H_sel, -1, 0)

[Copilot]

The public API method signatures were reordered (e.g., generate(features, size, split=...) instead of generate(features, split, size, ...)), which is a breaking change for callers using positional arguments. Suggested fix (mandatory if backward compatibility matters): keep the old positional order and introduce keyword-only parameters for the new preferred order, or accept both orders via deprecation handling (e.g., detect split passed as int/size passed as str).

Suggested change

	def generate(self, features, size, split='training', f=None, num=0, start_idx=0, progress_bar=True):
	def generate(self, features, split='training', size=None, f=None, num=0, start_idx=0, progress_bar=True):

[Copilot]

The public API method signatures were reordered (e.g., generate(features, size, split=...) instead of generate(features, split, size, ...)), which is a breaking change for callers using positional arguments. Suggested fix (mandatory if backward compatibility matters): keep the old positional order and introduce keyword-only parameters for the new preferred order, or accept both orders via deprecation handling (e.g., detect split passed as int/size passed as str).

[Copilot]

The public API method signatures were reordered (e.g., generate(features, size, split=...) instead of generate(features, split, size, ...)), which is a breaking change for callers using positional arguments. Suggested fix (mandatory if backward compatibility matters): keep the old positional order and introduce keyword-only parameters for the new preferred order, or accept both orders via deprecation handling (e.g., detect split passed as int/size passed as str).

[Copilot]

The public API method signatures were reordered (e.g., generate(features, size, split=...) instead of generate(features, split, size, ...)), which is a breaking change for callers using positional arguments. Suggested fix (mandatory if backward compatibility matters): keep the old positional order and introduce keyword-only parameters for the new preferred order, or accept both orders via deprecation handling (e.g., detect split passed as int/size passed as str).

[Copilot]

The new calc_transfer() function introduces non-trivial behavior (FFT length selection, zero-padding, bin mapping, and grid validation) but no direct unit tests are added alongside it. Adding targeted tests (e.g., 1D/2D/3D IR shapes, time_axis variations, fftfreq=None, integer-bin vs Hz input, and exact-grid vs off-grid errors) would help prevent regressions and validate the new assumptions.