/
pure_function.py
565 lines (446 loc) · 15.1 KB
/
pure_function.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
import torch
import inspect
from typing import Callable, List, Tuple, Union, Sequence
from deepchem.utils.attribute_utils import set_attr, del_attr
from deepchem.utils.differentiation_utils import EditableModule
from deepchem.utils.misc_utils import Uniquifier
from contextlib import contextmanager
from abc import abstractmethod
class PureFunction(object):
"""
PureFunction class wraps methods to make it stateless and expose the pure
function to take inputs of the original inputs (`params`) and the object's
states (`objparams`).
For functions, this class only acts as a thin wrapper.
Restore stack stores list of (objparams, identical) everytime the objparams
are set, it will store the old objparams and indication if the old and new
objparams are identical.
For Using this Class we first need to implement `_get_all_obj_params_init`
and `_set_all_obj_params`.
Examples
--------
>>> class WrapperFunction(PureFunction):
... def _get_all_obj_params_init(self):
... return []
... def _set_all_obj_params(self, objparams):
... pass
>>> def fcn(x, y):
... return x + y
>>> pfunc = WrapperFunction(fcn)
>>> pfunc(1, 2)
3
"""
def __init__(self, fcntocall: Callable):
"""Initialize the PureFunction.
Parameters
----------
fcntocall: Callable
The function to be wrapped
"""
self._state_change_allowed = True
self._allobjparams = self._get_all_obj_params_init()
self._uniq = Uniquifier(self._allobjparams)
self._cur_objparams = self._uniq.get_unique_objs()
self._fcntocall = fcntocall
self._restore_stack: List[Tuple[List, bool]] = []
def __call__(self, *params):
"""Call the wrapped function with the current object parameters and
the input parameters.
Parameters
----------
params: tuple
The input parameters of the wrapped function
Returns
-------
Any
The output of the wrapped function
"""
return self._fcntocall(*params)
@abstractmethod
def _get_all_obj_params_init(self) -> List:
"""Get the initial object parameters.
Returns
-------
List
The initial object parameters
"""
pass
@abstractmethod
def _set_all_obj_params(self, allobjparams: List):
"""Set the object parameters.
Parameters
----------
allobjparams: List
The object parameters to be set
"""
pass
def objparams(self) -> List:
"""Get the current object parameters.
Returns
-------
List
The current object parameters
"""
return self._cur_objparams
def set_objparams(self, objparams: List):
"""Set the object parameters.
Parameters
----------
objparams: List
The object parameters to be set
TODO: check if identical with current object parameters
"""
identical = _check_identical_objs(objparams, self._cur_objparams)
self._restore_stack.append((self._cur_objparams, identical))
if not identical:
allobjparams = self._uniq.map_unique_objs(objparams)
self._set_all_obj_params(allobjparams)
self._cur_objparams = list(objparams)
def restore_objparams(self):
"""Restore the object parameters to the previous state."""
old_objparams, identical = self._restore_stack.pop(-1)
if not identical:
allobjparams = self._uniq.map_unique_objs(old_objparams)
self._set_all_obj_params(allobjparams)
self._cur_objparams = old_objparams
@contextmanager
def useobjparams(self, objparams: List):
"""Context manager to temporarily set the object parameters.
Parameters
----------
objparams: List
The object parameters to be set temporarily
"""
if not self._state_change_allowed:
raise RuntimeError("The state change is disabled")
try:
self.set_objparams(objparams)
yield
finally:
self.restore_objparams()
@contextmanager
def disable_state_change(self):
"""Context manager to temporarily disable the state change."""
try:
prev_status = self._state_change_allowed
self._state_change_allowed = False
yield
finally:
self._state_change_allowed = prev_status
class FunctionPureFunction(PureFunction):
"""Implementation of PureFunction for functions.
It just acts as a thin wrapper for the function.
Examples
--------
>>> def fcn(x, y):
... return x + y
>>> pfunc = FunctionPureFunction(fcn)
>>> pfunc(1, 2)
3
"""
def _get_all_obj_params_init(self) -> List:
"""Get the initial object parameters.
Returns
-------
List
The initial object parameters
"""
return []
def _set_all_obj_params(self, objparams: List):
"""Set the object parameters.
Parameters
----------
objparams: List
The object parameters to be set
"""
pass
class EditableModulePureFunction(PureFunction):
"""Implementation of PureFunction for EditableModule.
Examples
--------
>>> import torch
>>> from deepchem.utils.differentiation_utils import EditableModule, get_pure_function
>>> class A(EditableModule):
... def __init__(self, a):
... self.b = a*a
... def mult(self, x):
... return self.b * x
... def getparamnames(self, methodname, prefix=""):
... if methodname == "mult":
... return [prefix+"b"]
... else:
... raise KeyError()
>>> B = A(4)
>>> m = get_pure_function(B.mult)
>>> m.set_objparams([3])
>>> m(2)
6
"""
def __init__(self, obj: EditableModule, method: Callable):
"""Initialize the EditableModulePureFunction.
Parameters
----------
obj: EditableModule
The object to be wrapped
method: Callable
The method to be wrapped
"""
self.obj = obj
self.method = method
super().__init__(method)
def _get_all_obj_params_init(self) -> List:
"""Get the initial object parameters.
Returns
-------
List
The initial object parameters
"""
return list(self.obj.getparams(self.method.__name__))
def _set_all_obj_params(self, allobjparams: List):
"""Set the object parameters.
Parameters
----------
allobjparams: List
The object parameters to be set
"""
self.obj.setparams(self.method.__name__, *allobjparams)
class TorchNNPureFunction(PureFunction):
"""Implementation of PureFunction for torch.nn.Module.
Examples
--------
>>> import torch
>>> from deepchem.utils.differentiation_utils import get_pure_function
>>> class A(torch.nn.Module):
... def __init__(self, a):
... super().__init__()
... self.b = torch.nn.Parameter(torch.tensor(a*a))
... def forward(self, x):
... return self.b * x
>>> B = A(4.)
>>> m = get_pure_function(B.forward)
>>> m.set_objparams([3.])
>>> m(2)
6.0
"""
def __init__(self, obj: torch.nn.Module, method: Callable):
"""Initialize the TorchNNPureFunction.
Parameters
----------
obj: torch.nn.Module
Object to be wrapped
method: Callable
Method to be wrapped
"""
self.obj = obj
self.method = method
super().__init__(method)
def _get_all_obj_params_init(self) -> List:
"""get the tensors in the torch.nn.Module to be used as params
Returns
-------
List
The initial object parameters
"""
named_params = list(self.obj.named_parameters())
if len(named_params) == 0:
paramnames: List[str] = []
obj_params: List[Union[torch.Tensor, torch.nn.Parameter]] = []
else:
paramnames_temp, obj_params_temp = zip(*named_params)
paramnames = list(paramnames_temp)
obj_params = list(obj_params_temp)
self.names = paramnames
return obj_params
def _set_all_obj_params(self, objparams: List):
"""Set the object parameters.
Parameters
----------
objparams: List
The object parameters to be set
"""
for (name, param) in zip(self.names, objparams):
del_attr(
self.obj, name
) # delete required in case the param is not a torch.nn.Parameter
set_attr(self.obj, name, param)
def _check_identical_objs(objs1: List, objs2: List) -> bool:
"""Check if the two lists of objects are identical.
Examples
--------
>>> l1 = [2, 2, 3]
>>> l2 = [1, 2, 3]
>>> _check_identical_objs(l1, l2)
False
Parameters
----------
objs1: List
The first list of objects
objs2: List
The second list of objects
Returns
-------
bool
True if the two lists of objects are identical, False otherwise
"""
for obj1, obj2 in zip(objs1, objs2):
if id(obj1) != id(obj2):
return False
return True
class SingleSiblingPureFunction(PureFunction):
"""Implementation of PureFunction for a sibling method
A sibling method is a method that is virtually belong to the same object,
but behaves differently.
Changing the state of the decorated function will also change the state of
``pfunc`` and its other siblings.
"""
def __init__(self, fcn: Callable, fcntocall: Callable):
"""Initialize the SingleSiblingPureFunction.
Parameters
----------
fcn: Callable
The sibling method to be wrapped
fcntocall: Callable
The method to be wrapped
"""
self.pfunc = get_pure_function(fcn)
super().__init__(fcntocall)
def _get_all_obj_params_init(self) -> List:
"""Get the initial object parameters.
Returns
-------
List
The initial object parameters
"""
return self.pfunc._get_all_obj_params_init()
def _set_all_obj_params(self, allobjparams: List):
"""Set the object parameters.
Parameters
----------
allobjparams: List
The object parameters to be set
"""
self.pfunc._set_all_obj_params(allobjparams)
class MultiSiblingPureFunction(PureFunction):
"""Implementation of PureFunction for multiple sibling methods
A sibling method is a method that is virtually belong to the same object,
but behaves differently.
Changing the state of the decorated function will also change the state of
``pfunc`` and its other siblings.
"""
def __init__(self, fcns: Sequence[Callable], fcntocall: Callable):
"""Initialize the MultiSiblingPureFunction.
Parameters
----------
fcns: Sequence[Callable]
The sibling methods to be wrapped
fcntocall: Callable
The method to be wrapped
"""
self.pfuncs = [get_pure_function(fcn) for fcn in fcns]
self.npfuncs = len(self.pfuncs)
super().__init__(fcntocall)
def _get_all_obj_params_init(self) -> List:
"""Get the initial object parameters.
Returns
-------
List
The initial object parameters
"""
res: List[Union[torch.Tensor, torch.nn.Parameter]] = []
self.cumsum_idx = [0] * (self.npfuncs + 1)
for i, pfunc in enumerate(self.pfuncs):
objparams = pfunc._get_all_obj_params_init()
res = res + objparams
self.cumsum_idx[i + 1] = self.cumsum_idx[i] + len(objparams)
return res
def _set_all_obj_params(self, allobjparams: List):
"""Set the object parameters.
Parameters
----------
allobjparams: List
The object parameters to be set
"""
for i, pfunc in enumerate(self.pfuncs):
pfunc._set_all_obj_params(
allobjparams[self.cumsum_idx[i]:self.cumsum_idx[i + 1]])
def get_pure_function(fcn) -> PureFunction:
"""Get the pure function form of the function or method ``fcn``.
Examples
--------
>>> import torch
>>> from deepchem.utils.differentiation_utils import get_pure_function
>>> def fcn(x, y):
... return x + y
>>> pfunc = get_pure_function(fcn)
>>> pfunc(1, 2)
3
Parameters
----------
fcn: function or method
Function or method to be converted into a ``PureFunction`` by exposing
the hidden parameters affecting its outputs.
Returns
-------
PureFunction
The pure function wrapper
"""
errmsg = "The input function must be a function, a method of " \
"torch.nn.Module, a method of xitorch.EditableModule, or a sibling method"
if isinstance(fcn, PureFunction):
return fcn
elif inspect.isfunction(fcn) or isinstance(fcn, torch.jit.ScriptFunction):
return FunctionPureFunction(fcn)
# if it is a method from an object, unroll the parameters and add
# the object's parameters as well
elif inspect.ismethod(fcn) or hasattr(fcn, "__call__"):
if inspect.ismethod(fcn):
obj = fcn.__self__
else:
obj = fcn
fcn = fcn.__call__
if isinstance(obj, EditableModule):
return EditableModulePureFunction(obj, fcn)
elif isinstance(obj, torch.nn.Module):
return TorchNNPureFunction(obj, fcn)
else:
raise RuntimeError(errmsg)
else:
raise RuntimeError(errmsg)
def make_sibling(*pfuncs) -> Callable[[Callable], PureFunction]:
"""
Used as a decor to mark the decorated function as a sibling method of the
input ``pfunc``.
Sibling method is a method that is virtually belong to the same object, but
behaves differently.
Changing the state of the decorated function will also change the state of
``pfunc`` and its other siblings.
Examples
--------
>>> import torch
>>> from deepchem.utils.differentiation_utils import make_sibling
>>> def fcn1(x, y):
... return x + y
>>> def fcn2(x, y):
... return x - y
>>> pfunc1 = get_pure_function(fcn1)
>>> pfunc2 = get_pure_function(fcn2)
>>> @make_sibling(pfunc1)
... def fcn3(x, y):
... return x * y
>>> pfunc3(1, 2)
2
Parameters
----------
pfuncs: List[Callable]
The sibling methods to be wrapped
Returns
-------
Callable[[Callable], PureFunction]
The decorator function
"""
if len(pfuncs) == 0:
raise TypeError("At least 1 function is required as the argument")
elif len(pfuncs) == 1:
return lambda fcn: SingleSiblingPureFunction(pfuncs[0], fcntocall=fcn)
else:
return lambda fcn: MultiSiblingPureFunction(pfuncs, fcntocall=fcn)