Repository navigation
Expand file tree
/
Copy pathaesaraf.py
More file actions
1173 lines (937 loc) · 35.7 KB
/
Copy pathaesaraf.py
File metadata and controls
1173 lines (937 loc) · 35.7 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
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
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
# Copyright 2020 The PyMC Developers
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import warnings
from typing import (
Callable,
Dict,
Generator,
Iterable,
List,
Optional,
Sequence,
Set,
Tuple,
Union,
)
import aesara
import aesara.tensor as at
import numpy as np
import pandas as pd
import scipy.sparse as sps
from aesara import scalar
from aesara.compile import Function, Mode, get_mode
from aesara.gradient import grad
from aesara.graph import node_rewriter, rewrite_graph
from aesara.graph.basic import (
Apply,
Constant,
Variable,
clone_get_equiv,
graph_inputs,
vars_between,
walk,
)
from aesara.graph.fg import FunctionGraph
from aesara.graph.op import Op
from aesara.sandbox.rng_mrg import MRG_RandomStream as RandomStream
from aesara.scalar.basic import Cast
from aesara.tensor.basic import _as_tensor_variable
from aesara.tensor.elemwise import Elemwise
from aesara.tensor.random.op import RandomVariable
from aesara.tensor.random.var import (
RandomGeneratorSharedVariable,
RandomStateSharedVariable,
)
from aesara.tensor.rewriting.basic import topo_constant_folding
from aesara.tensor.rewriting.shape import ShapeFeature
from aesara.tensor.sharedvar import SharedVariable, TensorSharedVariable
from aesara.tensor.subtensor import AdvancedIncSubtensor, AdvancedIncSubtensor1
from aesara.tensor.var import TensorConstant, TensorVariable
from pymc.exceptions import NotConstantValueError
from pymc.logprob.transforms import RVTransform
from pymc.logprob.utils import CheckParameterValue
from pymc.vartypes import continuous_types, isgenerator, typefilter
PotentialShapeType = Union[int, np.ndarray, Sequence[Union[int, Variable]], TensorVariable]
__all__ = [
"gradient",
"hessian",
"hessian_diag",
"inputvars",
"cont_inputs",
"floatX",
"intX",
"smartfloatX",
"jacobian",
"CallableTensor",
"join_nonshared_inputs",
"make_shared_replacements",
"generator",
"set_at_rng",
"at_rng",
"convert_observed_data",
"compile_pymc",
"constant_fold",
]
def convert_observed_data(data):
"""Convert user provided dataset to accepted formats."""
if hasattr(data, "to_numpy") and hasattr(data, "isnull"):
# typically, but not limited to pandas objects
vals = data.to_numpy()
null_data = data.isnull()
if hasattr(null_data, "to_numpy"):
# pandas Series
mask = null_data.to_numpy()
else:
# pandas Index
mask = null_data
if mask.any():
# there are missing values
ret = np.ma.MaskedArray(vals, mask)
else:
ret = vals
elif isinstance(data, np.ndarray):
if isinstance(data, np.ma.MaskedArray):
if not data.mask.any():
# empty mask
ret = data.filled()
else:
# already masked and rightly so
ret = data
else:
# already a ndarray, but not masked
mask = np.isnan(data)
if np.any(mask):
ret = np.ma.MaskedArray(data, mask)
else:
# no masking required
ret = data
elif isinstance(data, Variable):
ret = data
elif sps.issparse(data):
ret = data
elif isgenerator(data):
ret = generator(data)
else:
ret = np.asarray(data)
# type handling to enable index variables when data is int:
if hasattr(data, "dtype"):
if "int" in str(data.dtype):
return intX(ret)
# otherwise, assume float:
else:
return floatX(ret)
# needed for uses of this function other than with pm.Data:
else:
return floatX(ret)
@_as_tensor_variable.register(pd.Series)
@_as_tensor_variable.register(pd.DataFrame)
def dataframe_to_tensor_variable(df: pd.DataFrame, *args, **kwargs) -> TensorVariable:
return at.as_tensor_variable(df.to_numpy(), *args, **kwargs)
def extract_obs_data(x: TensorVariable) -> np.ndarray:
"""Extract data from observed symbolic variables.
Raises
------
TypeError
"""
if isinstance(x, Constant):
return x.data
if isinstance(x, SharedVariable):
return x.get_value()
if x.owner and isinstance(x.owner.op, Elemwise) and isinstance(x.owner.op.scalar_op, Cast):
array_data = extract_obs_data(x.owner.inputs[0])
return array_data.astype(x.type.dtype)
if x.owner and isinstance(x.owner.op, (AdvancedIncSubtensor, AdvancedIncSubtensor1)):
array_data = extract_obs_data(x.owner.inputs[0])
mask_idx = tuple(extract_obs_data(i) for i in x.owner.inputs[2:])
mask = np.zeros_like(array_data)
mask[mask_idx] = 1
return np.ma.MaskedArray(array_data, mask)
raise TypeError(f"Data cannot be extracted from {x}")
def walk_model(
graphs: Iterable[TensorVariable],
stop_at_vars: Optional[Set[TensorVariable]] = None,
expand_fn: Callable[[TensorVariable], Iterable[TensorVariable]] = lambda var: [],
) -> Generator[TensorVariable, None, None]:
"""Walk model graphs and yield their nodes.
Parameters
==========
graphs
The graphs to walk.
stop_at_vars
A list of variables at which the walk will terminate.
expand_fn
A function that returns the next variable(s) to be traversed.
"""
if stop_at_vars is None:
stop_at_vars = set()
def expand(var):
new_vars = expand_fn(var)
if var.owner and var not in stop_at_vars:
new_vars.extend(reversed(var.owner.inputs))
return new_vars
yield from walk(graphs, expand, bfs=False)
def _replace_rvs_in_graphs(
graphs: Iterable[TensorVariable],
replacement_fn: Callable[[TensorVariable], Dict[TensorVariable, TensorVariable]],
**kwargs,
) -> Tuple[List[TensorVariable], Dict[TensorVariable, TensorVariable]]:
"""Replace random variables in graphs
This will *not* recompute test values.
Parameters
==========
graphs
The graphs in which random variables are to be replaced.
Returns
=======
Tuple containing the transformed graphs and a ``dict`` of the replacements
that were made.
"""
replacements = {}
def expand_replace(var):
new_nodes = []
if var.owner:
# Call replacement_fn to update replacements dict inplace and, optionally,
# specify new nodes that should also be walked for replacements. This
# includes `value` variables that are not simple input variables, and may
# contain other `random` variables in their graphs (e.g., IntervalTransform)
new_nodes.extend(replacement_fn(var, replacements))
return new_nodes
# This iteration populates the replacements
for var in walk_model(graphs, expand_fn=expand_replace, **kwargs):
pass
if replacements:
inputs = [i for i in graph_inputs(graphs) if not isinstance(i, Constant)]
equiv = {k: k for k in replacements.keys()}
equiv = clone_get_equiv(inputs, graphs, False, False, equiv)
fg = FunctionGraph(
[equiv[i] for i in inputs],
[equiv[o] for o in graphs],
clone=False,
)
# replacements have to be done in reverse topological order so that nested
# expressions get recursively replaced correctly
toposort = fg.toposort()
sorted_replacements = sorted(
tuple(replacements.items()),
key=lambda pair: toposort.index(pair[0].owner),
reverse=True,
)
fg.replace_all(sorted_replacements, import_missing=True)
graphs = list(fg.outputs)
return graphs, replacements
def rvs_to_value_vars(
graphs: Iterable[Variable],
apply_transforms: bool = True,
**kwargs,
) -> List[Variable]:
"""Clone and replace random variables in graphs with their value variables.
This will *not* recompute test values in the resulting graphs.
Parameters
==========
graphs
The graphs in which to perform the replacements.
apply_transforms
If ``True``, apply each value variable's transform.
"""
warnings.warn(
"rvs_to_value_vars is deprecated. Use model.replace_rvs_by_values instead",
FutureWarning,
)
def populate_replacements(
random_var: TensorVariable, replacements: Dict[TensorVariable, TensorVariable]
) -> List[TensorVariable]:
# Populate replacements dict with {rv: value} pairs indicating which graph
# RVs should be replaced by what value variables.
value_var = getattr(
random_var.tag, "observations", getattr(random_var.tag, "value_var", None)
)
# No value variable to replace RV with
if value_var is None:
return []
transform = getattr(value_var.tag, "transform", None)
if transform is not None and apply_transforms:
# We want to replace uses of the RV by the back-transformation of its value
value_var = transform.backward(value_var, *random_var.owner.inputs)
replacements[random_var] = value_var
# Also walk the graph of the value variable to make any additional replacements
# if that is not a simple input variable
return [value_var]
# Clone original graphs
inputs = [i for i in graph_inputs(graphs) if not isinstance(i, Constant)]
equiv = clone_get_equiv(inputs, graphs, False, False, {})
graphs = [equiv[n] for n in graphs]
graphs, _ = _replace_rvs_in_graphs(
graphs,
replacement_fn=populate_replacements,
**kwargs,
)
return graphs
def replace_rvs_by_values(
graphs: Sequence[TensorVariable],
*,
rvs_to_values: Dict[TensorVariable, TensorVariable],
rvs_to_transforms: Dict[TensorVariable, RVTransform],
**kwargs,
) -> List[TensorVariable]:
"""Clone and replace random variables in graphs with their value variables.
This will *not* recompute test values in the resulting graphs.
Parameters
----------
graphs
The graphs in which to perform the replacements.
rvs_to_values
Mapping between the original graph RVs and respective value variables
rvs_to_transforms
Mapping between the original graph RVs and respective value transforms
"""
# Clone original graphs so that we don't modify variables in place
inputs = [i for i in graph_inputs(graphs) if not isinstance(i, Constant)]
equiv = clone_get_equiv(inputs, graphs, False, False, {})
graphs = [equiv[n] for n in graphs]
# Get needed mappings for equivalent cloned variables
equiv_rvs_to_values = {}
equiv_rvs_to_transforms = {}
for rv, value in rvs_to_values.items():
equiv_rv = equiv.get(rv, rv)
equiv_rvs_to_values[equiv_rv] = equiv.get(value, value)
equiv_rvs_to_transforms[equiv_rv] = rvs_to_transforms[rv]
def poulate_replacements(rv, replacements):
# Populate replacements dict with {rv: value} pairs indicating which graph
# RVs should be replaced by what value variables.
# No value variable to replace RV with
value = equiv_rvs_to_values.get(rv, None)
if value is None:
return []
transform = equiv_rvs_to_transforms.get(rv, None)
if transform is not None:
# We want to replace uses of the RV by the back-transformation of its value
value = transform.backward(value, *rv.owner.inputs)
# The value may have a less precise type than the rv. In this case
# filter_variable will add a SpecifyShape to ensure they are consistent
value = rv.type.filter_variable(value, allow_convert=True)
value.name = rv.name
replacements[rv] = value
# Also walk the graph of the value variable to make any additional
# replacements if that is not a simple input variable
return [value]
graphs, _ = _replace_rvs_in_graphs(
graphs,
replacement_fn=poulate_replacements,
**kwargs,
)
return graphs
def inputvars(a):
"""
Get the inputs into Aesara variables
Parameters
----------
a: Aesara variable
Returns
-------
r: list of tensor variables that are inputs
"""
return [
v
for v in graph_inputs(makeiter(a))
if isinstance(v, TensorVariable) and not isinstance(v, TensorConstant)
]
def cont_inputs(a):
"""
Get the continuous inputs into Aesara variables
Parameters
----------
a: Aesara variable
Returns
-------
r: list of tensor variables that are continuous inputs
"""
return typefilter(inputvars(a), continuous_types)
def floatX(X):
"""
Convert an Aesara tensor or numpy array to aesara.config.floatX type.
"""
try:
return X.astype(aesara.config.floatX)
except AttributeError:
# Scalar passed
return np.asarray(X, dtype=aesara.config.floatX)
_conversion_map = {"float64": "int32", "float32": "int16", "float16": "int8", "float8": "int8"}
def intX(X):
"""
Convert a aesara tensor or numpy array to aesara.tensor.int32 type.
"""
intX = _conversion_map[aesara.config.floatX]
try:
return X.astype(intX)
except AttributeError:
# Scalar passed
return np.asarray(X, dtype=intX)
def smartfloatX(x):
"""
Converts numpy float values to floatX and leaves values of other types unchanged.
"""
if str(x.dtype).startswith("float"):
x = floatX(x)
return x
"""
Aesara derivative functions
"""
def gradient1(f, v):
"""flat gradient of f wrt v"""
return at.flatten(grad(f, v, disconnected_inputs="warn"))
empty_gradient = at.zeros(0, dtype="float32")
def gradient(f, vars=None):
if vars is None:
vars = cont_inputs(f)
if vars:
return at.concatenate([gradient1(f, v) for v in vars], axis=0)
else:
return empty_gradient
def jacobian1(f, v):
"""jacobian of f wrt v"""
f = at.flatten(f)
idx = at.arange(f.shape[0], dtype="int32")
def grad_i(i):
return gradient1(f[i], v)
return aesara.map(grad_i, idx)[0]
def jacobian(f, vars=None):
if vars is None:
vars = cont_inputs(f)
if vars:
return at.concatenate([jacobian1(f, v) for v in vars], axis=1)
else:
return empty_gradient
def jacobian_diag(f, x):
idx = at.arange(f.shape[0], dtype="int32")
def grad_ii(i, f, x):
return grad(f[i], x)[i]
return aesara.scan(
grad_ii, sequences=[idx], n_steps=f.shape[0], non_sequences=[f, x], name="jacobian_diag"
)[0]
@aesara.config.change_flags(compute_test_value="ignore")
def hessian(f, vars=None):
return -jacobian(gradient(f, vars), vars)
@aesara.config.change_flags(compute_test_value="ignore")
def hessian_diag1(f, v):
g = gradient1(f, v)
idx = at.arange(g.shape[0], dtype="int32")
def hess_ii(i):
return gradient1(g[i], v)[i]
return aesara.map(hess_ii, idx)[0]
@aesara.config.change_flags(compute_test_value="ignore")
def hessian_diag(f, vars=None):
if vars is None:
vars = cont_inputs(f)
if vars:
return -at.concatenate([hessian_diag1(f, v) for v in vars], axis=0)
else:
return empty_gradient
def makeiter(a):
if isinstance(a, (tuple, list)):
return a
else:
return [a]
class IdentityOp(scalar.UnaryScalarOp):
@staticmethod
def st_impl(x):
return x
def impl(self, x):
return x
def grad(self, inp, grads):
return grads
def c_code(self, node, name, inp, out, sub):
return f"{out[0]} = {inp[0]};"
def __eq__(self, other):
return isinstance(self, type(other))
def __hash__(self):
return hash(type(self))
scalar_identity = IdentityOp(scalar.upgrade_to_float, name="scalar_identity")
identity = Elemwise(scalar_identity, name="identity")
def make_shared_replacements(point, vars, model):
"""
Makes shared replacements for all *other* variables than the ones passed.
This way functions can be called many times without setting unchanging variables. Allows us
to use func.trust_input by removing the need for DictToArrayBijection and kwargs.
Parameters
----------
point: dictionary mapping variable names to sample values
vars: list of variables not to make shared
model: model
Returns
-------
Dict of variable -> new shared variable
"""
othervars = set(model.value_vars) - set(vars)
return {
var: aesara.shared(point[var.name], var.name + "_shared", shape=var.type.shape)
for var in othervars
}
def join_nonshared_inputs(
point: Dict[str, np.ndarray],
outputs: List[TensorVariable],
inputs: List[TensorVariable],
shared_inputs: Optional[Dict[TensorVariable, TensorSharedVariable]] = None,
make_inputs_shared: bool = False,
) -> Tuple[List[TensorVariable], TensorVariable]:
"""
Create new outputs and input TensorVariables where the non-shared inputs are joined
in a single raveled vector input.
Parameters
----------
point : dict of {str : array_like}
Dictionary that maps each input variable name to a numerical variable. The values
are used to extract the shape of each input variable to establish a correct
mapping between joined and original inputs. The shape of each variable is
assumed to be fixed.
outputs : list of TensorVariable
List of output TensorVariables whose non-shared inputs will be replaced
by a joined vector input.
inputs : list of TensorVariable
List of input TensorVariables which will be replaced by a joined vector input.
shared_inputs : dict of {TensorVariable : TensorSharedVariable}, optional
Dict of TensorVariable and their associated TensorSharedVariable in
subgraph replacement.
make_inputs_shared : bool, default False
Whether to make the joined vector input a shared variable.
Returns
-------
new_outputs : list of TensorVariable
List of new outputs `outputs` TensorVariables that depend on `joined_inputs` and new shared variables as inputs.
joined_inputs : TensorVariable
Joined input vector TensorVariable for the `new_outputs`
Examples
--------
Join the inputs of a simple Aesara graph.
.. code-block:: python
import aesara.tensor as at
import numpy as np
from pymc.aesaraf import join_nonshared_inputs
# Original non-shared inputs
x = at.scalar("x")
y = at.vector("y")
# Original output
out = x + y
print(out.eval({x: np.array(1), y: np.array([1, 2, 3])})) # [2, 3, 4]
# New output and inputs
[new_out], joined_inputs = join_nonshared_inputs(
point={ # Only shapes matter
"x": np.zeros(()),
"y": np.zeros(3),
},
outputs=[out],
inputs=[x, y],
)
print(new_out.eval({
joined_inputs: np.array([1, 1, 2, 3]),
})) # [2, 3, 4]
Join the input value variables of a model logp.
.. code-block:: python
import pymc as pm
with pm.Model() as model:
mu_pop = pm.Normal("mu_pop")
sigma_pop = pm.HalfNormal("sigma_pop")
mu = pm.Normal("mu", mu_pop, sigma_pop, shape=(3, ))
y = pm.Normal("y", mu, 1.0, observed=[0, 1, 2])
print(model.compile_logp()({
"mu_pop": 0,
"sigma_pop_log__": 1,
"mu": [0, 1, 2],
})) # -12.691227342634292
initial_point = model.initial_point()
inputs = model.value_vars
[logp], joined_inputs = join_nonshared_inputs(
point=initial_point,
outputs=[model.logp()],
inputs=inputs,
)
print(logp.eval({
joined_inputs: [0, 1, 0, 1, 2],
})) # -12.691227342634292
Same as above but with the `mu_pop` value variable being shared.
.. code-block:: python
from aesara import shared
mu_pop_input, *other_inputs = inputs
shared_mu_pop_input = shared(0.0)
[logp], other_joined_inputs = join_nonshared_inputs(
point=initial_point,
outputs=[model.logp()],
inputs=other_inputs,
shared_inputs={
mu_pop_input: shared_mu_pop_input
},
)
print(logp.eval({
other_joined_inputs: [1, 0, 1, 2],
})) # -12.691227342634292
"""
if not inputs:
raise ValueError("Empty list of input variables.")
raveled_inputs = at.concatenate([var.ravel() for var in inputs])
if not make_inputs_shared:
tensor_type = raveled_inputs.type
joined_inputs = tensor_type("joined_inputs")
else:
joined_values = np.concatenate([point[var.name].ravel() for var in inputs])
joined_inputs = aesara.shared(joined_values, "joined_inputs")
if aesara.config.compute_test_value != "off":
joined_inputs.tag.test_value = raveled_inputs.tag.test_value
replace: Dict[TensorVariable, TensorVariable] = {}
last_idx = 0
for var in inputs:
shape = point[var.name].shape
arr_len = np.prod(shape, dtype=int)
replace[var] = joined_inputs[last_idx : last_idx + arr_len].reshape(shape).astype(var.dtype)
last_idx += arr_len
if shared_inputs is not None:
replace.update(shared_inputs)
new_outputs = [
aesara.clone_replace(output, replace, rebuild_strict=False) for output in outputs
]
return new_outputs, joined_inputs
class PointFunc:
"""Wraps so a function so it takes a dict of arguments instead of arguments."""
def __init__(self, f):
self.f = f
def __call__(self, state):
return self.f(**state)
class CallableTensor:
"""Turns a symbolic variable with one input into a function that returns symbolic arguments
with the one variable replaced with the input.
"""
def __init__(self, tensor):
self.tensor = tensor
def __call__(self, input):
"""Replaces the single input of symbolic variable to be the passed argument.
Parameters
----------
input: TensorVariable
"""
(oldinput,) = inputvars(self.tensor)
return aesara.clone_replace(self.tensor, {oldinput: input}, rebuild_strict=False)
class GeneratorOp(Op):
"""
Generator Op is designed for storing python generators inside aesara graph.
__call__ creates TensorVariable
It has 2 new methods
- var.set_gen(gen): sets new generator
- var.set_default(value): sets new default value (None erases default value)
If generator is exhausted, variable will produce default value if it is not None,
else raises `StopIteration` exception that can be caught on runtime.
Parameters
----------
gen: generator that implements __next__ (py3) or next (py2) method
and yields np.arrays with same types
default: np.array with the same type as generator produces
"""
__props__ = ("generator",)
def __init__(self, gen, default=None):
from pymc.data import GeneratorAdapter
super().__init__()
if not isinstance(gen, GeneratorAdapter):
gen = GeneratorAdapter(gen)
self.generator = gen
self.set_default(default)
def make_node(self, *inputs):
gen_var = self.generator.make_variable(self)
return Apply(self, [], [gen_var])
def perform(self, node, inputs, output_storage, params=None):
if self.default is not None:
output_storage[0][0] = next(self.generator, self.default)
else:
output_storage[0][0] = next(self.generator)
def do_constant_folding(self, fgraph, node):
return False
__call__ = aesara.config.change_flags(compute_test_value="off")(Op.__call__)
def set_gen(self, gen):
from pymc.data import GeneratorAdapter
if not isinstance(gen, GeneratorAdapter):
gen = GeneratorAdapter(gen)
if not gen.tensortype == self.generator.tensortype:
raise ValueError("New generator should yield the same type")
self.generator = gen
def set_default(self, value):
if value is None:
self.default = None
else:
value = np.asarray(value, self.generator.tensortype.dtype)
t1 = (False,) * value.ndim
t2 = self.generator.tensortype.broadcastable
if not t1 == t2:
raise ValueError("Default value should have the same type as generator")
self.default = value
def generator(gen, default=None):
"""
Generator variable with possibility to set default value and new generator.
If generator is exhausted variable will produce default value if it is not None,
else raises `StopIteration` exception that can be caught on runtime.
Parameters
----------
gen: generator that implements __next__ (py3) or next (py2) method
and yields np.arrays with same types
default: np.array with the same type as generator produces
Returns
-------
TensorVariable
It has 2 new methods
- var.set_gen(gen): sets new generator
- var.set_default(value): sets new default value (None erases default value)
"""
return GeneratorOp(gen, default)()
_at_rng = RandomStream()
def at_rng(random_seed=None):
"""
Get the package-level random number generator or new with specified seed.
Parameters
----------
random_seed: int
If not None
returns *new* aesara random generator without replacing package global one
Returns
-------
`aesara.tensor.random.utils.RandomStream` instance
`aesara.tensor.random.utils.RandomStream`
instance passed to the most recent call of `set_at_rng`
"""
if random_seed is None:
return _at_rng
else:
ret = RandomStream(random_seed)
return ret
def set_at_rng(new_rng):
"""
Set the package-level random number generator.
Parameters
----------
new_rng: `aesara.tensor.random.utils.RandomStream` instance
The random number generator to use.
"""
# pylint: disable=global-statement
global _at_rng
# pylint: enable=global-statement
if isinstance(new_rng, int):
new_rng = RandomStream(new_rng)
_at_rng = new_rng
def floatX_array(x):
return floatX(np.array(x))
def ix_(*args):
"""
Aesara np.ix_ analog
See numpy.lib.index_tricks.ix_ for reference
"""
out = []
nd = len(args)
for k, new in enumerate(args):
if new is None:
out.append(slice(None))
new = at.as_tensor(new)
if new.ndim != 1:
raise ValueError("Cross index must be 1 dimensional")
new = new.reshape((1,) * k + (new.size,) + (1,) * (nd - k - 1))
out.append(new)
return tuple(out)
def largest_common_dtype(tensors):
dtypes = {
str(t.dtype) if hasattr(t, "dtype") else smartfloatX(np.asarray(t)).dtype for t in tensors
}
return np.stack([np.ones((), dtype=dtype) for dtype in dtypes]).dtype
@node_rewriter(tracks=[CheckParameterValue])
def local_remove_check_parameter(fgraph, node):
"""Rewrite that removes CheckParameterValue
This is used when compile_rv_inplace
"""
if isinstance(node.op, CheckParameterValue):
return [node.inputs[0]]
@node_rewriter(tracks=[CheckParameterValue])
def local_check_parameter_to_ninf_switch(fgraph, node):
if isinstance(node.op, CheckParameterValue):
logp_expr, *logp_conds = node.inputs
if len(logp_conds) > 1:
logp_cond = at.all(logp_conds)
else:
(logp_cond,) = logp_conds
out = at.switch(logp_cond, logp_expr, -np.inf)
out.name = node.op.msg
if out.dtype != node.outputs[0].dtype:
out = at.cast(out, node.outputs[0].dtype)
return [out]
aesara.compile.optdb["canonicalize"].register(
"local_remove_check_parameter",
local_remove_check_parameter,
use_db_name_as_tag=False,
)
aesara.compile.optdb["canonicalize"].register(
"local_check_parameter_to_ninf_switch",
local_check_parameter_to_ninf_switch,
use_db_name_as_tag=False,
)
def find_rng_nodes(
variables: Iterable[Variable],
) -> List[Union[RandomStateSharedVariable, RandomGeneratorSharedVariable]]:
"""Return RNG variables in a graph"""
return [
node
for node in graph_inputs(variables)
if isinstance(node, (RandomStateSharedVariable, RandomGeneratorSharedVariable))
]
def replace_rng_nodes(outputs: Sequence[TensorVariable]) -> Sequence[TensorVariable]:
"""Replace any RNG nodes upstream of outputs by new RNGs of the same type