300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558 | def run_benchmark_manifest(
manifest_csv: Path,
*,
out_dir: Path,
aggregate_out: Path | None = None,
provenance_out: Path | None = None,
plot_out: Path | None = None,
chance: float | None = None,
default_label_column: str | None = None,
default_group_column: str | None = None,
default_picks: str = "data",
default_tmin: float | None = None,
default_tmax: float | None = None,
default_window_ms: float = 20.0,
default_step_ms: float = 10.0,
default_n_splits: int = 5,
default_max_iter: int = 1000,
default_decoder: str = "logistic",
default_emission_mode: str = "calibrated",
default_feature_preprocessor: str = "none",
default_pca_components: str | None = None,
default_normalization: str = "none",
default_baseline_window: tuple[float, float] | None = DEFAULT_DECODE_BASELINE_WINDOW,
default_tune_hyperparameters: bool = False,
default_tuning_cv_splits: int = 3,
default_tuning_scoring: str = "accuracy",
default_tuning_c_grid: str | None = None,
default_temporal_train_window: TemporalTrainWindow | None = None,
calibration_dir: Path | None = None,
calibration_bins: int = 10,
observation_dir: Path | None = None,
observation_ensemble_dir: Path | None = None,
observation_ensemble_source_decoders: tuple[str, ...] = DEFAULT_ENSEMBLE_SOURCE_DECODERS,
observation_ensemble_weights: tuple[float, ...] | None = None,
observation_ensemble_source_emission_mode: str | None = "calibrated",
observation_ensemble_baseline_window: tuple[float, float] | None = DEFAULT_ENSEMBLE_BASELINE_WINDOW,
observation_ensemble_baseline_group_columns: tuple[str, ...] = DEFAULT_ENSEMBLE_BASELINE_GROUP_COLUMNS,
temporal_smoothing_dir: Path | None = None,
temporal_smoothing_fit_window: tuple[float, float] | None = DEFAULT_FIT_WINDOW,
temporal_smoothing_stay_grid_size: int = 200,
temporal_smoothing_emission_suffix: str = DEFAULT_EMISSION_SUFFIX,
resume: bool = False,
) -> BenchmarkRun:
"""Run a manifest-defined benchmark and optionally aggregate and plot results."""
manifest = pd.read_csv(manifest_csv)
if "subject" not in manifest.columns or "epochs" not in manifest.columns:
raise ValueError("Manifest must contain 'subject' and 'epochs' columns.")
manifest_dir = manifest_csv.parent
out_dir.mkdir(parents=True, exist_ok=True)
if temporal_smoothing_dir is not None and observation_dir is None:
observation_dir = out_dir / "observations"
if observation_ensemble_dir is not None and observation_dir is None:
observation_dir = out_dir / "observations"
result_csvs: list[Path] = []
calibration_csvs: list[Path] = []
observation_csvs: list[Path] = []
skipped_existing = 0
for _, row in manifest.iterrows():
subject = _string_value(row, "subject")
if subject is None:
raise ValueError("Manifest contains a row with an empty subject.")
label_column = _string_value(row, "label_column", default_label_column)
if label_column is None:
raise ValueError(f"Subject '{subject}' has no label column.")
decoder = normalize_decoder_name(_string_value(row, "decoder", default_decoder) or default_decoder)
emission_mode = _string_value(row, "emission_mode", default_emission_mode) or default_emission_mode
if emission_mode != "both":
emission_mode = normalize_emission_mode(emission_mode)
feature_preprocessor = normalize_feature_preprocessor(
_string_value(row, "feature_preprocessor", default_feature_preprocessor) or default_feature_preprocessor
)
pca_components = _string_value(row, "pca_components", default_pca_components)
tune_hyperparameters = _bool_value(row, "tune_hyperparameters", default_tune_hyperparameters)
normalization = normalize_epoch_normalization(_string_value(row, "normalization", default_normalization) or default_normalization)
baseline_window = _baseline_window_value(row, default_baseline_window)
tuning_cv_splits = _int_value(row, "tuning_cv_splits", default_tuning_cv_splits)
tuning_scoring = normalize_tuning_scoring(
_string_value(row, "tuning_scoring", default_tuning_scoring) or default_tuning_scoring
)
tuning_c_grid = _string_value(row, "tuning_c_grid", default_tuning_c_grid)
temporal_train_window = _temporal_train_window_value(row, default_temporal_train_window)
output_stem = _output_stem(
subject,
decoder,
emission_mode,
has_decoder_column="decoder" in manifest.columns,
has_emission_mode_column="emission_mode" in manifest.columns,
variant=_string_value(row, "variant"),
feature_preprocessor=feature_preprocessor,
pca_components=pca_components,
tune_hyperparameters=tune_hyperparameters,
tuning_scoring=tuning_scoring,
temporal_train_window=temporal_train_window,
normalization=normalization,
baseline_window=baseline_window,
has_feature_preprocessor_column="feature_preprocessor" in manifest.columns,
has_pca_components_column="pca_components" in manifest.columns,
has_tune_hyperparameters_column="tune_hyperparameters" in manifest.columns,
has_tuning_scoring_column="tuning_scoring" in manifest.columns,
has_normalization_column="normalization" in manifest.columns,
has_baseline_window_column=bool(
{"baseline_window", "baseline_window_start", "baseline_window_stop"}.intersection(
manifest.columns
)
),
has_temporal_train_window_column=bool(
{"temporal_train_window", "temporal_train_window_start", "temporal_train_window_stop"}.intersection(
manifest.columns
)
),
)
output_csv = _resolve_path(_string_value(row, "out_csv"), manifest_dir)
if output_csv is None:
output_csv = out_dir / f"{output_stem}_time_decode.csv"
calibration_out_csv = _resolve_path(_string_value(row, "calibration_out_csv"), manifest_dir)
if calibration_out_csv is None and calibration_dir is not None:
calibration_out_csv = calibration_dir / f"{output_stem}_calibration_bins.csv"
observation_out_csv = _resolve_path(_string_value(row, "observation_out_csv"), manifest_dir)
if observation_out_csv is None and observation_dir is not None:
observation_out_csv = observation_dir / f"{output_stem}_observations.csv"
if (
resume
and _usable_file(output_csv)
and (calibration_out_csv is None or _usable_file(calibration_out_csv))
and (observation_out_csv is None or _usable_file(observation_out_csv))
):
result_csvs.append(output_csv)
if calibration_out_csv is not None:
calibration_csvs.append(calibration_out_csv)
if observation_out_csv is not None:
observation_csvs.append(observation_out_csv)
skipped_existing += 1
continue
metadata_csv = _prepare_or_resolve_metadata(row, manifest_dir, out_dir, subject)
results = run_time_resolved_decode(
epochs_path=_required_path(row, "epochs", manifest_dir),
metadata_csv=metadata_csv,
label_column=label_column,
group_column=_string_value(row, "group_column", default_group_column),
out_path=output_csv,
picks=_string_value(row, "picks", default_picks) or default_picks,
tmin=_float_value(row, "tmin", default_tmin),
tmax=_float_value(row, "tmax", default_tmax),
window_ms=_float_value(row, "window_ms", default_window_ms) or default_window_ms,
step_ms=_float_value(row, "step_ms", default_step_ms) or default_step_ms,
n_splits=_int_value(row, "n_splits", default_n_splits),
max_iter=_int_value(row, "max_iter", default_max_iter),
decoder=decoder,
emission_mode=emission_mode,
feature_preprocessor=feature_preprocessor,
pca_components=pca_components,
normalization=normalization,
baseline_window=baseline_window,
tune_hyperparameters=tune_hyperparameters,
tuning_cv_splits=tuning_cv_splits,
tuning_scoring=tuning_scoring,
tuning_c_grid=tuning_c_grid,
temporal_train_window=temporal_train_window,
calibration_out_path=calibration_out_csv,
calibration_bins=_int_value(row, "calibration_bins", calibration_bins),
observation_out_path=observation_out_csv,
subject=subject,
)
if calibration_out_csv is not None:
calibration_csvs.append(calibration_out_csv)
if observation_out_csv is not None:
observation_csvs.append(observation_out_csv)
if "subject" not in results.columns:
results.insert(0, "subject", subject)
else:
results["subject"] = subject
output_csv.parent.mkdir(parents=True, exist_ok=True)
results.to_csv(output_csv, index=False)
result_csvs.append(output_csv)
aggregate_result_csvs = list(result_csvs)
aggregate_observation_csvs = list(observation_csvs)
ensemble_observation_csv: Path | None = None
ensemble_metric_csv: Path | None = None
if observation_ensemble_dir is not None:
ensemble_observation_csv, ensemble_metric_csv = _write_observation_ensemble(
observation_csvs,
out_dir=observation_ensemble_dir,
resume=resume,
source_decoders=tuple(normalize_decoder_name(decoder) for decoder in observation_ensemble_source_decoders),
weights=observation_ensemble_weights,
source_emission_mode=observation_ensemble_source_emission_mode,
baseline_window=observation_ensemble_baseline_window,
baseline_group_columns=observation_ensemble_baseline_group_columns,
calibration_bins=calibration_bins,
)
aggregate_result_csvs.append(ensemble_metric_csv)
aggregate_observation_csvs.append(ensemble_observation_csv)
smoothed_observation_csv: Path | None = None
smoothed_metric_csv: Path | None = None
if temporal_smoothing_dir is not None:
if not observation_csvs:
raise ValueError("Temporal smoothing requires probability observations; pass --observation-dir.")
smoothed_observation_csv = temporal_smoothing_dir / "smoothed_observations.csv"
smoothed_metric_csv = temporal_smoothing_dir / "smoothed_metrics.csv"
if not (resume and _usable_file(smoothed_observation_csv) and _usable_file(smoothed_metric_csv)):
smooth_probability_observations(
observation_csvs,
fit_window=temporal_smoothing_fit_window,
stay_grid_size=temporal_smoothing_stay_grid_size,
emission_suffix=temporal_smoothing_emission_suffix,
ece_bins=calibration_bins,
out_observations=smoothed_observation_csv,
out_metrics=smoothed_metric_csv,
)
aggregate_result_csvs.append(smoothed_metric_csv)
aggregate_observation_csvs.append(smoothed_observation_csv)
if aggregate_out is None:
aggregate_out = out_dir / "summary.csv"
aggregate = aggregate_time_decode_csvs(
aggregate_result_csvs,
out_path=aggregate_out,
observation_csv_paths=aggregate_observation_csvs or None,
)
aggregate_path: Path | None = aggregate_out
if provenance_out is None:
provenance_out = out_dir / "provenance.csv"
provenance = write_provenance_table(
aggregate,
aggregate_result_csvs,
provenance_out,
)
provenance_path: Path | None = provenance_out
plot_path: Path | None = None
if plot_out is not None:
plot_time_decode_results(
aggregate_out,
out_path=plot_out,
chance=chance,
title=f"NeuRepTrace benchmark ({int(provenance['n_subjects'].max())} subject(s))",
)
plot_path = plot_out
return BenchmarkRun(
result_csvs=result_csvs,
aggregate_csv=aggregate_path,
provenance_csv=provenance_path,
plot_path=plot_path,
calibration_csvs=calibration_csvs,
observation_csvs=observation_csvs,
skipped_existing=skipped_existing,
smoothed_observation_csv=smoothed_observation_csv,
smoothed_metric_csv=smoothed_metric_csv,
ensemble_observation_csv=ensemble_observation_csv,
ensemble_metric_csv=ensemble_metric_csv,
)
|