/
builders.py
995 lines (860 loc) · 37.4 KB
/
builders.py
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
"""Define new Ops from existing Ops"""
from collections import OrderedDict
from collections.abc import Sequence
from copy import copy
from functools import partial
from typing import cast
import pytensor.tensor as pt
from pytensor.compile.function import function
from pytensor.compile.function.pfunc import rebuild_collect_shared
from pytensor.compile.mode import optdb
from pytensor.compile.sharedvalue import SharedVariable
from pytensor.configdefaults import config
from pytensor.gradient import DisconnectedType, Rop, grad
from pytensor.graph.basic import (
Apply,
Constant,
NominalVariable,
Variable,
graph_inputs,
io_connection_pattern,
)
from pytensor.graph.fg import FunctionGraph
from pytensor.graph.null_type import NullType
from pytensor.graph.op import HasInnerGraph, Op
from pytensor.graph.replace import clone_replace
from pytensor.graph.rewriting.basic import in2out, node_rewriter
from pytensor.graph.utils import MissingInputError
from pytensor.tensor.rewriting.shape import ShapeFeature
def infer_shape(outs, inputs, input_shapes):
"""
Compute the shape of the outputs given the shape of the inputs of an PyTensor
graph.
We do it this way to avoid compiling the inner function just to get
the shape. Changes to ShapeFeature could require changes in this function.
"""
# We use a ShapeFeature because it has all the necessary logic
# inside. We don't use the full ShapeFeature interface, but we
# let it initialize itself with an empty fgraph, otherwise we will
# need to do it manually
for inp, inp_shp in zip(inputs, input_shapes):
if inp_shp is not None and len(inp_shp) != inp.type.ndim:
assert len(inp_shp) == inp.type.ndim
shape_feature = ShapeFeature()
shape_feature.on_attach(FunctionGraph([], []))
# Initialize shape_of with the input shapes
for inp, inp_shp in zip(inputs, input_shapes):
shape_feature.set_shape(inp, inp_shp)
def local_traverse(out):
"""
Go back in the graph, from out, adding computable shapes to shape_of.
"""
if out in shape_feature.shape_of:
# Its shape is already known
return
elif out.owner is None:
# This is an input of the graph
shape_feature.init_r(out)
else:
# Recurse over inputs
for inp in out.owner.inputs:
if inp not in shape_feature.shape_of:
local_traverse(inp)
# shape_feature.on_import does not actually use an fgraph
# It will call infer_shape and set_shape appropriately
dummy_fgraph = None
shape_feature.on_import(dummy_fgraph, out.owner, reason="dummy")
ret = []
for o in outs:
local_traverse(o)
ret.append(shape_feature.shape_of[o])
return ret
def construct_nominal_fgraph(
inputs: Sequence[Variable], outputs: Sequence[Variable]
) -> tuple[
FunctionGraph,
Sequence[Variable],
dict[Variable, Variable],
dict[Variable, Variable],
]:
"""Construct an inner-`FunctionGraph` with ordered nominal inputs."""
implicit_shared_inputs = []
dummy_inputs = [inp.type() for inp in inputs]
dummy_implicit_shared_inputs = []
for var in graph_inputs(outputs, inputs):
if var in inputs:
continue
if isinstance(var, SharedVariable):
# We allow shared inputs to be added automatically to the graph
implicit_shared_inputs.append(var)
dummy_implicit_shared_inputs.append(var.type())
elif not isinstance(var, Constant):
raise MissingInputError(f"NominalGraph is missing an input: {var}")
replacements = dict(
zip(
inputs + implicit_shared_inputs, dummy_inputs + dummy_implicit_shared_inputs
)
)
new = rebuild_collect_shared(
cast(Sequence[Variable], outputs),
inputs=inputs + implicit_shared_inputs,
replace=replacements,
copy_inputs_over=False,
)
(
local_inputs,
local_outputs,
(clone_d, update_d, update_expr, new_shared_inputs),
) = new
assert len(local_inputs) == len(inputs) + len(implicit_shared_inputs)
assert len(local_outputs) == len(outputs)
assert not update_d
assert not update_expr
assert not new_shared_inputs
fgraph = FunctionGraph(local_inputs, local_outputs, clone=False)
# The inputs need to be `NominalVariable`s so that we can merge
# inner-graphs
nominal_local_inputs = tuple(
NominalVariable(n, var.type) for n, var in enumerate(local_inputs)
)
fgraph.replace_all(zip(local_inputs, nominal_local_inputs))
for i, inp in enumerate(fgraph.inputs):
nom_inp = nominal_local_inputs[i]
fgraph.inputs[i] = nom_inp
fgraph.clients.pop(inp, None)
fgraph.add_input(nom_inp)
return fgraph, implicit_shared_inputs, update_d, update_expr
class OpFromGraph(Op, HasInnerGraph):
r"""
This creates an `Op` from inputs and outputs lists of variables.
The signature is similar to :func:`pytensor.function <pytensor.function>`
and the resulting `Op`'s perform will do the same operation as::
orig_function(inputs, outputs, **kwargs)
Currently does not support ``updates`` or ``givens`` argument.
.. TODO:
- examples for a multi-layer mlp. where?
- __hash__, __eq__ otherwise won't merge, try
is_same_graph_with_merge(op1.local_outputs, op2,
local_outputs)
- c_code() to remove the double overhead?
- grad() make it support DisconnectedType and the new interface
- add support for NullType and DisconnectedType when R_op supports them
- check how it works with updates.
- Add support to pickle this Op.
- Add support/test with random generator
- Add optimization to removing unused inputs/outputs
- Add optimization to work inplace on inputs when not inline
Notes
-----
- We support shared variables in the inner graph. This is automatic
and invisible to the user. They can be as input to the node or in
the inner graph.
- We support unused inputs. This is needed for the grad.
- We support nested OpFromGraph.
- ``inline=True`` will cause better runtime optimization at the cost
of compilation time. Currently only works with ``fast_compile`` or
``fast_run`` mode.
- For overriding, it's recommended to provide pure functions (no side
effects like setting global variable) as callable(s). The callable(s)
supplied for overriding gradient/rop will be called only once at the
first call to grad/R_op, and will be converted to OpFromGraph instances.
Examples
--------
Example 1:
.. code-block:: python
from pytensor import function, tensor as pt
from pytensor.compile.builders import OpFromGraph
x, y, z = pt.scalars('xyz')
e = x + y * z
op = OpFromGraph([x, y, z], [e])
# op behaves like a normal pytensor op
e2 = op(x, y, z) + op(z, y, x)
fn = function([x, y, z], [e2])
Example 2 with shared variable:
.. code-block:: python
import numpy as np
import pytensor
from pytensor import config, function, tensor as pt
from pytensor.compile.builders import OpFromGraph
x, y, z = pt.scalars('xyz')
s = pytensor.shared(np.random.random((2, 2)).astype(config.floatX))
e = x + y * z + s
op = OpFromGraph([x, y, z], [e])
# op behaves like a normal pytensor op
e2 = op(x, y, z) + op(z, y, x)
fn = function([x, y, z], [e2])
Example 3 override gradient
.. code-block:: python
from pytensor import function, tensor as pt, grad
from pytensor.compile.builders import OpFromGraph
x, y, z = pt.scalars('xyz')
e = x + y * z
def rescale_dy(inps, grads):
x, y, z = inps
g, = grads
return z*2
op = OpFromGraph(
[x, y, z], [e], grad_overrides=['default', rescale_dy, 'default']
e2 = op(x, y, z)
dx, dy, dz = grad(e2, [x, y, z])
fn = function([x, y, z], [dx, dy, dz])
# the gradient wrt y is now doubled
fn(2., 3., 4.) # [1., 8., 3.]
"""
TYPE_ERR_MSG = (
"L_op/gradient override should be (single or list of)"
"'default' | OpFromGraph | callable | Variable "
"with NullType or DisconnectedType, got %s"
)
STYPE_ERR_MSG = (
"Overriding Variable instance can only have type"
" of DisconnectedType or NullType, got %s"
)
LOP_TYPE_ERR_MSG = 'L_op type can only be "grad" or "lop", got %s.'
OV_INP_LEN_ERR_MSG = "expect overrider with %d inputs, got %d"
@staticmethod
def _filter_grad_var(grad, inp):
# Returns (filtered_var, overrider_var)
# Args:
# grad: gradient Variable
# inp: the corresponding input of gradient Variable
#
# a grad() call could return instance of NullType() or DisconnectedType()
# which cannot be directly used in OfG
#
# Since we always use an OfG instance as self._lop_op, the current
# workaround is to "remember" the special cases of the gradient and
# replace them after self._lop_op is called.
#
# This helper function changes invalid types into a filtered_var,
# and provides a overrider_var to be replaced at grad() call
#
# For now, this converts NullType or DisconnectedType into zeros_like.
# other types are unmodified: overrider_var -> None
if isinstance(grad.type, NullType | DisconnectedType):
if hasattr(inp, "zeros_like"):
return inp.zeros_like(), grad
else:
return pt.constant(0.0), grad
else:
return grad, None
@staticmethod
def _filter_rop_var(inpJ, out):
# mostly similar to _filter_grad_var
if isinstance(inpJ.type, NullType):
return out.zeros_like(), inpJ
if isinstance(inpJ.type, DisconnectedType):
# since R_op does not have DisconnectedType yet, we will just
# make them zeros.
return out.zeros_like(), None
else:
return inpJ, None
def __init__(
self,
inputs: list[Variable],
outputs: list[Variable],
*,
inline: bool = False,
lop_overrides: str = "default",
grad_overrides: str = "default",
rop_overrides: str = "default",
connection_pattern: list[list[bool]] | None = None,
strict: bool = False,
name: str | None = None,
**kwargs,
):
"""
Parameters
----------
inputs
The inputs to the graph.
outputs
The outputs to the graph.
inline
Defaults to ``False``
``True`` : Cause the :class:`Op`'s original graph being used during
compilation, the :class:`Op` will not be visible in the compiled
graph but rather its internal graph.
``False`` : will use a pre-compiled function inside.
grad_overrides
Defaults to ``'default'``.
This argument is mutually exclusive with ``lop_overrides``.
``'default'`` : Do not override, use default grad() result
`OpFromGraph`: Override with another `OpFromGraph`, should
accept inputs as the same order and types of ``inputs`` and ``output_grads``
arguments as one would specify in :meth:`Op.grad`() method.
`callable`: Should take two args: ``inputs`` and ``output_grads``.
Each argument is expected to be a list of :class:`Variable `.
Must return list of :class:`Variable `.
lop_overrides
Defaults to ``'default'``.
This argument is mutually exclusive with ``grad_overrides``.
These options are similar to the ``grad_overrides`` above, but for
the :meth:`Op.L_op` method.
``'default'``: Do not override, use the default :meth:`Op.L_op` result
`OpFromGraph`: Override with another `OpFromGraph`, should
accept inputs as the same order and types of ``inputs``,
``outputs`` and ``output_grads`` arguments as one would specify in
:meth:`Op.grad` method.
`callable`: Should take three args: ``inputs``, ``outputs`` and ``output_grads``.
Each argument is expected to be a list of :class:`Variable`.
Must return list of :class:`Variable`.
`NullType` instance: Treat as non-differentiable
`DisconnectedType` instance: Treat as disconnected gradient,
numerically gives zero
``list``: Each `OpFromGraph`/callable must return a single
:class:`Variable`. Each list element corresponds to gradient of
a specific input, length of list must be equal to number of inputs.
rop_overrides
One of ``{'default', OpFromGraph, callable, Variable}``.
Defaults to ``'default'``.
``'default'``: Do not override, use the default :meth:`Op.R_op` result
`OpFromGraph`: Override with another `OpFromGraph`, should
accept inputs as the same order and types of ``inputs`` and ``eval_points``
arguments as one would specify in :meth:`Op.R_op` method.
`callable`: Should take two args: ``inputs`` and ``eval_points``.
Each argument is expected to be a list of :class:`Variable`. Must
return list of :class:`Variable`.
`NullType` instance: Treat as non-differentiable `DisconnectedType`
instance: Treat as zero since `DisconnectedType` is not yet supported
in :meth:`Op.R_op`.
``list``:
Each :class:`OpFromGraph`/callable must return a single
:class:`Variable <pytensor.graph.basic.Variable>`. Each list element
corresponds to a specific output of :meth:`Op.R_op`, length of list
must be equal to number of outputs. connection_pattern If not
``None``, this will be used as the connection_pattern for this
:class:`Op`.
strict: bool, default False
If true, it raises when any variables needed to compute the inner graph
are not provided as explici inputs. This can only happen for graphs with
shared variables.
name
A name for debugging purposes.
kwargs
Check :func:`pytensor.function` for more arguments, only works when not
inline.
"""
ignore_unused_inputs = kwargs.get("on_unused_input", False) == "ignore"
if not ignore_unused_inputs and len(inputs) != len(set(inputs)):
var_counts = {var: inputs.count(var) for var in inputs}
duplicated_inputs = [var for var, count in var_counts.items() if count > 1]
raise ValueError(
f"There following variables were provided more than once as inputs to the OpFromGraph, resulting in an "
f"invalid graph: {duplicated_inputs}. Use dummy variables or var.copy() to distinguish "
f"variables when creating the OpFromGraph graph."
)
if not (isinstance(inputs, list) and isinstance(outputs, list)):
raise TypeError("Inputs and outputs must be lists")
for out in outputs:
if not isinstance(out, Variable):
raise TypeError(
f"Inputs and outputs must be Variable instances; got {out}"
)
if "updates" in kwargs or "givens" in kwargs:
raise NotImplementedError("Updates and givens are not supported")
self.is_inline = inline
self.fgraph, self.shared_inputs, _, _ = construct_nominal_fgraph(
inputs, outputs
)
if strict and self.shared_inputs:
raise ValueError(
"All variables needed to compute inner-graph must be provided as inputs under strict=True. "
f"The inner-graph implicitly depends on the following shared variables {self.shared_inputs}"
)
self.kwargs = kwargs
self.input_types = [inp.type for inp in inputs]
self.output_types = [out.type for out in outputs]
self.lop_overrides = lop_overrides
self.grad_overrides = grad_overrides
self.rop_overrides = rop_overrides
if lop_overrides != "default":
if grad_overrides != "default":
raise ValueError(
"lop_overrides and grad_overrides are mutually exclusive"
)
else:
self.set_lop_overrides(lop_overrides)
self._lop_type = "lop"
elif grad_overrides != "default":
self.set_lop_overrides(grad_overrides)
self._lop_type = "grad"
else:
self.set_lop_overrides("default")
self._lop_type = "lop"
self.set_rop_overrides(rop_overrides)
self._connection_pattern = connection_pattern
if name is not None:
assert isinstance(name, str), "name must be None or string object"
self.name = name
def __eq__(self, other):
# TODO: recognize a copy
return self is other
def __hash__(self):
# TODO: use internal variables in hash
return hash(type(self))
def __str__(self):
name = self.__class__.__name__ if self.name is None else self.name
is_inline = self.is_inline
return "{name}{{inline={is_inline}}}".format(**locals())
@config.change_flags(compute_test_value="off")
def _recompute_lop_op(self):
"""
converts self._lop_op from user supplied form to type(self) instance
"""
local_inputs = self.inner_inputs
local_outputs = self.inner_outputs
inp_len = len(local_inputs)
lop_op = self._lop_op
if isinstance(lop_op, OpFromGraph):
if self._lop_op_is_cached:
return
assert self._lop_type in ("lop", "grad"), (
self.LOP_TYPE_ERR_MSG % self._lop_type
)
if self._lop_type == "grad":
needed_ninps = inp_len + len(local_outputs)
ninps = len(lop_op.inner_inputs)
if needed_ninps != ninps:
raise ValueError(self.OV_INP_LEN_ERR_MSG % (needed_ninps, ninps))
# make a wrapper callable
def lop_op(inps, grads):
return self._lop_op(*(inps + grads))
elif self._lop_type == "lop":
# OfG can be directly used in L_op format
needed_ninps = inp_len + 2 * len(local_outputs)
ninps = len(lop_op.inner_inputs)
if needed_ninps != ninps:
raise ValueError(self.OV_INP_LEN_ERR_MSG % (needed_ninps, ninps))
self._lop_op_is_cached = True
self._lop_op_stypes_l = [None] * inp_len
self._lop_op.kwargs["on_unused_input"] = "ignore"
return
output_grads = [out_t() for out_t in self.output_types]
fn_grad = partial(
grad,
cost=None,
disconnected_inputs="ignore",
return_disconnected="Disconnected",
null_gradients="return",
known_grads=OrderedDict(zip(local_outputs, output_grads)),
)
assert self._lop_type in ("lop", "grad"), self.LOP_TYPE_ERR_MSG % self._lop_type
if self._lop_type == "lop":
callable_args = (local_inputs, local_outputs, output_grads)
elif self._lop_type == "grad":
callable_args = (local_inputs, output_grads)
# we need to convert _lop_op into an OfG instance
if lop_op == "default":
gdefaults_l = fn_grad(wrt=local_inputs)
all_grads_l, all_grads_ov_l = zip(
*[
OpFromGraph._filter_grad_var(grad, inp)
for grad, inp in zip(gdefaults_l, local_inputs)
]
)
all_grads_l = list(all_grads_l)
all_grads_ov_l = list(all_grads_ov_l)
elif isinstance(lop_op, Variable):
if isinstance(lop_op.type, DisconnectedType | NullType):
all_grads_l = [inp.zeros_like() for inp in local_inputs]
all_grads_ov_l = [lop_op.type() for _ in range(inp_len)]
else:
raise ValueError(self.STYPE_ERR_MSG % lop_op.type)
elif isinstance(lop_op, list):
goverrides_l = lop_op
if len(goverrides_l) != inp_len:
raise ValueError(
f"Need to override {int(inp_len)} gradients, got {len(goverrides_l)}",
goverrides_l,
)
# compute non-overriding downsteam grads from upstreams grads
# it's normal some input may be disconnected, thus the 'ignore'
wrt_l = [
lin for lin, gov in zip(local_inputs, goverrides_l) if gov == "default"
]
gdefaults = iter(fn_grad(wrt=wrt_l) if wrt_l else [])
# combine overriding gradients
all_grads_l = []
all_grads_ov_l = []
for inp, fn_gov in zip(local_inputs, goverrides_l):
if fn_gov == "default":
gnext, gnext_ov = OpFromGraph._filter_grad_var(next(gdefaults), inp)
all_grads_l.append(gnext)
all_grads_ov_l.append(gnext_ov)
elif isinstance(fn_gov, Variable):
if isinstance(fn_gov.type, DisconnectedType | NullType):
all_grads_l.append(inp.zeros_like())
all_grads_ov_l.append(fn_gov.type())
else:
raise ValueError(self.STYPE_ERR_MSG % fn_gov.type)
else:
if not callable(fn_gov):
raise TypeError(self.TYPE_ERR_MSG % fn_gov)
gov, gov_ov = OpFromGraph._filter_grad_var(
fn_gov(*callable_args), inp
)
all_grads_l.append(gov)
all_grads_ov_l.append(gov_ov)
else:
# callable case
if not callable(lop_op):
raise TypeError(self.TYPE_ERR_MSG % lop_op)
goverrides_l = lop_op(*callable_args)
if not isinstance(goverrides_l, list):
raise TypeError(
"Gradient/L_op overriding function should return a list, "
f'got "{type(goverrides_l)}"'
)
all_grads_l, all_grads_ov_l = zip(
*[
OpFromGraph._filter_grad_var(grad, inp)
for grad, inp in zip(goverrides_l, local_inputs)
]
)
if len(all_grads_l) != len(local_inputs):
raise ValueError(
"Gradient/L_op overriding function should return list of "
f"{int(inp_len)} outputs, got {len(all_grads_l)}"
)
all_grads_l = list(all_grads_l)
all_grads_ov_l = list(all_grads_ov_l)
self._lop_op = type(self)(
inputs=local_inputs + local_outputs + output_grads,
outputs=all_grads_l,
inline=self.is_inline,
name=(None if self.name is None else self.name + "_" + self._lop_type),
on_unused_input="ignore",
)
self._lop_op_stypes_l = all_grads_ov_l
self._lop_op_is_cached = True
self._lop_type = "lop"
@config.change_flags(compute_test_value="off")
def _recompute_rop_op(self):
"""
converts self._rop_op from user supplied form to type(self) instance
"""
local_inputs = self.inner_inputs
local_outputs = self.inner_outputs
out_len = len(local_outputs)
rop_op = self._rop_op
if isinstance(rop_op, OpFromGraph):
if not self._rop_op_is_cached:
self._rop_op_is_cached = True
self._rop_op_stypes_l = [None] * out_len
return
eval_points = [inp_t() for inp_t in self.input_types]
fn_rop = partial(Rop, wrt=local_inputs, eval_points=eval_points)
TYPE_ERR_MSG = (
"R_op overrides should be (single or list of)"
"OpFromGraph | 'default' | None | 0 | callable, got %s"
)
STYPE_ERR_MSG = (
"Overriding Variable instance can only have type"
" of DisconnectedType or NullType, got %s"
)
if rop_op == "default":
rdefaults_l = fn_rop(f=local_outputs)
all_rops_l, all_rops_ov_l = zip(
*[
OpFromGraph._filter_rop_var(rop, out)
for rop, out in zip(rdefaults_l, local_outputs)
]
)
all_rops_l = list(all_rops_l)
all_rops_ov_l = list(all_rops_ov_l)
elif isinstance(rop_op, Variable):
if isinstance(rop_op.type, NullType):
all_rops_l = [inp.zeros_like() for inp in local_inputs]
all_rops_ov_l = [rop_op.type() for _ in range(out_len)]
elif isinstance(rop_op.type, DisconnectedType):
all_rops_l = [inp.zeros_like() for inp in local_inputs]
all_rops_ov_l = [None] * out_len
else:
raise ValueError(STYPE_ERR_MSG % rop_op.type)
elif isinstance(rop_op, list):
roverrides_l = rop_op
if len(roverrides_l) != out_len:
raise ValueError(
f"Need to override {int(out_len)} Rop, got {len(roverrides_l)}",
roverrides_l,
)
# get outputs that does not have Rop override
odefaults_l = [
lo for lo, rov in zip(local_outputs, roverrides_l) if rov == "default"
]
rdefaults_l = fn_rop(f=odefaults_l)
rdefaults = iter(rdefaults_l if odefaults_l else [])
# combine overriding Rops
all_rops_l = []
all_rops_ov_l = []
for out, fn_rov in zip(local_outputs, roverrides_l):
if fn_rov == "default":
rnext, rnext_ov = OpFromGraph._filter_rop_var(next(rdefaults), out)
all_rops_l.append(rnext)
all_rops_ov_l.append(rnext_ov)
elif isinstance(fn_rov, Variable):
if isinstance(fn_rov.type, NullType):
all_rops_l.append(out.zeros_like())
all_rops_ov_l.append(fn_rov.type())
if isinstance(fn_rov.type, DisconnectedType):
all_rops_l.append(out.zeros_like())
all_rops_ov_l.append(None)
else:
raise ValueError(STYPE_ERR_MSG % fn_rov.type)
else:
if not callable(fn_rov):
raise TypeError(TYPE_ERR_MSG % fn_rov)
rov, rov_ov = OpFromGraph._filter_rop_var(
fn_rov(local_inputs, eval_points), out
)
all_rops_l.append(rov)
all_rops_ov_l.append(rov_ov)
else:
if not callable(rop_op):
raise TypeError(TYPE_ERR_MSG % rop_op)
roverrides_l = rop_op(local_inputs, eval_points)
if not isinstance(roverrides_l, list):
raise TypeError(
"Rop overriding function should return a list, "
f'got "{type(roverrides_l)}"'
)
all_rops_l, all_rops_ov_l = zip(
*[
OpFromGraph._filter_rop_var(rop, out)
for rop, out in zip(roverrides_l, local_outputs)
]
)
if len(all_rops_l) != out_len:
raise ValueError(
(
f"Rop overriding function {self._rop_op} should return list of "
f"{int(out_len)} outputs, got {len(all_rops_l)}",
),
rop_op,
)
all_rops_l = list(all_rops_l)
all_rops_ov_l = list(all_rops_ov_l)
self._rop_op = type(self)(
inputs=local_inputs + eval_points,
outputs=all_rops_l,
inline=self.is_inline,
name=(None if self.name is None else self.name + "_rop"),
on_unused_input="ignore",
)
self._rop_op_stypes_l = all_rops_ov_l
self._rop_op_is_cached = True
def get_lop_op(self):
if not self._lop_op_is_cached:
self._recompute_lop_op()
return self._lop_op
def get_rop_op(self):
if not self._rop_op_is_cached:
self._recompute_rop_op()
return self._rop_op
def set_grad_overrides(self, grad_overrides):
"""
Set gradient overrides.
This will completely remove any previously set L_op/gradient overrides
"""
self._lop_op = grad_overrides
self._lop_op_is_cached = False
self._lop_type = "grad"
self._lop_is_default = grad_overrides == "default"
def set_lop_overrides(self, lop_overrides):
"""
Set L_op overrides
This will completely remove any previously set L_op/gradient overrides
"""
self._lop_op = lop_overrides
self._lop_op_is_cached = False
self._lop_type = "lop"
self._lop_is_default = lop_overrides == "default"
def set_rop_overrides(self, rop_overrides):
"""
Set R_op overrides
This will completely remove any previously set R_op overrides
"""
self._rop_op = rop_overrides
self._rop_op_is_cached = False
self._rop_is_default = rop_overrides == "default"
def L_op(self, inputs, outputs, output_grads):
if not self._lop_op_is_cached:
self._recompute_lop_op()
inps = list(inputs) + list(outputs) + list(output_grads)
ret_ofg_l = self._lop_op(*inps, return_list=True)
ret_l = [
ret_ofg if ov is None else ov
for ret_ofg, ov in zip(ret_ofg_l, self._lop_op_stypes_l)
]
return ret_l
def R_op(self, inputs, eval_points):
if not self._rop_op_is_cached:
self._recompute_rop_op()
ret_ofg_l = self._rop_op(*(list(inputs) + list(eval_points)), return_list=True)
ret_l = [
ret_ofg if ov is None else ov
for ret_ofg, ov in zip(ret_ofg_l, self._rop_op_stypes_l)
]
return ret_l
def __call__(self, *inputs, **kwargs):
# The user interface doesn't expect the shared variable inputs of the
# inner-graph, but, since `Op.make_node` does (and `Op.__call__`
# dispatches to `Op.make_node`), we need to compensate here
num_expected_inps = len(self.inner_inputs) - len(self.shared_inputs)
if len(inputs) == num_expected_inps:
actual_inputs = inputs + tuple(self.shared_inputs)
return super().__call__(*actual_inputs, **kwargs)
elif len(inputs) == len(self.inner_inputs):
return super().__call__(*inputs, **kwargs)
else:
raise ValueError(f"Expected at least {num_expected_inps} input(s)")
def make_node(self, *inputs):
# The `inputs` received here should correspond to the inputs in the
# `Apply` nodes we produce below
if len(inputs) != len(self.inner_inputs):
raise ValueError(f"Expected {len(self.inner_inputs)} input(s)")
num_expected_inps = len(self.inner_inputs) - len(self.shared_inputs)
non_shared_inputs = inputs[:num_expected_inps]
non_shared_inputs = [
inp_t.filter_variable(inp)
for inp, inp_t in zip(non_shared_inputs, self.input_types)
]
new_shared_inputs = inputs[num_expected_inps:]
inner_and_input_shareds = list(zip(self.shared_inputs, new_shared_inputs))
if not all(inp_s == inn_s for inn_s, inp_s in inner_and_input_shareds):
# The shared variables are not equal to the original shared
# variables, so we construct a new `Op` that uses the new shared
# variables instead.
replace = dict(
zip(self.inner_inputs[num_expected_inps:], new_shared_inputs)
)
# If the new shared variables are inconsistent with the inner-graph,
# such errors should arise in this step
new_inner_outputs = clone_replace(
self.inner_outputs, replace=replace, copy_inputs_over=True
)
# It's possible that the new shared variable inputs aren't actually
# shared variables. When they aren't we need to add them as new
# inputs.
unshared_inputs = [
inp for inp in new_shared_inputs if not isinstance(inp, SharedVariable)
]
new_inner_inputs = self.inner_inputs[:num_expected_inps] + unshared_inputs
new_op = type(self)(
inputs=new_inner_inputs,
outputs=new_inner_outputs,
inline=self.is_inline,
lop_overrides=self.lop_overrides,
grad_overrides=self.grad_overrides,
rop_overrides=self.rop_overrides,
connection_pattern=self._connection_pattern,
name=self.name,
**self.kwargs,
)
new_inputs = (
list(non_shared_inputs) + unshared_inputs + new_op.shared_inputs
)
else:
new_op = self
new_inputs = list(non_shared_inputs) + new_op.shared_inputs
apply_node = Apply(
new_op,
new_inputs,
[type() for type in new_op.output_types],
)
return apply_node
def connection_pattern(self, node):
"""
Return connection pattern of subfgraph defined by inputs and outputs.
"""
if self._connection_pattern is not None:
return self._connection_pattern
inp_len = len(self.inner_inputs)
out_len = len(self.inner_outputs)
cpmat_self = io_connection_pattern(self.inner_inputs, self.inner_outputs)
lop_op = self.get_lop_op()
cpmat_grad = io_connection_pattern(
lop_op.inner_inputs[inp_len:], lop_op.inner_outputs
)
# cpmat_self |= cpmat_grad.T
# cpmat_self &= out_is_disconnected
for i, t in enumerate(self._lop_op_stypes_l):
if t is not None:
if isinstance(t.type, DisconnectedType):
for o in range(out_len):
cpmat_self[i][o] = False
for o in range(out_len):
cpmat_self[i][o] |= cpmat_grad[o][i]
# TODO in case DisconnectedType is implemented for R_op,
# self._rop_op_stypes_l self._rop_op should considered for
# connection_pattern
return list(map(list, cpmat_self))
def infer_shape(self, fgraph, node, shapes):
# TODO: Use `fgraph.shape_feature` to do this instead.
out_shapes = infer_shape(self.inner_outputs, self.inner_inputs, shapes)
# Clone the output shape so that shape are computed from outer inputs.
# Note:
# Here we could do it more simply like:
# `ret = [pytensor.clone_replace(shp, replace=repl) for shp in out_shp]`
# But doing it multiple time could duplicate common subgraph between
# each shape call. PyTensor optimizer will clean this up later, but this
# will make extra work for the optimizer.
repl = dict(zip(self.inner_inputs, node.inputs))
clone_out_shapes = [s for s in out_shapes if isinstance(s, tuple)]
cloned = clone_replace(sum(clone_out_shapes, ()), replace=repl)
ret = []
used = 0
for i, out_shape in enumerate(out_shapes):
if out_shape is None:
ret.append(None)
else:
nb = len(out_shape)
ret.append(cloned[used : used + nb])
used += nb
return ret
@property
def fn(self):
"""Lazily compile the inner function graph."""
if getattr(self, "_fn", None) is not None:
return self._fn
self._fn = function(self.inner_inputs, self.inner_outputs, **self.kwargs)
self._fn.trust_input = True
return self._fn
@property
def inner_inputs(self):
return self.fgraph.inputs
@property
def inner_outputs(self):
return self.fgraph.outputs
def clone(self):
res = copy(self)
res.fgraph = res.fgraph.clone()
return res
def perform(self, node, inputs, outputs):
variables = self.fn(*inputs)
assert len(variables) == len(outputs)
for output, variable in zip(outputs, variables):
output[0] = variable
@node_rewriter([OpFromGraph])
def inline_ofg_expansion(fgraph, node):
"""
This optimization expands internal graph of OpFromGraph.
Only performed if node.op.is_inline == True
Doing so can improve optimization at the cost of compilation speed.
"""
op = node.op
if not isinstance(op, OpFromGraph):
return False
if not op.is_inline:
return False
return clone_replace(op.inner_outputs, dict(zip(op.inner_inputs, node.inputs)))
# We want to run this before the first merge optimizer
# and before the first scan optimizer.
optdb.register(
"inline_ofg_expansion",
in2out(inline_ofg_expansion),
"fast_compile",
"fast_run",
position=-0.01,
)