test(diarization): добавлен стенд сравнения CPU-моделей
- Зачем: - требовалось сравнить облегченные модели диаризации на старом CPU без изменения рабочей конфигурации. - Что: - добавлен воспроизводимый и возобновляемый прогон матрицы Pyannote FP32/INT8 с WeSpeaker и TitaNet. - зафиксированы измерения 108 ячеек, ограничения эксперимента и решение сохранить модель по умолчанию. - локальные записи в каталоге tmp исключены из индекса Git. - Проверка: - .venv/bin/python -m pytest. - /tmp/local-transcriber-ruff-bin/ruff check scripts/benchmarks/diarization_calibration.py tests/test_diarization_calibration.py.
This commit is contained in:
Dmitry Dementiev committed
1 parent
95f295c37c
commit
1e14f2687a
8 files changed
+2704
-234
No files matched your search
@@ -0,0 +1,851 @@
|
||||
import hashlib
|
||||
import importlib.util
|
||||
import json
|
||||
import sys
|
||||
from pathlib import Path
|
||||
from types import SimpleNamespace
|
||||
|
||||
import pytest
|
||||
|
||||
SCRIPT = (
|
||||
Path(__file__).parents[1] / "scripts" / "benchmarks" / "diarization_calibration.py"
|
||||
)
|
||||
SPEC = importlib.util.spec_from_file_location("diarization_calibration", SCRIPT)
|
||||
assert SPEC is not None and SPEC.loader is not None
|
||||
calibration = importlib.util.module_from_spec(SPEC)
|
||||
sys.modules[SPEC.name] = calibration
|
||||
SPEC.loader.exec_module(calibration)
|
||||
|
||||
|
||||
def _write(path, content):
|
||||
path.write_bytes(content)
|
||||
return {
|
||||
"path": str(path),
|
||||
"sha256": hashlib.sha256(content).hexdigest(),
|
||||
"size_bytes": len(content),
|
||||
}
|
||||
|
||||
|
||||
def _manifest(tmp_path):
|
||||
fp32 = _write(tmp_path / "fp32.onnx", b"fp32")
|
||||
int8 = _write(tmp_path / "int8.onnx", b"int8")
|
||||
wespeaker = _write(tmp_path / "wespeaker.onnx", b"wespeaker")
|
||||
titanet = _write(tmp_path / "titanet.onnx", b"titanet")
|
||||
segmentations = [
|
||||
{
|
||||
"id": "pyannote-fp32",
|
||||
"precision": "fp32",
|
||||
"source_url": "https://example/fp32",
|
||||
"license": "MIT",
|
||||
**fp32,
|
||||
},
|
||||
{
|
||||
"id": "pyannote-int8",
|
||||
"precision": "int8",
|
||||
"source_url": "https://example/int8",
|
||||
"license": "MIT",
|
||||
**int8,
|
||||
},
|
||||
]
|
||||
embeddings = [
|
||||
{
|
||||
"id": "wespeaker",
|
||||
"source_url": "https://example/wespeaker",
|
||||
"license": "CC-BY-4.0",
|
||||
**wespeaker,
|
||||
},
|
||||
{
|
||||
"id": "titanet-small",
|
||||
"source_url": "https://example/titanet",
|
||||
"license": "Apache-2.0",
|
||||
**titanet,
|
||||
},
|
||||
]
|
||||
recordings = []
|
||||
for index, (recording_id, speakers) in enumerate(
|
||||
(("data-test", 3), ("t2-bdma", 2), ("yantar", 2))
|
||||
):
|
||||
media = _write(tmp_path / f"{recording_id}.mp4", recording_id.encode())
|
||||
words_content = json.dumps(
|
||||
[{"start": 0.0, "end": 0.5, "text": " Слово"}]
|
||||
).encode()
|
||||
words = _write(tmp_path / f"{recording_id}.words.json", words_content)
|
||||
reference = None
|
||||
if index:
|
||||
reference_content = b"**[00:00] Speaker 1:**\n"
|
||||
ref = _write(tmp_path / f"{recording_id}.md", reference_content)
|
||||
reference = {
|
||||
"path": ref["path"],
|
||||
"sha256": ref["sha256"],
|
||||
"format": "hypescribe_markdown",
|
||||
}
|
||||
recordings.append(
|
||||
{
|
||||
"id": recording_id,
|
||||
"path": media["path"],
|
||||
"sha256": media["sha256"],
|
||||
"start": 0.0,
|
||||
"duration": 10.0,
|
||||
"expected_speakers": speakers,
|
||||
"reference": reference,
|
||||
"asr_words": {
|
||||
"path": words["path"],
|
||||
"sha256": words["sha256"],
|
||||
"format": "local_transcriber_words_v1",
|
||||
},
|
||||
}
|
||||
)
|
||||
return {
|
||||
"schema_version": 2,
|
||||
"segmentation_models": segmentations,
|
||||
"embedding_models": embeddings,
|
||||
"combinations": [
|
||||
{
|
||||
"id": "fp32-wespeaker",
|
||||
"segmentation_id": "pyannote-fp32",
|
||||
"embedding_id": "wespeaker",
|
||||
"role": "baseline",
|
||||
"thresholds": [0.8, 0.9],
|
||||
},
|
||||
{
|
||||
"id": "fp32-titanet",
|
||||
"segmentation_id": "pyannote-fp32",
|
||||
"embedding_id": "titanet-small",
|
||||
"role": "candidate",
|
||||
"thresholds": [0.8],
|
||||
},
|
||||
{
|
||||
"id": "int8-wespeaker",
|
||||
"segmentation_id": "pyannote-int8",
|
||||
"embedding_id": "wespeaker",
|
||||
"role": "candidate",
|
||||
"thresholds": [0.8],
|
||||
},
|
||||
{
|
||||
"id": "int8-titanet",
|
||||
"segmentation_id": "pyannote-int8",
|
||||
"embedding_id": "titanet-small",
|
||||
"role": "candidate",
|
||||
"thresholds": [0.8],
|
||||
},
|
||||
],
|
||||
"recordings": recordings,
|
||||
"threads": 4,
|
||||
"rss_sample_interval_ms": 5,
|
||||
}
|
||||
|
||||
|
||||
def _environment():
|
||||
return {
|
||||
"sherpa_onnx_version": "1.13.5",
|
||||
"onnxruntime_version": "1",
|
||||
"numpy_version": "2",
|
||||
}
|
||||
|
||||
|
||||
def _complete_cell(spec, clusters, purity=None, residual=0.0):
|
||||
return {
|
||||
**spec,
|
||||
"status": "complete",
|
||||
"diagnostics": {
|
||||
"clusters": clusters,
|
||||
"mapped_speaker_purity": purity,
|
||||
"residual_share_after_expected": residual,
|
||||
},
|
||||
}
|
||||
|
||||
|
||||
def test_manifest_validates_required_matrix_and_files(tmp_path):
|
||||
assert calibration.validate_manifest(_manifest(tmp_path))["schema_version"] == 2
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"mutation", ["missing_combination", "duplicate_id", "unsorted_threshold"]
|
||||
)
|
||||
def test_manifest_rejects_invalid_matrix_and_ids(tmp_path, mutation):
|
||||
manifest = _manifest(tmp_path)
|
||||
if mutation == "missing_combination":
|
||||
manifest["combinations"].pop()
|
||||
elif mutation == "duplicate_id":
|
||||
manifest["recordings"][1]["id"] = manifest["recordings"][0]["id"]
|
||||
else:
|
||||
manifest["combinations"][0]["thresholds"] = [0.9, 0.8]
|
||||
with pytest.raises(ValueError):
|
||||
calibration.validate_manifest(manifest, verify_files=False)
|
||||
|
||||
|
||||
def test_manifest_rejects_hash_mismatch_before_work(tmp_path):
|
||||
manifest = _manifest(tmp_path)
|
||||
manifest["recordings"][0]["sha256"] = "0" * 64
|
||||
with pytest.raises(ValueError, match="SHA-256"):
|
||||
calibration.validate_manifest(manifest)
|
||||
|
||||
|
||||
@pytest.mark.parametrize("model_kind", ["segmentation", "embedding"])
|
||||
def test_manifest_rejects_multiple_exact_model_ids_per_matrix_axis(
|
||||
tmp_path, model_kind
|
||||
):
|
||||
manifest = _manifest(tmp_path)
|
||||
if model_kind == "segmentation":
|
||||
duplicate = dict(manifest["segmentation_models"][0])
|
||||
duplicate["id"] = "pyannote-fp32-other"
|
||||
manifest["segmentation_models"].append(duplicate)
|
||||
manifest["combinations"][1]["segmentation_id"] = duplicate["id"]
|
||||
else:
|
||||
duplicate = dict(manifest["embedding_models"][0])
|
||||
duplicate["id"] = "wespeaker-other"
|
||||
manifest["embedding_models"].append(duplicate)
|
||||
manifest["combinations"][2]["embedding_id"] = duplicate["id"]
|
||||
|
||||
with pytest.raises(ValueError, match="ровно один"):
|
||||
calibration.validate_manifest(manifest, verify_files=False)
|
||||
|
||||
|
||||
def test_complete_run_rejects_substitute_corpus(tmp_path):
|
||||
manifest = _manifest(tmp_path)
|
||||
with pytest.raises(ValueError, match="Корпус"):
|
||||
calibration.validate_original_corpus(manifest)
|
||||
for recording in manifest["recordings"]:
|
||||
recording.update(calibration.ORIGINAL_CORPUS[recording["id"]])
|
||||
calibration.validate_original_corpus(manifest)
|
||||
|
||||
|
||||
def test_experiment_identity_ignores_paths_but_rejects_semantic_resume(tmp_path):
|
||||
manifest = _manifest(tmp_path)
|
||||
first = calibration.make_experiment_id(manifest, _environment())
|
||||
moved = json.loads(json.dumps(manifest))
|
||||
moved["recordings"][0]["path"] = "/different/private/path.mp4"
|
||||
assert calibration.make_experiment_id(moved, _environment()) == first
|
||||
output = tmp_path / "result.json"
|
||||
calibration.save_output(output, {"schema_version": 2, "experiment_id": "f" * 64})
|
||||
with pytest.raises(ValueError, match="другому семантическому"):
|
||||
calibration._load_or_create_output(output, manifest, _environment(), first)
|
||||
|
||||
|
||||
def test_resume_rejects_different_measurement_host_without_path_identity(tmp_path):
|
||||
manifest = _manifest(tmp_path)
|
||||
first_environment = {
|
||||
**_environment(),
|
||||
"started_at_utc": "2026-08-29T00:00:00+00:00",
|
||||
"platform": "host-a",
|
||||
"cpu": "cpu-a",
|
||||
"logical_cpus": 16,
|
||||
"python_version": "3.13",
|
||||
}
|
||||
experiment_id = calibration.make_experiment_id(manifest, first_environment)
|
||||
output_path = tmp_path / "result.json"
|
||||
calibration.save_output(
|
||||
output_path,
|
||||
calibration._new_output(manifest, first_environment, experiment_id),
|
||||
)
|
||||
moved = json.loads(json.dumps(manifest))
|
||||
moved["recordings"][0]["path"] = "/moved/private/recording.mp4"
|
||||
assert calibration.make_experiment_id(moved, first_environment) == experiment_id
|
||||
second_environment = {
|
||||
**first_environment,
|
||||
"started_at_utc": "2026-08-30T00:00:00+00:00",
|
||||
"platform": "host-b",
|
||||
"cpu": "cpu-b",
|
||||
"logical_cpus": 2,
|
||||
}
|
||||
|
||||
with pytest.raises(ValueError, match="другой platform/cpu/logical_cpus"):
|
||||
calibration._load_or_create_output(
|
||||
output_path, moved, second_environment, experiment_id
|
||||
)
|
||||
|
||||
|
||||
def test_cell_expansion_is_rectangular_deterministic_and_path_independent(tmp_path):
|
||||
manifest = _manifest(tmp_path)
|
||||
experiment_id = calibration.make_experiment_id(manifest, _environment())
|
||||
cells = calibration.expand_automatic_cells(manifest, experiment_id)
|
||||
assert len(cells) == 15
|
||||
assert [
|
||||
(cell["combination_id"], cell["recording_id"], cell["threshold"])
|
||||
for cell in cells[:4]
|
||||
] == [
|
||||
("fp32-wespeaker", "data-test", 0.8),
|
||||
("fp32-wespeaker", "t2-bdma", 0.8),
|
||||
("fp32-wespeaker", "yantar", 0.8),
|
||||
("fp32-wespeaker", "data-test", 0.9),
|
||||
]
|
||||
assert all(cell["num_clusters"] == -1 for cell in cells)
|
||||
moved = json.loads(json.dumps(manifest))
|
||||
for model in moved["segmentation_models"] + moved["embedding_models"]:
|
||||
model["path"] = f"/moved/models/{model['id']}.onnx"
|
||||
for recording in moved["recordings"]:
|
||||
recording["path"] = f"/moved/media/{recording['id']}.mp4"
|
||||
recording["asr_words"]["path"] = f"/moved/words/{recording['id']}.json"
|
||||
if recording["reference"]:
|
||||
recording["reference"]["path"] = f"/moved/refs/{recording['id']}.md"
|
||||
moved_experiment_id = calibration.make_experiment_id(moved, _environment())
|
||||
assert moved_experiment_id == experiment_id
|
||||
assert calibration.expand_automatic_cells(moved, moved_experiment_id) == cells
|
||||
|
||||
|
||||
def test_threshold_selection_requires_complete_rectangle_and_uses_shared_gate(tmp_path):
|
||||
manifest = _manifest(tmp_path)
|
||||
combination = manifest["combinations"][0]
|
||||
specs = calibration.expand_automatic_cells(manifest, "e" * 64)[:6]
|
||||
assert calibration.select_threshold(combination, manifest["recordings"], []) is None
|
||||
counts = [3, 2, 3, 3, 2, 2]
|
||||
purities = [None, 0.7, 0.9, None, 0.8, 0.8]
|
||||
cells = [
|
||||
_complete_cell(spec, count, purity)
|
||||
for spec, count, purity in zip(specs, counts, purities, strict=True)
|
||||
]
|
||||
selection = calibration.select_threshold(combination, manifest["recordings"], cells)
|
||||
assert selection["selected_threshold"] == 0.9
|
||||
assert selection["meets_count_gate"] is True
|
||||
known = calibration.derive_known_cells(manifest, "e" * 64, selection)
|
||||
assert [cell["num_clusters"] for cell in known] == [3, 2, 2]
|
||||
assert {cell["threshold"] for cell in known} == {0.9}
|
||||
|
||||
|
||||
def test_threshold_selection_records_failed_gate(tmp_path):
|
||||
manifest = _manifest(tmp_path)
|
||||
combination = manifest["combinations"][1]
|
||||
specs = [
|
||||
cell
|
||||
for cell in calibration.expand_automatic_cells(manifest, "e" * 64)
|
||||
if cell["combination_id"] == combination["id"]
|
||||
]
|
||||
cells = [
|
||||
_complete_cell(spec, count, purity)
|
||||
for spec, count, purity in zip(specs, [4, 3, 3], [None, 0.8, 0.8], strict=True)
|
||||
]
|
||||
assert (
|
||||
calibration.select_threshold(combination, manifest["recordings"], cells)[
|
||||
"meets_count_gate"
|
||||
]
|
||||
is False
|
||||
)
|
||||
|
||||
|
||||
def test_stage_timing_parser_accepts_native_lines():
|
||||
log = (
|
||||
"[I] OfflineSpeakerDiarization: segmentation 1.000 s\n"
|
||||
"[I] OfflineSpeakerDiarization: embedding 2.000 s\n"
|
||||
"[I] OfflineSpeakerDiarization: clustering 0.100 s\n"
|
||||
"[I] OfflineSpeakerDiarization: total 3.100 s, audio 10.000 s, RTF 0.310"
|
||||
)
|
||||
stages = calibration.parse_stage_timings(log, "1.13.5")
|
||||
assert stages["engine_total_seconds"] == 3.1
|
||||
assert stages["engine_audio_seconds"] == 10.0
|
||||
assert stages["engine_rtf"] == 0.31
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"log",
|
||||
[
|
||||
"OfflineSpeakerDiarization: segmentation 1.0 s",
|
||||
"OfflineSpeakerDiarization: segmentation 1.0 s\nOfflineSpeakerDiarization: segmentation 1.0 s\nOfflineSpeakerDiarization: embedding 1.0 s\nOfflineSpeakerDiarization: clustering 1.0 s\nOfflineSpeakerDiarization: total 3.0 s, audio 4 s, RTF .75",
|
||||
"OfflineSpeakerDiarization: segmentation -1.0 s\nOfflineSpeakerDiarization: embedding 1.0 s\nOfflineSpeakerDiarization: clustering 1.0 s\nOfflineSpeakerDiarization: total 1.0 s, audio 4 s, RTF .25",
|
||||
],
|
||||
)
|
||||
def test_stage_timing_parser_rejects_missing_duplicate_and_negative(log):
|
||||
with pytest.raises(ValueError):
|
||||
calibration.parse_stage_timings(log, "1.13.5")
|
||||
|
||||
|
||||
def test_stage_timing_parser_rejects_extra_malformed_record():
|
||||
log = (
|
||||
"OfflineSpeakerDiarization: segmentation ??? s\n"
|
||||
"OfflineSpeakerDiarization: segmentation 1.000 s\n"
|
||||
"OfflineSpeakerDiarization: embedding 2.000 s\n"
|
||||
"OfflineSpeakerDiarization: clustering 0.100 s\n"
|
||||
"OfflineSpeakerDiarization: total 3.100 s, audio 10.000 s, RTF 0.310"
|
||||
)
|
||||
with pytest.raises(ValueError, match="повреждённую"):
|
||||
calibration.parse_stage_timings(log, "1.13.5")
|
||||
|
||||
|
||||
def test_measurement_uses_injected_cpu_clock_and_rss_once():
|
||||
process = SimpleNamespace()
|
||||
calls = []
|
||||
cpu_times = iter(
|
||||
[SimpleNamespace(user=1.0, system=2.0), SimpleNamespace(user=2.0, system=2.5)]
|
||||
)
|
||||
|
||||
def read_cpu_times():
|
||||
calls.append("cpu")
|
||||
return next(cpu_times)
|
||||
|
||||
process.cpu_times = read_cpu_times
|
||||
|
||||
class Sampler:
|
||||
def __init__(self, actual_process, interval):
|
||||
assert actual_process is process
|
||||
assert interval == 0.01
|
||||
|
||||
def start(self):
|
||||
calls.append("start")
|
||||
|
||||
def stop(self):
|
||||
calls.append("stop")
|
||||
return 1234
|
||||
|
||||
clock = iter([10.0, 12.0]).__next__
|
||||
result, metrics = calibration.measure_call(
|
||||
lambda: calls.append("process") or "ok",
|
||||
process,
|
||||
clock=clock,
|
||||
sampler_factory=Sampler,
|
||||
interval_seconds=0.01,
|
||||
logical_cpus=4,
|
||||
)
|
||||
assert result == "ok"
|
||||
assert calls == ["start", "cpu", "process", "cpu", "stop"]
|
||||
assert metrics["cpu_total_seconds"] == 1.5
|
||||
assert metrics["average_cpu_percent_machine"] == 18.75
|
||||
assert metrics["peak_rss_bytes"] == 1234
|
||||
|
||||
|
||||
def test_measurement_preserves_operation_error_when_sampler_stop_also_fails():
|
||||
process = SimpleNamespace(
|
||||
cpu_times=iter(
|
||||
[
|
||||
SimpleNamespace(user=1.0, system=2.0),
|
||||
SimpleNamespace(user=1.1, system=2.1),
|
||||
]
|
||||
).__next__
|
||||
)
|
||||
|
||||
class Sampler:
|
||||
def __init__(self, process, interval):
|
||||
pass
|
||||
|
||||
def start(self):
|
||||
pass
|
||||
|
||||
def stop(self):
|
||||
raise RuntimeError("sampler failed")
|
||||
|
||||
def operation():
|
||||
raise ValueError("operation failed")
|
||||
|
||||
with pytest.raises(ValueError, match="operation failed") as error:
|
||||
calibration.measure_call(
|
||||
operation,
|
||||
process,
|
||||
clock=iter([1.0, 2.0]).__next__,
|
||||
sampler_factory=Sampler,
|
||||
interval_seconds=0.01,
|
||||
logical_cpus=4,
|
||||
)
|
||||
assert any("sampler failed" in note for note in error.value.__notes__)
|
||||
|
||||
|
||||
def test_decode_cache_identity_covers_hash_bounds_and_decoder():
|
||||
recording = {"sha256": "a" * 64, "start": 1.0, "duration": 2.0}
|
||||
original = calibration.decode_cache_key(recording)
|
||||
assert calibration.decode_cache_key({**recording, "start": 2.0}) != original
|
||||
assert calibration.decode_cache_key({**recording, "sha256": "b" * 64}) != original
|
||||
assert calibration.decode_cache_key(recording, "another-decoder") != original
|
||||
|
||||
|
||||
def test_decode_clip_declares_wav_format_for_temporary_output(tmp_path, monkeypatch):
|
||||
recording = {
|
||||
"id": "synthetic",
|
||||
"path": str(tmp_path / "source.mp4"),
|
||||
"sha256": "a" * 64,
|
||||
"start": 1.0,
|
||||
"duration": 2.0,
|
||||
}
|
||||
commands = []
|
||||
|
||||
def run(command, *, check):
|
||||
assert check is True
|
||||
commands.append(command)
|
||||
Path(command[-1]).touch()
|
||||
|
||||
monkeypatch.setattr(calibration.subprocess, "run", run)
|
||||
monkeypatch.setattr(calibration, "read_wav", lambda path: None)
|
||||
|
||||
output = calibration.decode_clip(recording, tmp_path / "work")
|
||||
|
||||
assert output.is_file()
|
||||
assert not output.with_suffix(".wav.tmp").exists()
|
||||
assert commands[0][-3:-1] == ["-f", "wav"]
|
||||
|
||||
|
||||
def test_hypescribe_parser_mapping_and_empty_diagnostics(tmp_path):
|
||||
reference = tmp_path / "reference.md"
|
||||
reference.write_text(
|
||||
"**[00:10] Speaker 1:**\ntext\n**[00:15] Speaker 2:**\n", encoding="utf-8"
|
||||
)
|
||||
recording = {
|
||||
"id": "clip",
|
||||
"path": str(tmp_path / "clip.mp4"),
|
||||
"sha256": "a" * 64,
|
||||
"start": 12.0,
|
||||
"duration": 6.0,
|
||||
"expected_speakers": 2,
|
||||
"reference": {"path": str(reference)},
|
||||
}
|
||||
turns = calibration.read_reference_turns(recording)
|
||||
assert turns == [
|
||||
{"speaker": "1", "start": 0.0, "end": 3.0},
|
||||
{"speaker": "2", "start": 3.0, "end": 6.0},
|
||||
]
|
||||
mapping = calibration.best_mapping(
|
||||
[
|
||||
{"speaker": 7, "start": 0.0, "end": 3.0},
|
||||
{"speaker": 8, "start": 3.0, "end": 6.0},
|
||||
],
|
||||
turns,
|
||||
)
|
||||
assert mapping["mapped_speaker_purity"] == 1.0
|
||||
assert calibration.best_mapping([], turns) is None
|
||||
summary = calibration.summarize_segments([], recording)
|
||||
assert summary["clusters"] == 0 and summary["residual_share_after_expected"] is None
|
||||
|
||||
|
||||
def test_asr_sidecar_validation_text_invariance_and_unknown_words(tmp_path):
|
||||
sidecar = tmp_path / "words.json"
|
||||
sidecar.write_text(
|
||||
json.dumps(
|
||||
[
|
||||
{"start": 0.0, "end": 0.5, "text": "Привет"},
|
||||
{"start": 0.5, "end": 0.6, "text": "!"},
|
||||
]
|
||||
),
|
||||
encoding="utf-8",
|
||||
)
|
||||
words = calibration.load_asr_words(sidecar, 2.0)
|
||||
recording = {"duration": 2.0, "expected_speakers": 1}
|
||||
diagnostics, invariant = calibration.evaluate_segments(
|
||||
[{"speaker": 0, "start": 0.0, "end": 0.5}], recording, words, []
|
||||
)
|
||||
assert diagnostics["unassigned_word_count"] == 1
|
||||
assert diagnostics["mapped_speaker_purity"] is None
|
||||
assert invariant["equal"] is True and invariant["first_mismatch_index"] is None
|
||||
sidecar.write_text(
|
||||
json.dumps(
|
||||
[
|
||||
{"start": 1.0, "end": 1.1, "text": "a"},
|
||||
{"start": 0.5, "end": 0.6, "text": "b"},
|
||||
]
|
||||
),
|
||||
encoding="utf-8",
|
||||
)
|
||||
with pytest.raises(ValueError, match="границы"):
|
||||
calibration.load_asr_words(sidecar, 2.0)
|
||||
|
||||
|
||||
def test_asr_sidecar_rejects_empty_baseline(tmp_path):
|
||||
sidecar = tmp_path / "words.json"
|
||||
sidecar.write_text("[]", encoding="utf-8")
|
||||
|
||||
with pytest.raises(ValueError, match="хотя бы одно слово"):
|
||||
calibration.load_asr_words(sidecar, 300.0)
|
||||
|
||||
|
||||
def test_text_invariance_records_first_mismatch_without_text():
|
||||
transcript = SimpleNamespace(turns=[SimpleNamespace(text="другой")])
|
||||
words = [SimpleNamespace(text="текст")]
|
||||
evidence = calibration.text_invariance(words, transcript)
|
||||
assert evidence["equal"] is False
|
||||
assert evidence["first_mismatch_index"] == 0
|
||||
assert "текст" not in json.dumps(evidence, ensure_ascii=False)
|
||||
|
||||
|
||||
def test_text_invariance_ignores_unicode_whitespace_at_turn_boundaries():
|
||||
words = [
|
||||
SimpleNamespace(text="Вопрос?"),
|
||||
SimpleNamespace(text=" — "),
|
||||
SimpleNamespace(text="ответ"),
|
||||
]
|
||||
transcript = SimpleNamespace(
|
||||
turns=[SimpleNamespace(text="Вопрос?"), SimpleNamespace(text="— ответ")]
|
||||
)
|
||||
|
||||
evidence = calibration.text_invariance(words, transcript)
|
||||
|
||||
assert calibration.canonical_non_whitespace_sequence(words) == "Вопрос?—ответ"
|
||||
assert evidence["equal"] is True
|
||||
assert evidence["first_mismatch_index"] is None
|
||||
|
||||
|
||||
@pytest.mark.parametrize("output", ["Вопрос?ответ", "Вопрос!—ответ"])
|
||||
def test_text_invariance_rejects_non_whitespace_deletion_or_substitution(output):
|
||||
words = [SimpleNamespace(text="Вопрос?"), SimpleNamespace(text="— ответ")]
|
||||
transcript = SimpleNamespace(turns=[SimpleNamespace(text=output)])
|
||||
|
||||
evidence = calibration.text_invariance(words, transcript)
|
||||
|
||||
assert evidence["equal"] is False
|
||||
assert evidence["input_sha256"] != evidence["output_sha256"]
|
||||
assert evidence["first_mismatch_index"] is not None
|
||||
|
||||
|
||||
def test_atomic_persistence_leaves_no_temporary_file(tmp_path):
|
||||
output = tmp_path / "nested" / "result.json"
|
||||
calibration.save_output(output, {"private": "value"})
|
||||
assert json.loads(output.read_text()) == {"private": "value"}
|
||||
assert list(output.parent.glob("*.tmp")) == []
|
||||
|
||||
|
||||
def test_atomic_persistence_preserves_target_and_cleans_temp_on_dump_failure(tmp_path):
|
||||
output = tmp_path / "result.json"
|
||||
calibration.save_output(output, {"stable": True})
|
||||
|
||||
with pytest.raises(TypeError):
|
||||
calibration.save_output(output, {"unserializable": object()})
|
||||
|
||||
assert json.loads(output.read_text(encoding="utf-8")) == {"stable": True}
|
||||
assert list(output.parent.glob("*.tmp")) == []
|
||||
|
||||
|
||||
def test_worker_failure_is_persisted_and_retried(tmp_path, monkeypatch):
|
||||
output = {"cells": []}
|
||||
output_path = tmp_path / "result.json"
|
||||
spec = {
|
||||
"cell_id": "cell",
|
||||
"combination_id": "c",
|
||||
"recording_id": "r",
|
||||
"mode": "automatic",
|
||||
"threshold": 0.8,
|
||||
"num_clusters": -1,
|
||||
}
|
||||
attempts = []
|
||||
|
||||
def execute(*args, **kwargs):
|
||||
attempts.append(1)
|
||||
if len(attempts) == 1:
|
||||
raise RuntimeError("boom")
|
||||
return {**spec, "status": "complete"}
|
||||
|
||||
monkeypatch.setattr(calibration, "_execute_cell", execute)
|
||||
calibration._run_specs([spec], {}, output, {}, output_path, tmp_path)
|
||||
assert output["cells"][0]["status"] == "failed"
|
||||
assert output["cells"][0]["metrics"] is None
|
||||
calibration._run_specs([spec], {}, output, {}, output_path, tmp_path)
|
||||
assert output["cells"][0]["status"] == "complete"
|
||||
assert len(attempts) == 2
|
||||
|
||||
|
||||
def test_execute_cell_rejects_text_mismatch_without_persisting_private_text(
|
||||
tmp_path, monkeypatch
|
||||
):
|
||||
manifest = _manifest(tmp_path)
|
||||
output = {"environment": {"logical_cpus": 4, "sherpa_onnx_version": "1.13.5"}}
|
||||
spec = calibration.expand_automatic_cells(manifest, "e" * 64)[0]
|
||||
recording = manifest["recordings"][0]
|
||||
private_text = " СОВЕРШЕННО-СЕКРЕТНЫЙ-ТЕКСТ"
|
||||
Path(recording["asr_words"]["path"]).write_text(
|
||||
json.dumps([{"start": 0.0, "end": 0.5, "text": private_text}]),
|
||||
encoding="utf-8",
|
||||
)
|
||||
monkeypatch.setattr(
|
||||
calibration,
|
||||
"run_worker",
|
||||
lambda *args, **kwargs: (
|
||||
{"metrics": {"wall_seconds": 1.0}, "segments": []},
|
||||
(
|
||||
"OfflineSpeakerDiarization: segmentation 0.300 s\n"
|
||||
"OfflineSpeakerDiarization: embedding 0.600 s\n"
|
||||
"OfflineSpeakerDiarization: clustering 0.100 s\n"
|
||||
"OfflineSpeakerDiarization: total 1.000 s, "
|
||||
"audio 10.000 s, RTF 0.100"
|
||||
),
|
||||
),
|
||||
)
|
||||
monkeypatch.setattr(calibration, "read_reference_turns", lambda *args: [])
|
||||
|
||||
def evaluate(segments, recording, words, reference_turns):
|
||||
source = calibration.canonical_non_whitespace_sequence(words)
|
||||
assert private_text.strip() in source
|
||||
output = source[:-1]
|
||||
return {}, {
|
||||
"input_sha256": hashlib.sha256(source.encode()).hexdigest(),
|
||||
"output_sha256": hashlib.sha256(output.encode()).hexdigest(),
|
||||
"equal": False,
|
||||
"first_mismatch_index": len(output),
|
||||
}
|
||||
|
||||
monkeypatch.setattr(calibration, "evaluate_segments", evaluate)
|
||||
|
||||
with pytest.raises(RuntimeError, match="first_mismatch_index=") as error:
|
||||
calibration._execute_cell(
|
||||
spec,
|
||||
manifest,
|
||||
output,
|
||||
{recording["id"]: tmp_path / "decoded.wav"},
|
||||
tmp_path,
|
||||
)
|
||||
assert private_text.strip() not in str(error.value)
|
||||
|
||||
|
||||
def test_execute_cell_rejects_engine_audio_duration_mismatch(tmp_path, monkeypatch):
|
||||
manifest = _manifest(tmp_path)
|
||||
output = {"environment": {"logical_cpus": 4, "sherpa_onnx_version": "1.13.5"}}
|
||||
spec = calibration.expand_automatic_cells(manifest, "e" * 64)[0]
|
||||
recording = manifest["recordings"][0]
|
||||
monkeypatch.setattr(
|
||||
calibration,
|
||||
"run_worker",
|
||||
lambda *args, **kwargs: (
|
||||
{"metrics": {"wall_seconds": 1.0}, "segments": []},
|
||||
(
|
||||
"OfflineSpeakerDiarization: segmentation 0.300 s\n"
|
||||
"OfflineSpeakerDiarization: embedding 0.600 s\n"
|
||||
"OfflineSpeakerDiarization: clustering 0.100 s\n"
|
||||
"OfflineSpeakerDiarization: total 1.000 s, "
|
||||
"audio 9.000 s, RTF 0.111"
|
||||
),
|
||||
),
|
||||
)
|
||||
|
||||
with pytest.raises(RuntimeError, match="Длительность audio"):
|
||||
calibration._execute_cell(
|
||||
spec,
|
||||
manifest,
|
||||
output,
|
||||
{recording["id"]: tmp_path / "decoded.wav"},
|
||||
tmp_path,
|
||||
)
|
||||
|
||||
|
||||
def test_main_fails_closed_when_a_cell_failed(tmp_path, monkeypatch):
|
||||
manifest = _manifest(tmp_path)
|
||||
for recording in manifest["recordings"]:
|
||||
recording.update(calibration.ORIGINAL_CORPUS[recording["id"]])
|
||||
manifest_path = tmp_path / "manifest.json"
|
||||
manifest_path.write_text(json.dumps(manifest), encoding="utf-8")
|
||||
output_path = tmp_path / "result.json"
|
||||
work_dir = tmp_path / "work"
|
||||
monkeypatch.setattr(calibration, "validate_manifest", lambda *args, **kwargs: None)
|
||||
monkeypatch.setattr(
|
||||
calibration,
|
||||
"environment_snapshot",
|
||||
lambda: {
|
||||
**_environment(),
|
||||
"started_at_utc": "2026-08-29T00:00:00+00:00",
|
||||
"platform": "test",
|
||||
"cpu": "test",
|
||||
"logical_cpus": 4,
|
||||
"python_version": "3.13",
|
||||
},
|
||||
)
|
||||
monkeypatch.setattr(
|
||||
calibration,
|
||||
"decode_clip",
|
||||
lambda recording, target: tmp_path / f"{recording['id']}.wav",
|
||||
)
|
||||
|
||||
def run_specs(specs, manifest, output, decoded, output_path, work_dir):
|
||||
for spec in specs:
|
||||
calibration._replace_cell(
|
||||
output, calibration._failed_cell(spec, RuntimeError("boom"))
|
||||
)
|
||||
calibration.save_output(output_path, output)
|
||||
|
||||
monkeypatch.setattr(calibration, "_run_specs", run_specs)
|
||||
|
||||
with pytest.raises(RuntimeError, match="незавершённых ячеек 15"):
|
||||
calibration.main(
|
||||
[
|
||||
"--manifest",
|
||||
str(manifest_path),
|
||||
"--output",
|
||||
str(output_path),
|
||||
"--work-dir",
|
||||
str(work_dir),
|
||||
]
|
||||
)
|
||||
|
||||
|
||||
def test_main_fails_closed_for_complete_rectangle_without_mapped_purity(
|
||||
tmp_path, monkeypatch
|
||||
):
|
||||
manifest = _manifest(tmp_path)
|
||||
for recording in manifest["recordings"]:
|
||||
recording.update(calibration.ORIGINAL_CORPUS[recording["id"]])
|
||||
manifest_path = tmp_path / "manifest.json"
|
||||
manifest_path.write_text(json.dumps(manifest), encoding="utf-8")
|
||||
output_path = tmp_path / "result.json"
|
||||
work_dir = tmp_path / "work"
|
||||
monkeypatch.setattr(calibration, "validate_manifest", lambda *args, **kwargs: None)
|
||||
monkeypatch.setattr(
|
||||
calibration,
|
||||
"environment_snapshot",
|
||||
lambda: {
|
||||
**_environment(),
|
||||
"started_at_utc": "2026-08-29T00:00:00+00:00",
|
||||
"platform": "test",
|
||||
"cpu": "test",
|
||||
"logical_cpus": 4,
|
||||
"python_version": "3.13",
|
||||
},
|
||||
)
|
||||
monkeypatch.setattr(
|
||||
calibration,
|
||||
"decode_clip",
|
||||
lambda recording, target: tmp_path / f"{recording['id']}.wav",
|
||||
)
|
||||
modes = []
|
||||
|
||||
def run_specs(specs, manifest, output, decoded, output_path, work_dir):
|
||||
modes.extend({spec["mode"] for spec in specs})
|
||||
expected = {
|
||||
recording["id"]: recording["expected_speakers"]
|
||||
for recording in manifest["recordings"]
|
||||
}
|
||||
for spec in specs:
|
||||
calibration._replace_cell(
|
||||
output,
|
||||
_complete_cell(spec, expected[spec["recording_id"]], purity=None),
|
||||
)
|
||||
calibration.save_output(output_path, output)
|
||||
|
||||
monkeypatch.setattr(calibration, "_run_specs", run_specs)
|
||||
|
||||
with pytest.raises(RuntimeError, match="свип непригоден"):
|
||||
calibration.main(
|
||||
[
|
||||
"--manifest",
|
||||
str(manifest_path),
|
||||
"--output",
|
||||
str(output_path),
|
||||
"--work-dir",
|
||||
str(work_dir),
|
||||
]
|
||||
)
|
||||
|
||||
output = json.loads(output_path.read_text(encoding="utf-8"))
|
||||
assert modes == ["automatic"]
|
||||
assert {cell["mode"] for cell in output["cells"]} == {"automatic"}
|
||||
assert output["threshold_selection"] == []
|
||||
|
||||
|
||||
def test_run_worker_fake_protocol_reports_failure_and_success(tmp_path):
|
||||
request = {"cell_id": "abc"}
|
||||
commands = []
|
||||
|
||||
def failed(command, **kwargs):
|
||||
commands.append(command)
|
||||
return SimpleNamespace(returncode=1, stdout="", stderr="worker failed\n")
|
||||
|
||||
with pytest.raises(RuntimeError, match="worker failed"):
|
||||
calibration.run_worker(request, tmp_path, runner=failed)
|
||||
|
||||
def success(command, **kwargs):
|
||||
commands.append(command)
|
||||
return SimpleNamespace(
|
||||
returncode=0,
|
||||
stdout='{"metrics":{},"segments":[]}',
|
||||
stderr="native log",
|
||||
)
|
||||
|
||||
payload, stderr = calibration.run_worker(request, tmp_path, runner=success)
|
||||
assert payload["segments"] == [] and stderr == "native log"
|
||||
for command in commands:
|
||||
assert command[:3] == [
|
||||
sys.executable,
|
||||
str(SCRIPT.resolve()),
|
||||
"--worker",
|
||||
]
|
||||
request_path = Path(command[3])
|
||||
assert request_path.is_file()
|
||||
assert json.loads(request_path.read_text(encoding="utf-8")) == request
|
||||
Reference in new issue
Block a user