Coverage for src / bioimageio / core / proc_ops.py: 79%

1import collections.abc 

2from dataclasses import InitVar, dataclass, field 

3from functools import partial 

4from typing import ( 

5 Any, 

6 Callable, 

7 Collection, 

8 Literal, 

9 Mapping, 

10 Optional, 

11 Sequence, 

12 Set, 

13 Tuple, 

14 Union, 

15) 

16 

17import numpy as np 

18import scipy 

19import xarray as xr 

20from numpy.typing import NDArray 

21from typing_extensions import Self, assert_never 

22 

23from bioimageio.spec.model import v0_4, v0_5 

24from bioimageio.spec.model.v0_5 import ( 

25 _convert_proc, # pyright: ignore[reportPrivateUsage] 

26) 

27 

28from ._op_base import BlockwiseOperator, SamplewiseOperator, SimpleOperator 

29from ._ops_cellpose import CellposeFlowDynamics 

30from ._ops_stardist import StardistPostprocessing2D as StardistPostprocessing2D 

31from ._ops_stardist import StardistPostprocessing3D as StardistPostprocessing3D 

32from .axis import AxisId 

33from .common import DTypeStr, MemberId 

34from .digest_spec import get_member_id, import_callable 

35from .sample import Sample, SampleBlock 

36from .stat_calculators import StatsCalculator 

37from .stat_measures import ( 

38 DatasetMean, 

39 DatasetQuantile, 

40 DatasetStd, 

41 MeanMeasure, 

42 Measure, 

43 MeasureValue, 

44 SampleMean, 

45 SampleQuantile, 

46 SampleStd, 

47 Stat, 

48 StdMeasure, 

49) 

50from .tensor import Tensor 

51 

52 

53def _convert_axis_ids( 

54 axes: v0_4.AxesInCZYX, 

55 mode: Literal["per_sample", "per_dataset"], 

56) -> Tuple[AxisId, ...]: 

57 if not isinstance(axes, str): 

58 return tuple(axes) 

59 

60 if mode == "per_sample": 

61 ret = [] 

62 elif mode == "per_dataset": 

63 ret = [v0_5.BATCH_AXIS_ID] 

64 else: 

65 assert_never(mode) 

66 

67 ret.extend([AxisId(a) for a in axes]) 

68 return tuple(ret) 

69 

70 

71@dataclass 

72class AddKnownDatasetStats(BlockwiseOperator): 

73 dataset_stats: Mapping[Measure, MeasureValue] 

74 

75 def __post_init__(self): 

76 # keep only dataset measures 

77 self.dataset_stats = { 

78 k: v for k, v in self.dataset_stats.items() if k.scope == "dataset" 

79 } 

80 

81 @property 

82 def required_measures(self) -> Collection[Measure]: 

83 return set() 

84 

85 def __call__(self, sample: Union[Sample, SampleBlock]) -> None: 

86 sample.stat.update(self.dataset_stats) 

87 

88 

89# @dataclass 

90# class UpdateStats(Operator): 

91# """Calculates sample and/or dataset measures""" 

92 

93# measures: Union[Sequence[Measure], Set[Measure], Mapping[Measure, MeasureValue]] 

94# """sample and dataset `measuers` to be calculated by this operator. Initial/fixed 

95# dataset measure values may be given, see `keep_updating_dataset_stats` for details. 

96# """ 

97# keep_updating_dataset_stats: Optional[bool] = None 

98# """indicates if operator calls should keep updating dataset statistics or not 

99 

100# default (None): if `measures` is a `Mapping` (i.e. initial measure values are 

101# given) no further updates to dataset statistics is conducted, otherwise (w.o. 

102# initial measure values) dataset statistics are updated by each processed sample. 

103# """ 

104# _keep_updating_dataset_stats: bool = field(init=False) 

105# _stats_calculator: StatsCalculator = field(init=False) 

106 

107# @property 

108# def required_measures(self) -> Collection[Measure]: 

109# return set() 

110 

111# def __post_init__(self): 

112# self._stats_calculator = StatsCalculator(self.measures) 

113# if self.keep_updating_dataset_stats is None: 

114# self._keep_updating_dataset_stats = not isinstance(self.measures, collections.abc.Mapping) 

115# else: 

116# self._keep_updating_dataset_stats = self.keep_updating_dataset_stats 

117 

118# def __call__(self, sample_block: SampleBlockWithOrigin> None: 

119# if self._keep_updating_dataset_stats: 

120# sample.stat.update(self._stats_calculator.update_and_get_all(sample)) 

121# else: 

122# sample.stat.update(self._stats_calculator.skip_update_and_get_all(sample)) 

123 

124 

125@dataclass 

126class UpdateStats(SamplewiseOperator): 

127 """Calculates sample and/or dataset measures""" 

128 

129 stats_calculator: StatsCalculator 

130 """`StatsCalculator` to be used by this operator.""" 

131 keep_updating_initial_dataset_stats: bool = False 

132 """indicates if operator calls should keep updating initial dataset statistics or not; 

133 if the `stats_calculator` was not provided with any initial dataset statistics, 

134 these are always updated with every new sample. 

135 """ 

136 _keep_updating_dataset_stats: bool = field(init=False) 

137 

138 @property 

139 def required_measures(self) -> Collection[Measure]: 

140 return set() 

141 

142 def __post_init__(self): 

143 self._keep_updating_dataset_stats = ( 

144 self.keep_updating_initial_dataset_stats 

145 or not self.stats_calculator.has_dataset_measures 

146 ) 

147 

148 def __call__(self, sample: Sample) -> None: 

149 if self._keep_updating_dataset_stats: 

150 sample.stat.update(self.stats_calculator.update_and_get_all(sample)) 

151 else: 

152 sample.stat.update(self.stats_calculator.skip_update_and_get_all(sample)) 

153 

154 

155@dataclass 

156class Binarize(SimpleOperator): 

157 """'output = tensor > threshold'.""" 

158 

159 threshold: Union[float, Sequence[float]] 

160 axis: Optional[AxisId] = None 

161 

162 def _apply(self, x: Tensor, stat: Stat) -> Tensor: 

163 return x > self.threshold 

164 

165 @property 

166 def required_measures(self) -> Collection[Measure]: 

167 return set() 

168 

169 def get_output_shape( 

170 self, input_shape: Mapping[AxisId, int] 

171 ) -> Mapping[AxisId, int]: 

172 return input_shape 

173 

174 @classmethod 

175 def from_proc_descr( 

176 cls, descr: Union[v0_4.BinarizeDescr, v0_5.BinarizeDescr], member_id: MemberId 

177 ) -> Self: 

178 if isinstance(descr.kwargs, (v0_4.BinarizeKwargs, v0_5.BinarizeKwargs)): 

179 return cls( 

180 input=member_id, output=member_id, threshold=descr.kwargs.threshold 

181 ) 

182 elif isinstance(descr.kwargs, v0_5.BinarizeAlongAxisKwargs): 

183 return cls( 

184 input=member_id, 

185 output=member_id, 

186 threshold=descr.kwargs.threshold, 

187 axis=descr.kwargs.axis, 

188 ) 

189 else: 

190 assert_never(descr.kwargs) 

191 

192 

193@dataclass 

194class Clip(SimpleOperator): 

195 min: Optional[Union[float, SampleQuantile, DatasetQuantile]] = None 

196 """minimum value for clipping""" 

197 max: Optional[Union[float, SampleQuantile, DatasetQuantile]] = None 

198 """maximum value for clipping""" 

199 

200 def __post_init__(self): 

201 if self.min is None and self.max is None: 

202 raise ValueError("missing min or max value") 

203 

204 if ( 

205 isinstance(self.min, float) 

206 and isinstance(self.max, float) 

207 and self.min >= self.max 

208 ): 

209 raise ValueError(f"expected min < max, but {self.min} >= {self.max}") 

210 

211 if isinstance(self.min, (SampleQuantile, DatasetQuantile)) and isinstance( 

212 self.max, (SampleQuantile, DatasetQuantile) 

213 ): 

214 if self.min.axes != self.max.axes: 

215 raise NotImplementedError( 

216 f"expected min and max quantiles with same axes, but got {self.min.axes} and {self.max.axes}" 

217 ) 

218 if self.min.q >= self.max.q: 

219 raise ValueError( 

220 f"expected min quantile < max quantile, but {self.min.q} >= {self.max.q}" 

221 ) 

222 

223 @property 

224 def required_measures(self): 

225 return { 

226 arg 

227 for arg in (self.min, self.max) 

228 if isinstance(arg, (SampleQuantile, DatasetQuantile)) 

229 } 

230 

231 def _apply(self, x: Tensor, stat: Stat) -> Tensor: 

232 if isinstance(self.min, (SampleQuantile, DatasetQuantile)): 

233 min_value = stat[self.min] 

234 if isinstance(min_value, (int, float)): 

235 # use clip for scalar value 

236 min_clip_arg = min_value 

237 else: 

238 # clip does not support non-scalar values 

239 x = Tensor.from_xarray( 

240 x.data.where(x.data >= min_value.data, min_value.data) 

241 ) 

242 min_clip_arg = None 

243 else: 

244 min_clip_arg = self.min 

245 

246 if isinstance(self.max, (SampleQuantile, DatasetQuantile)): 

247 max_value = stat[self.max] 

248 if isinstance(max_value, (int, float)): 

249 # use clip for scalar value 

250 max_clip_arg = max_value 

251 else: 

252 # clip does not support non-scalar values 

253 x = Tensor.from_xarray( 

254 x.data.where(x.data <= max_value.data, max_value.data) 

255 ) 

256 max_clip_arg = None 

257 else: 

258 max_clip_arg = self.max 

259 

260 if min_clip_arg is not None or max_clip_arg is not None: 

261 x = x.clip(min_clip_arg, max_clip_arg) 

262 

263 return x 

264 

265 def get_output_shape( 

266 self, input_shape: Mapping[AxisId, int] 

267 ) -> Mapping[AxisId, int]: 

268 return input_shape 

269 

270 @classmethod 

271 def from_proc_descr( 

272 cls, descr: Union[v0_4.ClipDescr, v0_5.ClipDescr], member_id: MemberId 

273 ) -> Self: 

274 if isinstance(descr, v0_5.ClipDescr): 

275 dataset_mode, axes = _get_axes(descr.kwargs) 

276 if dataset_mode: 

277 Quantile = DatasetQuantile 

278 else: 

279 Quantile = partial(SampleQuantile, method="inverted_cdf") 

280 

281 if descr.kwargs.min is not None: 

282 min_arg = descr.kwargs.min 

283 elif descr.kwargs.min_percentile is not None: 

284 min_arg = Quantile( 

285 q=descr.kwargs.min_percentile / 100, 

286 axes=axes, 

287 member_id=member_id, 

288 ) 

289 else: 

290 min_arg = None 

291 

292 if descr.kwargs.max is not None: 

293 max_arg = descr.kwargs.max 

294 elif descr.kwargs.max_percentile is not None: 

295 max_arg = Quantile( 

296 q=descr.kwargs.max_percentile / 100, 

297 axes=axes, 

298 member_id=member_id, 

299 ) 

300 else: 

301 max_arg = None 

302 

303 elif isinstance(descr, v0_4.ClipDescr): 

304 min_arg = descr.kwargs.min 

305 max_arg = descr.kwargs.max 

306 else: 

307 assert_never(descr) 

308 

309 return cls( 

310 input=member_id, 

311 output=member_id, 

312 min=min_arg, 

313 max=max_arg, 

314 ) 

315 

316 

317@dataclass 

318class EnsureDtype(SimpleOperator): 

319 dtype: DTypeStr 

320 

321 @classmethod 

322 def from_proc_descr(cls, descr: v0_5.EnsureDtypeDescr, member_id: MemberId): 

323 return cls(input=member_id, output=member_id, dtype=descr.kwargs.dtype) 

324 

325 def get_descr(self): 

326 return v0_5.EnsureDtypeDescr(kwargs=v0_5.EnsureDtypeKwargs(dtype=self.dtype)) 

327 

328 def get_output_shape( 

329 self, input_shape: Mapping[AxisId, int] 

330 ) -> Mapping[AxisId, int]: 

331 return input_shape 

332 

333 def _apply(self, x: Tensor, stat: Stat) -> Tensor: 

334 return x.astype(self.dtype) 

335 

336 @property 

337 def required_measures(self) -> Collection[Measure]: 

338 return set() 

339 

340 

341@dataclass 

342class ScaleLinear(SimpleOperator): 

343 gain: Union[float, xr.DataArray] = 1.0 

344 """multiplicative factor""" 

345 

346 offset: Union[float, xr.DataArray] = 0.0 

347 """additive term""" 

348 

349 def _apply(self, x: Tensor, stat: Stat) -> Tensor: 

350 return x * self.gain + self.offset 

351 

352 @property 

353 def required_measures(self) -> Collection[Measure]: 

354 return set() 

355 

356 def get_output_shape( 

357 self, input_shape: Mapping[AxisId, int] 

358 ) -> Mapping[AxisId, int]: 

359 return input_shape 

360 

361 @classmethod 

362 def from_proc_descr( 

363 cls, 

364 descr: Union[v0_4.ScaleLinearDescr, v0_5.ScaleLinearDescr], 

365 member_id: MemberId, 

366 ) -> Self: 

367 kwargs = descr.kwargs 

368 if isinstance(kwargs, v0_5.ScaleLinearKwargs): 

369 axis = None 

370 elif isinstance(kwargs, v0_5.ScaleLinearAlongAxisKwargs): 

371 axis = kwargs.axis 

372 elif isinstance(kwargs, v0_4.ScaleLinearKwargs): 

373 if kwargs.axes is not None: 

374 raise NotImplementedError( 

375 "model.v0_4.ScaleLinearKwargs with axes not implemented, please consider updating the model to v0_5." 

376 ) 

377 axis = None 

378 else: 

379 assert_never(kwargs) 

380 

381 if axis: 

382 gain = xr.DataArray(np.atleast_1d(kwargs.gain), dims=axis) 

383 offset = xr.DataArray(np.atleast_1d(kwargs.offset), dims=axis) 

384 else: 

385 assert isinstance(kwargs.gain, (float, int)) or len(kwargs.gain) == 1, ( 

386 kwargs.gain 

387 ) 

388 gain = ( 

389 kwargs.gain if isinstance(kwargs.gain, (float, int)) else kwargs.gain[0] 

390 ) 

391 assert isinstance(kwargs.offset, (float, int)) or len(kwargs.offset) == 1 

392 offset = ( 

393 kwargs.offset 

394 if isinstance(kwargs.offset, (float, int)) 

395 else kwargs.offset[0] 

396 ) 

397 

398 return cls(input=member_id, output=member_id, gain=gain, offset=offset) 

399 

400 

401@dataclass 

402class ScaleMeanVariance(SimpleOperator): 

403 axes: Optional[Sequence[AxisId]] = None 

404 reference_tensor: Optional[MemberId] = None 

405 eps: float = 1e-6 

406 mean: Union[SampleMean, DatasetMean] = field(init=False) 

407 std: Union[SampleStd, DatasetStd] = field(init=False) 

408 ref_mean: Union[SampleMean, DatasetMean] = field(init=False) 

409 ref_std: Union[SampleStd, DatasetStd] = field(init=False) 

410 

411 @property 

412 def required_measures(self): 

413 return {self.mean, self.std, self.ref_mean, self.ref_std} 

414 

415 def __post_init__(self): 

416 axes = None if self.axes is None else tuple(self.axes) 

417 ref_tensor = self.reference_tensor or self.input 

418 if axes is None or AxisId("batch") not in axes: 

419 Mean = SampleMean 

420 Std = SampleStd 

421 else: 

422 Mean = DatasetMean 

423 Std = DatasetStd 

424 

425 self.mean = Mean(member_id=self.input, axes=axes) 

426 self.std = Std(member_id=self.input, axes=axes) 

427 self.ref_mean = Mean(member_id=ref_tensor, axes=axes) 

428 self.ref_std = Std(member_id=ref_tensor, axes=axes) 

429 

430 def _apply(self, x: Tensor, stat: Stat) -> Tensor: 

431 mean = stat[self.mean] 

432 std = stat[self.std] + self.eps 

433 ref_mean = stat[self.ref_mean] 

434 ref_std = stat[self.ref_std] + self.eps 

435 return (x - mean) / std * ref_std + ref_mean 

436 

437 def get_output_shape( 

438 self, input_shape: Mapping[AxisId, int] 

439 ) -> Mapping[AxisId, int]: 

440 return input_shape 

441 

442 @classmethod 

443 def from_proc_descr( 

444 cls, 

445 descr: Union[v0_4.ScaleMeanVarianceDescr, v0_5.ScaleMeanVarianceDescr], 

446 member_id: MemberId, 

447 ) -> Self: 

448 kwargs = descr.kwargs 

449 _, axes = _get_axes(descr.kwargs) 

450 

451 return cls( 

452 input=member_id, 

453 output=member_id, 

454 reference_tensor=MemberId(str(kwargs.reference_tensor)), 

455 axes=axes, 

456 eps=kwargs.eps, 

457 ) 

458 

459 

460def _get_axes( 

461 kwargs: Union[ 

462 v0_4.ZeroMeanUnitVarianceKwargs, 

463 v0_5.ZeroMeanUnitVarianceKwargs, 

464 v0_4.ScaleRangeKwargs, 

465 v0_5.ScaleRangeKwargs, 

466 v0_4.ScaleMeanVarianceKwargs, 

467 v0_5.ScaleMeanVarianceKwargs, 

468 v0_5.ClipKwargs, 

469 ], 

470) -> Tuple[bool, Optional[Tuple[AxisId, ...]]]: 

471 if kwargs.axes is None: 

472 return True, None 

473 elif isinstance(kwargs.axes, str): 

474 axes = _convert_axis_ids(kwargs.axes, kwargs["mode"]) 

475 return AxisId("b") in axes, axes 

476 elif isinstance(kwargs.axes, collections.abc.Sequence): 

477 axes = tuple(kwargs.axes) 

478 return AxisId("batch") in axes, axes 

479 else: 

480 assert_never(kwargs.axes) 

481 

482 

483@dataclass 

484class ScaleRange(SimpleOperator): 

485 lower_quantile: InitVar[Optional[Union[SampleQuantile, DatasetQuantile]]] = None 

486 upper_quantile: InitVar[Optional[Union[SampleQuantile, DatasetQuantile]]] = None 

487 lower: Union[SampleQuantile, DatasetQuantile] = field(init=False) 

488 upper: Union[SampleQuantile, DatasetQuantile] = field(init=False) 

489 

490 eps: float = 1e-6 

491 

492 def __post_init__( 

493 self, 

494 lower_quantile: Optional[Union[SampleQuantile, DatasetQuantile]], 

495 upper_quantile: Optional[Union[SampleQuantile, DatasetQuantile]], 

496 ): 

497 if lower_quantile is None: 

498 tid = self.input if upper_quantile is None else upper_quantile.member_id 

499 self.lower = DatasetQuantile(q=0.0, member_id=tid) 

500 else: 

501 self.lower = lower_quantile 

502 

503 if upper_quantile is None: 

504 self.upper = DatasetQuantile(q=1.0, member_id=self.lower.member_id) 

505 else: 

506 self.upper = upper_quantile 

507 

508 assert self.lower.member_id == self.upper.member_id 

509 assert self.lower.q < self.upper.q 

510 assert self.lower.axes == self.upper.axes 

511 

512 @property 

513 def required_measures(self): 

514 return {self.lower, self.upper} 

515 

516 def get_output_shape( 

517 self, input_shape: Mapping[AxisId, int] 

518 ) -> Mapping[AxisId, int]: 

519 return input_shape 

520 

521 @classmethod 

522 def from_proc_descr( 

523 cls, 

524 descr: Union[v0_4.ScaleRangeDescr, v0_5.ScaleRangeDescr], 

525 member_id: MemberId, 

526 ): 

527 kwargs = descr.kwargs 

528 ref_tensor = ( 

529 member_id 

530 if kwargs.reference_tensor is None 

531 else MemberId(str(kwargs.reference_tensor)) 

532 ) 

533 dataset_mode, axes = _get_axes(descr.kwargs) 

534 if dataset_mode: 

535 Quantile = DatasetQuantile 

536 else: 

537 Quantile = partial(SampleQuantile, method="linear") 

538 

539 return cls( 

540 input=member_id, 

541 output=member_id, 

542 lower_quantile=Quantile( 

543 q=kwargs.min_percentile / 100, 

544 axes=axes, 

545 member_id=ref_tensor, 

546 ), 

547 upper_quantile=Quantile( 

548 q=kwargs.max_percentile / 100, 

549 axes=axes, 

550 member_id=ref_tensor, 

551 ), 

552 ) 

553 

554 def _apply(self, x: Tensor, stat: Stat) -> Tensor: 

555 lower = stat[self.lower] 

556 upper = stat[self.upper] 

557 return (x - lower) / (upper - lower + self.eps) 

558 

559 def get_descr(self): 

560 assert self.lower.axes == self.upper.axes 

561 assert self.lower.member_id == self.upper.member_id 

562 

563 return v0_5.ScaleRangeDescr( 

564 kwargs=v0_5.ScaleRangeKwargs( 

565 axes=self.lower.axes, 

566 min_percentile=self.lower.q * 100, 

567 max_percentile=self.upper.q * 100, 

568 eps=self.eps, 

569 reference_tensor=self.lower.member_id, 

570 ) 

571 ) 

572 

573 

574@dataclass 

575class Sigmoid(SimpleOperator): 

576 """1 / (1 + e^(-input)).""" 

577 

578 def _apply(self, x: Tensor, stat: Stat) -> Tensor: 

579 return Tensor(1.0 / (1.0 + np.exp(-x)), dims=x.dims) 

580 

581 @property 

582 def required_measures(self) -> Collection[Measure]: 

583 return {} 

584 

585 def get_output_shape( 

586 self, input_shape: Mapping[AxisId, int] 

587 ) -> Mapping[AxisId, int]: 

588 return input_shape 

589 

590 @classmethod 

591 def from_proc_descr( 

592 cls, descr: Union[v0_4.SigmoidDescr, v0_5.SigmoidDescr], member_id: MemberId 

593 ) -> Self: 

594 assert isinstance(descr, (v0_4.SigmoidDescr, v0_5.SigmoidDescr)) 

595 return cls(input=member_id, output=member_id) 

596 

597 def get_descr(self): 

598 return v0_5.SigmoidDescr() 

599 

600 

601@dataclass 

602class Softmax(SimpleOperator): 

603 """Softmax activation function.""" 

604 

605 axis: AxisId = AxisId("channel") 

606 

607 def _apply(self, x: Tensor, stat: Stat) -> Tensor: 

608 axis_idx = x.dims.index(self.axis) 

609 result = scipy.special.softmax(x.data, axis=axis_idx) 

610 result_xr = xr.DataArray(result, dims=x.dims) 

611 return Tensor.from_xarray(result_xr) 

612 

613 @property 

614 def required_measures(self) -> Collection[Measure]: 

615 return set() 

616 

617 def get_output_shape( 

618 self, input_shape: Mapping[AxisId, int] 

619 ) -> Mapping[AxisId, int]: 

620 return input_shape 

621 

622 @classmethod 

623 def from_proc_descr(cls, descr: v0_5.SoftmaxDescr, member_id: MemberId) -> Self: 

624 assert isinstance(descr, v0_5.SoftmaxDescr) 

625 return cls(input=member_id, output=member_id, axis=descr.kwargs.axis) 

626 

627 def get_descr(self): 

628 return v0_5.SoftmaxDescr(kwargs=v0_5.SoftmaxKwargs(axis=self.axis)) 

629 

630 

631@dataclass 

632class ZeroMeanUnitVariance(SimpleOperator): 

633 """normalize to zero mean, unit variance.""" 

634 

635 mean: MeanMeasure 

636 std: StdMeasure 

637 

638 eps: float = 1e-6 

639 

640 def __post_init__(self): 

641 assert self.mean.axes == self.std.axes 

642 

643 @property 

644 def required_measures(self) -> Set[Union[MeanMeasure, StdMeasure]]: 

645 return {self.mean, self.std} 

646 

647 def get_output_shape( 

648 self, input_shape: Mapping[AxisId, int] 

649 ) -> Mapping[AxisId, int]: 

650 return input_shape 

651 

652 @classmethod 

653 def from_proc_descr( 

654 cls, 

655 descr: Union[v0_4.ZeroMeanUnitVarianceDescr, v0_5.ZeroMeanUnitVarianceDescr], 

656 member_id: MemberId, 

657 ): 

658 dataset_mode, axes = _get_axes(descr.kwargs) 

659 

660 if dataset_mode: 

661 Mean = DatasetMean 

662 Std = DatasetStd 

663 else: 

664 Mean = SampleMean 

665 Std = SampleStd 

666 

667 return cls( 

668 input=member_id, 

669 output=member_id, 

670 mean=Mean(axes=axes, member_id=member_id), 

671 std=Std(axes=axes, member_id=member_id), 

672 ) 

673 

674 def _apply(self, x: Tensor, stat: Stat) -> Tensor: 

675 mean = stat[self.mean] 

676 std = stat[self.std] 

677 return (x - mean) / (std + self.eps) 

678 

679 def get_descr(self): 

680 return v0_5.ZeroMeanUnitVarianceDescr( 

681 kwargs=v0_5.ZeroMeanUnitVarianceKwargs(axes=self.mean.axes, eps=self.eps) 

682 ) 

683 

684 

685@dataclass 

686class FixedZeroMeanUnitVariance(SimpleOperator): 

687 """normalize to zero mean, unit variance with precomputed values.""" 

688 

689 mean: Union[float, xr.DataArray] 

690 std: Union[float, xr.DataArray] 

691 

692 eps: float = 1e-6 

693 

694 def __post_init__(self): 

695 assert ( 

696 isinstance(self.mean, (int, float)) 

697 or isinstance(self.std, (int, float)) 

698 or self.mean.dims == self.std.dims 

699 ) 

700 

701 @property 

702 def required_measures(self) -> Collection[Measure]: 

703 return set() 

704 

705 def get_output_shape( 

706 self, input_shape: Mapping[AxisId, int] 

707 ) -> Mapping[AxisId, int]: 

708 return input_shape 

709 

710 @classmethod 

711 def from_proc_descr( 

712 cls, 

713 descr: v0_5.FixedZeroMeanUnitVarianceDescr, 

714 member_id: MemberId, 

715 ) -> Self: 

716 if isinstance(descr.kwargs, v0_5.FixedZeroMeanUnitVarianceKwargs): 

717 dims = None 

718 elif isinstance(descr.kwargs, v0_5.FixedZeroMeanUnitVarianceAlongAxisKwargs): 

719 dims = (AxisId(descr.kwargs.axis),) 

720 else: 

721 assert_never(descr.kwargs) 

722 

723 return cls( 

724 input=member_id, 

725 output=member_id, 

726 mean=xr.DataArray(descr.kwargs.mean, dims=dims), 

727 std=xr.DataArray(descr.kwargs.std, dims=dims), 

728 ) 

729 

730 def get_descr(self): 

731 if isinstance(self.mean, (int, float)): 

732 assert isinstance(self.std, (int, float)) 

733 kwargs = v0_5.FixedZeroMeanUnitVarianceKwargs(mean=self.mean, std=self.std) 

734 else: 

735 assert isinstance(self.std, xr.DataArray) 

736 assert len(self.mean.dims) == 1 

737 kwargs = v0_5.FixedZeroMeanUnitVarianceAlongAxisKwargs( 

738 axis=AxisId(str(self.mean.dims[0])), 

739 mean=list(self.mean), 

740 std=list(self.std), 

741 ) 

742 

743 return v0_5.FixedZeroMeanUnitVarianceDescr(kwargs=kwargs) 

744 

745 def _apply(self, x: Tensor, stat: Stat) -> Tensor: 

746 return (x - self.mean) / (self.std + self.eps) 

747 

748 

749@dataclass 

750class CustomProcessing(SimpleOperator): 

751 """Execute a user-supplied custom processing callable. 

752 

753 Two styles are supported — callable class and factory function:: 

754 

755 # Callable class style 

756 class my_factory: 

757 def __init__(self, threshold=0.5): 

758 self.threshold = threshold 

759 def __call__(self, *arrays): 

760 return (arrays[0] > self.threshold).astype(np.uint8) 

761 

762 # Factory function style 

763 def my_factory(threshold=0.5): 

764 def run(*arrays): 

765 return (arrays[0] > threshold).astype(np.uint8) 

766 return run 

767 

768 Runtime protocol: ``custom_callable = my_factory(**kwargs)`` once at construction; 

769 ``result = custom_callable(tensor)`` once per sample. 

770 

771 Note: The custom callable may not change the shape of the input tensor. 

772 """ 

773 

774 custom_factory: Callable[..., Callable[[NDArray[Any]], NDArray[Any]]] 

775 

776 kwargs: Mapping[str, Any] 

777 """Keyword arguments forwarded to the custom factory.""" 

778 

779 # Initialised in __post_init__ 

780 _custom_callable: Any = field(init=False, repr=False) 

781 

782 def __post_init__(self) -> None: 

783 self._custom_callable = self.custom_factory(**self.kwargs) 

784 

785 def _apply(self, x: Tensor, stat: Stat) -> Tensor: 

786 return Tensor.from_numpy(self._custom_callable(x.to_numpy()), dims=x.dims)