Coverage for mlprodict/npy/onnx

1"""

2@file

3@brief Intermediate class between :epkg:`numpy` and :epkg:`onnx`.

5.. versionadded:: 0.6

6"""

7import logging

8import numpy

9from onnx.helper import make_tensor

10from ..onnx_tools.onnx2py_helper import guess_proto_dtype

11from .xop_variable import Variable

12from .xop import loadop, OnnxOperatorItem, OnnxOperatorTuple

13from .xop_variable import guess_numpy_type

15logger = logging.getLogger('xop')

18try:

19 numpy_bool = numpy.bool_

20except AttributeError: # pragma: no cover

21 numpy_bool = bool

22try:

23 numpy_str = numpy.str_

24except AttributeError: # pragma: no cover

25 numpy_str = str

28class OnnxVar:

29 """

30 Variables used into :epkg:`onnx` computation.

32 :param inputs: variable name or object

33 :param op: :epkg:`ONNX` operator

34 :param select_output: if multiple output are returned by

35 ONNX operator *op*, it takes only one specifed by this

36 argument

37 :param dtype: specifies the type of the variable

38 held by this class (*op* is None) in that case

39 :param kwargs: addition argument to give operator *op*

41 .. versionadded:: 0.6

42 """

43 __array_ufunc__ = None

45 def __init__(self, *inputs, op=None, select_output=None,

46 dtype=None, **kwargs):

47 logger.debug('OnnxVar(%d in, dtype=%r, op=%r, select_output=%r)',

48 len(inputs), dtype, op, select_output)

49 self.inputs = inputs

50 self.select_output = select_output

51 self.onnx_op = op

52 self.alg_ = None

53 self.onnx_op_kwargs = kwargs

54 if dtype is not None and (op is not None or len(inputs) != 1):

55 raise RuntimeError( # pragma: no cover

56 "dtype can only be used if op is None or len(inputs) == 1.")

57 for i, inp in enumerate(self.inputs):

58 if isinstance(inp, type):

59 raise TypeError( # pragma: no cover

60 "Unexpected type for input %d - %r." % (i, inp))

61 if not isinstance(inp, numpy.ndarray):

62 continue

63 if (inp.size > 0 and

64 isinstance(inp.ravel()[0], (numpy.ndarray, OnnxVar))):

65 raise TypeError( # pragma: no cover

66 "Unexpected type for input %d: %r, %r, "

67 "op=%r" % (i, type(inp), inp.ravel()[0], op))

68 self.dtype = self._guess_dtype(dtype, from_init=True)

70 def _guess_dtype(self, dtype, from_init=False):

71 "Guesses dtype when not specified."

72 if dtype is not None:

73 return dtype

74 dtypes = []

75 for i, inp in enumerate(self.inputs):

76 if isinstance(inp, str):

77 return None

78 if isinstance(inp, numpy.ndarray):

79 dtypes.append(inp.dtype)

80 elif isinstance(inp, Variable):

81 dtypes.append(inp.dtype)

82 elif isinstance(inp, OnnxVar):

83 dtypes.append(inp.dtype)

84 elif isinstance(inp, MultiOnnxVar):

85 dtypes.append(inp._guess_dtype(dtype))

86 elif isinstance(inp, (numpy.float32, numpy.float64,

87 numpy.int32, numpy.int64)):

88 dtypes.append(inp.dtype)

89 elif isinstance(inp, numpy_str):

90 dtypes.append(numpy_str)

91 elif isinstance(inp, numpy_bool):

92 dtypes.append(numpy_bool)

93 elif isinstance(inp, int):

94 dtypes.append(numpy.int64) # pragma: no cover

95 elif isinstance(inp, float):

96 dtypes.append(numpy.float64)

97 elif hasattr(inp, 'fit'):

98 # scikit-learn model

99 continue

100 elif hasattr(inp, '_guess_dtype'):

101 dtypes.append(inp._guess_dtype(dtype))

102 else:

103 try:

104 dtype = guess_numpy_type(inp)

105 except NotImplementedError as e:

106 raise TypeError( # pragma: no cover

107 "Unexpected type for input %i type=%r." % (

108 i, type(inp))) from e

109 dtypes.append(dtype)

110 dtypes = [_ for _ in dtypes if _ is not None]

111 unique = set(dtypes)

112 if len(unique) != 1:

113 return None

114 return dtypes[0]

115

116 def __repr__(self):

117 "usual"

118 args = []

119 for inp in self.inputs:

120 args.append(repr(inp))

121 if self.onnx_op is not None:

122 if isinstance(self.onnx_op, str):

123 args.append("op=%r" % self.onnx_op)

124 else:

125 args.append("op=%s" % self.onnx_op.__name__)

126 if self.select_output is not None:

127 args.append("select_output=%r" % self.select_output)

128 if self.dtype is not None and self.dtype != self._guess_dtype(None):

129 args.append("dtype=%r" % self.dtype)

130 for k, v in sorted(self.onnx_op_kwargs.items()):

131 args.append("%s=%r" % (k, v))

132 res = "%s(%s)" % (self.__class__.__name__, ", ".join(args))

133 return res

134

135 def set_onnx_name(self, name_type):

136 """

137 Forces this variable to get this name during

138

139 :param name_type: a tuple *(name, type)*

140 """

141 self.onnx_input_type_ = name_type

142

143 def to_algebra(self, op_version=None):

144 """

145 Converts the variable into an operator.

146 """

147 if self.alg_ is not None:

148 return self.alg_

149

150 if self.onnx_op is None:

151 logger.debug('OnnxVar.to_algebra:1(op_version=%r)', op_version)

152 if len(self.inputs) != 1:

153 raise RuntimeError( # pragma: no cover

154 "Unexpected number of inputs, 1 expected, "

155 "got {} instead.".format(self.inputs))

156 if self.dtype is None or hasattr(self.inputs[0], 'onnx_name'):

157 self.alg_ = Variable.from_skl2onnx(self.inputs[0])

158 elif isinstance(self.inputs[0], Variable):

159 self.alg_ = self.inputs[0]

160 else:

161 self.alg_ = Variable(self.inputs[0], self.dtype)

162 else:

163 logger.debug('OnnxVar.to_algebra:2(op_version=%r) - onnx_op=%r',

164 op_version, self.onnx_op)

165 if isinstance(self.onnx_op, str):

166 var = self._custom_op(*self.inputs, op_version=op_version,

167 **self.onnx_op_kwargs)

168 alg = var.to_algebra(op_version=op_version)

169 if not hasattr(self, 'alg_'):

170 raise RuntimeError( # pragma: no cover

171 "Missing attribute 'alg_'.")

172 self.alg_ = alg

173 return alg

174

175 new_inputs = []

176 for inp in self.inputs:

177 if hasattr(inp, 'fit'):

178 # scikit-learn model

179 new_inputs.append(inp)

180 elif isinstance(inp, (

181 int, float, str, numpy.ndarray, numpy.int32,

182 numpy.int64, numpy.float32, numpy.float64,

183 numpy_bool, numpy_str, numpy.int8, numpy.uint8,

184 numpy.int16, numpy.uint16, numpy.uint32,

185 numpy.uint64)):

186 if (inp.size > 0 and

187 isinstance(

188 inp.ravel()[0], # pylint: disable=E1101

189 (numpy.ndarray, OnnxVar))):

190 raise TypeError( # pragma: no cover

191 "Unexpected type for an input %r, %r."

192 "" % (type(inp), inp.ravel()[0])) # pylint: disable=E1101

193 new_inputs.append(inp)

194 else:

195 new_inputs.append(

196 inp.to_algebra(op_version=op_version))

197

198 res = self.onnx_op(*new_inputs, op_version=op_version,

199 **self.onnx_op_kwargs)

200 if self.select_output is None:

201 self.alg_ = res

202 else:

203 self.alg_ = res[self.select_output]

204 return self.alg_

205

206 def _custom_op(self, *args, op_version=None, runtime=None, **kwargs):

207 """

208 This could be handled before a call to this method

209 but this method can change the conversion of an non-existing

210 operator depending on the given opset.

211 """

212 if self.onnx_op == 'filter':

213 return self._custom_op_filter(*args, op_version=op_version,

214 runtime=runtime, **kwargs)

215 raise NotImplementedError( # pragma: no cover

216 "Unexpected custom operator %r." % self.onnx_op)

217

218 def _custom_op_filter(self, *args, op_version=None, runtime=None, **kwargs):

219 """

220 This could be handled before a call to this method

221 but this method can change the conversion of an non-existing

222 operator depending on the given opset.

223 """

224 OnnxSqueeze, OnnxTopK, OnnxGather, OnnxReduceSum = loadop(

225 'Squeeze', 'TopK', 'Gather', 'ReduceSum')

226 if len(args) != 2:

227 raise RuntimeError( # pragma: no cover

228 "Custom op 'filter' expects two inputs not %r." % len(args))

229 if len(kwargs) != 0:

230 raise RuntimeError( # pragma: no cover

231 "Custom op 'filter' expects no arguments but got %r." % kwargs)

232 mat, index = args

233 cast = OnnxVar(index.astype(numpy.int64), op=OnnxSqueeze)

234 n1 = OnnxVar(cast, op=OnnxReduceSum, keepdims=1)

235 indices = OnnxVar(cast, n1, op=OnnxTopK, select_output=1)

236 return OnnxVar(mat, indices, op=OnnxGather)

237

238 @property

239 def T(self):

240 "Transpose."

241 OnnxTranspose = loadop('Transpose')

242 return OnnxVar(self, op=OnnxTranspose)

243

244 def astype(self, dtype):

245 "Cast"

246 OnnxCast = loadop('Cast')

247 return OnnxVar(self, op=OnnxCast, to=guess_proto_dtype(dtype))

248

249 @property

250 def shape(self):

251 "Shape"

252 OnnxShape = loadop('Shape')

253 return OnnxVar(self, op=OnnxShape)

254

255 @property

256 def size(self):

257 "Size"

258 OnnxSize = loadop('Size')

259 return OnnxVar(self, op=OnnxSize)

260

261 def reshape(self, shape):

262 "Reshape"

263 OnnxReshape = loadop('Reshape')

264 if isinstance(shape, (tuple, list)):

265 shape = numpy.array(shape, dtype=numpy.int64)

266 return OnnxVar(self, shape, op=OnnxReshape)

267

268 def _make_array(self, y):

269 """Converts *y* into an array if not."""

270 if isinstance(y, (numpy.ndarray, OnnxVar)):

271 return y

272 if hasattr(y, 'dtype'):

273 return numpy.full((1, ), y, dtype=y.dtype)

274 if isinstance(y, str):

275 return numpy.array([y])

276 if isinstance(y, float):

277 return numpy.array([y], dtype=numpy.float32)

278 if isinstance(y, int):

279 return numpy.array([y], dtype=numpy.int64)

280 return y

281

282 def __add__(self, y):

283 "Addition."

284 y = self._make_array(y)

285 OnnxAdd = loadop('Add')

286 return OnnxVar(self, y, op=OnnxAdd)

287

288 def __radd__(self, y):

289 "Right Addition."

290 y = self._make_array(y)

291 OnnxIdentity, OnnxAdd = loadop('Identity', 'Add')

292 return OnnxVar(OnnxVar(y, op=OnnxIdentity), self, op=OnnxAdd)

293

294 def __sub__(self, y):

295 "Subtraction."

296 y = self._make_array(y)

297 OnnxSub = loadop('Sub')

298 return OnnxVar(self, y, op=OnnxSub)

299

300 def __rsub__(self, y):

301 "Right subtraction."

302 y = self._make_array(y)

303 OnnxIdentity, OnnxSub = loadop('Identity', 'Sub')

304 return OnnxVar(OnnxVar(y, op=OnnxIdentity), self, op=OnnxSub)

305

306 def __mul__(self, y):

307 "Multiplication."

308 y = self._make_array(y)

309 OnnxMul = loadop('Mul')

310 return OnnxVar(self, y, op=OnnxMul)

311

312 def __rmul__(self, y):

313 "Right multiplication."

314 y = self._make_array(y)

315 OnnxIdentity = loadop('Identity')

316 return OnnxVar(y, op=OnnxIdentity) * self

317

318 def __pow__(self, y):

319 "Power."

320 y = self._make_array(y)

321 OnnxPow = loadop('Pow')

322 return OnnxVar(self, y, op=OnnxPow)

323

324 def __mod__(self, y):

325 "Modulo."

326 y = self._make_array(y)

327 OnnxMod = loadop('Mod')

328 return OnnxVar(self, y, op=OnnxMod)

329

330 def __matmul__(self, y):

331 "Matrix multiplication."

332 y = self._make_array(y)

333 OnnxMatMul = loadop('MatMul')

334 return OnnxVar(self, y, op=OnnxMatMul)

335

336 def __truediv__(self, y):

337 "Division, no difference between `/` and `//`."

338 y = self._make_array(y)

339 OnnxDiv = loadop('Div')

340 return OnnxVar(self, y, op=OnnxDiv)

341

342 def __rtruediv__(self, y):

343 "Division, no difference between `/` and `//`."

344 y = self._make_array(y)

345 OnnxIdentity, OnnxDiv = loadop('Identity', 'Div')

346 return OnnxVar(OnnxVar(y, op=OnnxIdentity), self, op=OnnxDiv)

347

348 def __floordiv__(self, y):

349 "Division, no difference between `/` and `//`."

350 y = self._make_array(y)

351 OnnxDiv = loadop('Div')

352 return OnnxVar(self, y, op=OnnxDiv)

353

354 def __eq__(self, y):

355 "Equality."

356 y = self._make_array(y)

357 OnnxEqual = loadop('Equal')

358 return OnnxVar(self, y, op=OnnxEqual)

359

360 def __ne__(self, y):

361 "Difference."

362 y = self._make_array(y)

363 OnnxEqual, OnnxNot = loadop('Equal', 'Not')

364 return OnnxVar(OnnxVar(self, y, op=OnnxEqual), op=OnnxNot)

365

366 def __ge__(self, y):

367 "Greater or Equal."

368 y = self._make_array(y)

369 OnnxGreaterOrEqual = loadop('GreaterOrEqual')

370 return OnnxVar(self, y, op=OnnxGreaterOrEqual)

371

372 def __gt__(self, y):

373 "Greater."

374 y = self._make_array(y)

375 OnnxGreater = loadop('Greater')

376 return OnnxVar(self, y, op=OnnxGreater)

377

378 def __invert__(self):

379 "not."

380 OnnxNot = loadop('Not')

381 return OnnxVar(self, op=OnnxNot)

382

383 def __le__(self, y):

384 "Less or Equal."

385 y = self._make_array(y)

386 OnnxLessOrEqual = loadop('LessOrEqual')

387 return OnnxVar(self, y, op=OnnxLessOrEqual)

388

389 def __lt__(self, y):

390 "Less."

391 y = self._make_array(y)

392 OnnxLess = loadop('Less')

393 return OnnxVar(self, y, op=OnnxLess)

394

395 def __and__(self, y):

396 "And."

397 y = self._make_array(y)

398 OnnxAnd = loadop('And')

399 return OnnxVar(self, y, op=OnnxAnd)

400

401 def __or__(self, y):

402 "And."

403 y = self._make_array(y)

404 OnnxOr = loadop('Or')

405 return OnnxVar(self, y, op=OnnxOr)

406

407 def not_(self):

408 "Not."

409 OnnxNot = loadop('Not')

410 return OnnxVar(self, op=OnnxNot)

411

412 def __neg__(self):

413 "Neg."

414 OnnxNeg = loadop('Neg')

415 return OnnxVar(self, op=OnnxNeg)

416

417 def __getitem__(self, index):

418 """

419 Deals with multiple scenarios.

420

421 * *index* is an integer or a slice, a tuple of integers and slices,

422 example: `[0, 1]`, `[:5, :6]`, `[::2]` (**scenario 1**)

423 * *index* is an *ONNX* object (more precisely an instance of

424 @see cl OnnxVar), then the method assumes it is an array of

425 boolean to select a subset of the tensor along the first axis,

426 example: `mat[mat == 0]` (**scenario 2**)

427 """

428 if isinstance(index, OnnxVar):

429 # scenario 2

430 return OnnxVar(self, index, op='filter')

431

432 if isinstance(index, int):

433 # Use Gather instead.

434 OnnxGather = loadop('Gather')

435 return OnnxVar(

436 self, numpy.array(index, dtype=numpy.int64),

437 axis=0, op=OnnxGather)

438

439 if not isinstance(index, tuple):

440 index = (index, )

441

442 # only one integer?

443 ni = None

444 ax = None

445 for i, a in enumerate(index):

446 if isinstance(a, int):

447 if ni is None:

448 ni = i

449 ax = a

450 else:

451 ax = None

452 ni = None

453 break

454 if (isinstance(a, slice) and a.start is None and

455 a.stop is None and a.step is None):

456 continue

457 ax = None

458 ni = None

459 break

460 if ni is not None and ax is not None:

461 # Use Gather instead.

462 OnnxGather = loadop('Gather')

463 return OnnxVar(

464 self, numpy.array(ni, dtype=numpy.int64),

465 axis=ax, op=OnnxGather)

466

467 # scenario 1

468 starts = []

469 ends = []

470 axes = []

471 steps = []

472 axis_squeeze = []

473 needs_shape = []

474 for i, ind in enumerate(index):

475 if isinstance(ind, int):

476 starts.append(ind)

477 ends.append(ind + 1)

478 axes.append(i)

479 steps.append(1)

480 axis_squeeze.append(i)

481 continue

482 if isinstance(ind, slice):

483 if ind.start is None and ind.stop is None and ind.step is None:

484 continue

485 start = 0 if ind.start is None else ind.start

486 end = (None, i) if ind.stop is None else ind.stop

487 step = 1 if ind.step is None else ind.step

488 starts.append(start)

489 ends.append(end)

490 axes.append(i)

491 steps.append(step)

492 if isinstance(end, tuple):

493 needs_shape.append(len(ends) - 1)

494 elif isinstance(end, OnnxVar):

495 needs_shape.append(end)

496 continue

497 raise NotImplementedError( # pragma: no cover

498 "Not implemented for type %r." % type(ind))

499

500 if max(steps) == min(steps) == 1:

501 steps = None

502 else:

503 steps = numpy.array(steps, dtype=numpy.int64)

504

505 starts = numpy.array(starts, dtype=numpy.int64)

506 axes = numpy.array(axes, dtype=numpy.int64)

507

508 OnnxGather, OnnxSlice, OnnxSqueeze, OnnxConcat = loadop(

509 'Gather', 'Slice', 'Squeeze', 'Concat')

510 if len(needs_shape) > 0:

511 shape = self.shape

512 conc = []

513 for e in ends:

514 if isinstance(e, tuple):

515 conc.append(

516 OnnxVar(shape, numpy.array([e[1]], numpy.int64),

517 op=OnnxGather))

518 elif isinstance(e, OnnxVar):

519 conc.append(

520 e.reshape(numpy.array([-1], dtype=numpy.int64)))

521 else:

522 conc.append(numpy.array([e], dtype=numpy.int64))

523 if len(conc) > 1:

524 ends = OnnxVar(*conc, op=OnnxConcat, axis=0)

525 else:

526 ends = conc[0]

527 else:

528 ends = numpy.array(ends, dtype=numpy.int64)

529

530 if steps is None:

531 sliced = OnnxVar(self, starts, ends, axes, op=OnnxSlice)

532 else:

533 sliced = OnnxVar(self, starts, ends, axes, steps, op=OnnxSlice)

534 if len(axis_squeeze) > 0:

535 return OnnxVar(

536 sliced, numpy.array(axis_squeeze, dtype=numpy.int64),

537 op=OnnxSqueeze)

538 return sliced

539

540 def __setitem__(self, index, value):

541 """

542 Only supports vectors (1D tensor).

543

544 * *index* is an integer or a slice, a tuple of integers and slices,

545 example: `[0]`, `[:5]`, `[::2]` (**scenario 1**)

546 * *index* is an *ONNX* object (more precisely an instance of

547 @see cl OnnxVar), then the method assumes it is an array of

548 boolean to select a subset of the tensor along the first axis,

549 example: `mat[mat == 0]` (**scenario 2**)

550 This processing is applied before the operator it contains.

551 A copy should be made (Identity node or copy method).

552 """

553 OnnxIdentity = loadop('Identity')

554 if self.onnx_op is not None and self.onnx_op is not OnnxIdentity:

555 raise RuntimeError( # pragma: no cover

556 "A copy should be made before setting new values on a matrix. "

557 "Method copy() would do that.")

558

559 if isinstance(index, OnnxVar):

560 # scenario 2, example: cp[x < 0] = -1

561 return self._setitem2i_(index, value)

562 elif not isinstance(index, tuple):

563 index = (index, )

564

565 for i in index:

566 if isinstance(i, OnnxVar):

567 raise NotImplementedError( # pragma: no cover

568 "Unable to handle case such as cp[0, x < 0] = -1.")

569

570 # scenario 1

571 if len(index) == 1:

572 return self._setitem1i_(index[0], value)

573 raise NotImplementedError( # pragma: no cover

574 "Indices in %d dimensions are not implemented yet." % len(index))

575

576 def _setitem1i_(self, index, value):

577 sl = None

578 if isinstance(index, slice):

579 start = 0 if index.start is None else index.start

580 stop = index.stop

581 step = index.step

582 sl = [start, stop, step]

583 elif isinstance(index, int):

584 sl = [index, index + 1, 1]

585 else:

586 raise NotImplementedError( # pragma: no cover

587 "Unable to assign new values due to unexpected type %r."

588 "" % type(index))

589

590 if sl[1] is None and isinstance(value, numpy.ndarray):

591 sl[1] = sl[0] + value.size

592 OnnxConstantOfShape, OnnxScatterElements = loadop(

593 'ConstantOfShape', 'ScatterElements')

594 if sl[1] is None:

595 if sl[2] is not None and sl[2] != 1:

596 raise NotImplementedError( # pragma: no cover

597 "If the length is not known, step must be 1 not %d." % sl[2])

598 value = make_tensor(

599 "value", guess_proto_dtype(value.dtype), (1, ), [value]) # pylint: disable=E1101

600 inp = self.inputs[0]

601 if not isinstance(inp, OnnxVar):

602 raise RuntimeError( # pragma: no cover

603 "Input must be an instance of OnnxVar not %r." % type(inp))

604 cst = OnnxVar(inp.shape, op=OnnxConstantOfShape, value=value)

605 ext = inp[:sl[0]]

606 indices = numpy.arange(0, sl[0]).astype(numpy.int64)

607 add_step = OnnxVar(cst, indices, ext,

608 op=OnnxScatterElements, axis=0)

609 else:

610 indices = numpy.arange(sl[0], sl[1], sl[2]).astype(numpy.int64)

611 if isinstance(value, numpy.ndarray):

612 values = value

613 else:

614 values = numpy.full(indices.shape, value)

615 add_step = OnnxVar(self.inputs[0], indices, values,

616 op=OnnxScatterElements, axis=0)

617

618 self.inputs = [add_step]

619 return self

620

621 def _setitem2i_(self, index, value):

622 OnnxWhere = loadop('Where')

623 add_step = OnnxVar(index, value, self.inputs[0], op=OnnxWhere)

624 self.inputs = [add_step]

625 return self

626

627 def copy(self):

628 """

629 Returns a copy of self (use of Identity node).

630 """

631 OnnxIdentity = loadop('Identity')

632 return OnnxVar(self, op=OnnxIdentity)

633

634 def flatten(self, axis=0):

635 """

636 Flattens a matrix (see :epkg:`numpy:ndarray:flatten`).

637

638 :param axis: only flatten from axis to the end.

639 :return: @see cl OnnxVar.

640 """

641 OnnxFlatten, OnnxSqueeze = loadop('Flatten', 'Squeeze')

642 fl = OnnxVar(self, op=OnnxFlatten, axis=axis)

643 if axis == 0:

644 return OnnxVar(fl, numpy.array([0], dtype=numpy.int64),

645 op=OnnxSqueeze)

646 return fl

647

648

649class MultiOnnxVar:

650 """

651 Class used to return multiple @see cl OnnxVar

652 at the same time.

653 """

654

655 def __init__(self, *inputs, op=None, dtype=None, **kwargs):

656 "constructor"

657 logger.debug('MultiOnnxVar(%d in, dtype=%r, op=%r)',

658 len(inputs), dtype, op)

659 self.onxvar = OnnxVar(*inputs, op=op, dtype=None, **kwargs)

660 self.alg_ = None

661

662 def _guess_dtype(self, dtype):

663 "Guesses dtype when not specified."

664 return self.onxvar._guess_dtype(dtype)

665

666 @property

667 def inputs(self):

668 "Returns `self.onxvar.inputs`."

669 return self.onxvar.inputs

670

671 @property

672 def onnx_op(self):

673 "Returns `self.onxvar.onnx_op`."

674 return self.onxvar.onnx_op

675

676 @property

677 def onnx_op_kwargs(self):

678 "Returns `self.onxvar.onnx_op_kwargs`."

679 return self.onxvar.onnx_op_kwargs

680

681 def to_algebra(self, op_version=None):

682 """

683 Converts the variable into an operator.

684 """

685 if self.alg_ is None:

686 logger.debug('MultiOnnxVar.to_algebra(op_version=%r)',

687 op_version)

688 new_inputs = []

689 for inp in self.inputs:

690 if isinstance(inp, (

691 int, float, str, numpy.ndarray, numpy.int32,

692 numpy.int64, numpy.float32, numpy.float64,

693 numpy_bool, numpy_str, numpy.int8, numpy.uint8,

694 numpy.int16, numpy.uint16, numpy.uint32,

695 numpy.uint64)):

696 new_inputs.append(inp)

697 elif hasattr(inp, 'fit'):

698 # scikit-learn models

699 new_inputs.append(inp)

700 else:

701 new_inputs.append(

702 inp.to_algebra(op_version=op_version))

703

704 if self.onnx_op is None:

705 if len(new_inputs) == 1:

706 logger.debug('MultiOnnxVar.to_algebra:1:new_inputs[0]=%r',

707 new_inputs[0])

708 self.alg_ = OnnxOperatorTuple(new_inputs[0])

709 else:

710 logger.debug('MultiOnnxVar.to_algebra:2:new_inputs=%r',

711 new_inputs)

712 self.alg_ = OnnxOperatorTuple(

713 new_inputs[0], *(new_inputs[1:]))

714 else:

715 logger.debug('MultiOnnxVar.to_algebra:%s:new_inputs=%r',

716 self.onnx_op.__class__.__name__, new_inputs)

717 res = self.onnx_op( # pylint: disable=E1102

718 *new_inputs, op_version=op_version, **self.onnx_op_kwargs)

719 self.alg_ = OnnxOperatorTuple(res)

720 return self.alg_

721

722 def __getitem__(self, index):

723 """

724 Returns the ith elements.

725 """

726 return OnnxVar(self, index=index, op=OnnxOperatorItem)

Coverage for mlprodict/npy/onnx_variable.py: 94%

418 statements