/
fisher_factors.py
1830 lines (1460 loc) · 65.2 KB
/
fisher_factors.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
996
997
998
999
1000
# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
#
# 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.
# ==============================================================================
"""FisherFactor definitions."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import abc
import contextlib
import numpy as np
import six
from tensorflow.contrib.kfac.python.ops import linear_operator as lo
from tensorflow.contrib.kfac.python.ops import utils
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import ops as tf_ops
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import control_flow_ops
from tensorflow.python.ops import init_ops
from tensorflow.python.ops import linalg_ops
from tensorflow.python.ops import math_ops
from tensorflow.python.ops import random_ops
from tensorflow.python.ops import special_math_ops
from tensorflow.python.ops import variable_scope
from tensorflow.python.ops import variables
from tensorflow.python.training import moving_averages
from tensorflow.python.util import nest
# Whether to initialize covariance estimators at a zero matrix (or the identity
# matrix).
INIT_COVARIANCES_AT_ZERO = True
# Whether to zero-debias the moving averages.
ZERO_DEBIAS = True
# Whether to initialize inverse (and other such matrices computed from the cov
# matrices) to the zero matrix (or the identity matrix).
INIT_INVERSES_AT_ZERO = True
# When the number of inverses requested from a FisherFactor exceeds this value,
# the inverses are computed using an eigenvalue decomposition.
EIGENVALUE_DECOMPOSITION_THRESHOLD = 2
# Numerical eigenvalues computed from covariance matrix estimates are clipped to
# be at least as large as this value before they are used to compute inverses or
# matrix powers. Must be nonnegative.
EIGENVALUE_CLIPPING_THRESHOLD = 0.0
# Used to subsample the flattened extracted image patches. The number of
# outer products per row of the covariance matrix should not exceed this
# value. This parameter is used only if `_SUB_SAMPLE_OUTER_PRODUCTS` is True.
_MAX_NUM_OUTER_PRODUCTS_PER_COV_ROW = 1
# Used to subsample the inputs passed to the extract image patches. The batch
# size of number of inputs to extract image patches is multiplied by this
# factor. This parameter is used only if `_SUB_SAMPLE_INPUTS` is True.
_INPUTS_TO_EXTRACT_PATCHES_FACTOR = 0.5
# If True, then subsamples the tensor passed to compute the covaraince matrix.
_SUB_SAMPLE_OUTER_PRODUCTS = False
# If True, then subsamples the tensor passed to compute the covaraince matrix.
_SUB_SAMPLE_INPUTS = False
# TOWER_STRATEGY can be one of "concat" or "separate". If "concat", the data
# passed to the factors from the blocks will be concatenated across towers
# (lazilly via PartitionedTensor objects). Otherwise a tuple of tensors over
# towers will be passed in, and the factors will iterate over this and do the
# cov computations separately for each one, averaging the results together.
TOWER_STRATEGY = "concat"
def set_global_constants(init_covariances_at_zero=None,
zero_debias=None,
init_inverses_at_zero=None,
eigenvalue_decomposition_threshold=None,
eigenvalue_clipping_threshold=None,
max_num_outer_products_per_cov_row=None,
sub_sample_outer_products=None,
inputs_to_extract_patches_factor=None,
sub_sample_inputs=None,
tower_strategy=None):
"""Sets various global constants used by the classes in this module."""
global INIT_COVARIANCES_AT_ZERO
global ZERO_DEBIAS
global INIT_INVERSES_AT_ZERO
global EIGENVALUE_DECOMPOSITION_THRESHOLD
global EIGENVALUE_CLIPPING_THRESHOLD
global _MAX_NUM_OUTER_PRODUCTS_PER_COV_ROW
global _SUB_SAMPLE_OUTER_PRODUCTS
global _INPUTS_TO_EXTRACT_PATCHES_FACTOR
global _SUB_SAMPLE_INPUTS
global TOWER_STRATEGY
if init_covariances_at_zero is not None:
INIT_COVARIANCES_AT_ZERO = init_covariances_at_zero
if zero_debias is not None:
ZERO_DEBIAS = zero_debias
if init_inverses_at_zero is not None:
INIT_INVERSES_AT_ZERO = init_inverses_at_zero
if eigenvalue_decomposition_threshold is not None:
EIGENVALUE_DECOMPOSITION_THRESHOLD = eigenvalue_decomposition_threshold
if eigenvalue_clipping_threshold is not None:
EIGENVALUE_CLIPPING_THRESHOLD = eigenvalue_clipping_threshold
if max_num_outer_products_per_cov_row is not None:
_MAX_NUM_OUTER_PRODUCTS_PER_COV_ROW = max_num_outer_products_per_cov_row
if sub_sample_outer_products is not None:
_SUB_SAMPLE_OUTER_PRODUCTS = sub_sample_outer_products
if inputs_to_extract_patches_factor is not None:
_INPUTS_TO_EXTRACT_PATCHES_FACTOR = inputs_to_extract_patches_factor
if sub_sample_inputs is not None:
_SUB_SAMPLE_INPUTS = sub_sample_inputs
if tower_strategy is not None:
TOWER_STRATEGY = tower_strategy
def inverse_initializer(shape, dtype, partition_info=None): # pylint: disable=unused-argument
if INIT_INVERSES_AT_ZERO:
return array_ops.zeros(shape, dtype=dtype)
return linalg_ops.eye(num_rows=shape[0], dtype=dtype)
def covariance_initializer(shape, dtype, partition_info=None): # pylint: disable=unused-argument
if INIT_COVARIANCES_AT_ZERO:
return array_ops.zeros(shape, dtype=dtype)
return linalg_ops.eye(num_rows=shape[0], dtype=dtype)
def diagonal_covariance_initializer(shape, dtype, partition_info=None): # pylint: disable=unused-argument
if INIT_COVARIANCES_AT_ZERO:
return array_ops.zeros(shape, dtype=dtype)
return array_ops.ones(shape, dtype=dtype)
@contextlib.contextmanager
def place_on_device(device):
if device is not None and len(device):
with tf_ops.device(device):
yield
else:
yield
def compute_cov(tensor, tensor_right=None, normalizer=None):
"""Compute the empirical second moment of the rows of a 2D Tensor.
This function is meant to be applied to random matrices for which the true row
mean is zero, so that the true second moment equals the true covariance.
Args:
tensor: A 2D Tensor.
tensor_right: An optional 2D Tensor. If provided, this function computes
the matrix product tensor^T * tensor_right instead of tensor^T * tensor.
normalizer: optional scalar for the estimator (by default, the normalizer is
the number of rows of tensor).
Returns:
A square 2D Tensor with as many rows/cols as the number of input columns.
"""
if normalizer is None:
normalizer = array_ops.shape(tensor)[0]
if tensor_right is None:
cov = (
math_ops.matmul(tensor, tensor, transpose_a=True) / math_ops.cast(
normalizer, tensor.dtype))
return (cov + array_ops.transpose(cov)) / math_ops.cast(2.0, cov.dtype)
else:
return (math_ops.matmul(tensor, tensor_right, transpose_a=True) /
math_ops.cast(normalizer, tensor.dtype))
def append_homog(tensor):
"""Appends a homogeneous coordinate to the last dimension of a Tensor.
Args:
tensor: A Tensor.
Returns:
A Tensor identical to the input but one larger in the last dimension. The
new entries are filled with ones.
"""
rank = len(tensor.shape.as_list())
shape = array_ops.concat([array_ops.shape(tensor)[:-1], [1]], axis=0)
ones = array_ops.ones(shape, dtype=tensor.dtype)
return array_ops.concat([tensor, ones], axis=rank - 1)
def scope_string_from_params(params):
"""Builds a variable scope string name from the given parameters.
Supported parameters are:
* tensors
* booleans
* ints
* strings
* depth-1 tuples/lists of ints
* any depth tuples/lists of tensors
Other parameter types will throw an error.
Args:
params: A parameter or list of parameters.
Returns:
A string to use for the variable scope.
Raises:
ValueError: if params includes an unsupported type.
"""
params = params if isinstance(params, (tuple, list)) else (params,)
name_parts = []
for param in params:
if param is None:
name_parts.append("None")
elif isinstance(param, (tuple, list)):
if all([isinstance(p, int) for p in param]):
name_parts.append("-".join([str(p) for p in param]))
else:
name_parts.append(scope_string_from_name(param))
elif isinstance(param, (str, int, bool)):
name_parts.append(str(param))
elif isinstance(param, (tf_ops.Tensor, variables.Variable)):
name_parts.append(scope_string_from_name(param))
elif isinstance(param, utils.PartitionedTensor):
name_parts.append(scope_string_from_name(param.tensors))
else:
raise ValueError("Encountered an unsupported param type {}".format(
type(param)))
return "_".join(name_parts)
def scope_string_from_name(tensor):
if isinstance(tensor, (tuple, list)):
return "__".join([scope_string_from_name(t) for t in tensor])
# "gradients/add_4_grad/Reshape:0" -> "gradients_add_4_grad_Reshape"
return tensor.name.split(":")[0].replace("/", "_")
def scalar_or_tensor_to_string(val):
return repr(val) if np.isscalar(val) else scope_string_from_name(val)
def list_to_string(lst):
return "_".join(val if isinstance(val, six.string_types)
else scalar_or_tensor_to_string(val) for val in lst)
def graph_func_to_id(func):
"""Returns a hashable object that represents func's computation."""
# TODO(b/74201126): replace with Topohash of func's output
return func.func_id
def graph_func_to_string(func):
# TODO(b/74201126): replace with Topohash of func's output
return list_to_string(func.func_id)
def _subsample_for_cov_computation(array, name=None):
"""Subsamples the first dimension of the array.
`array`(A) is a tensor of shape `[batch_size, dim_2]`. Then the covariance
matrix(A^TA) is of shape `dim_2 ** 2`. Subsample only if the number of outer
products per row of the covariance matrix is greater than
`_MAX_NUM_OUTER_PRODUCTS_PER_COV_ROW`.
Args:
array: Tensor, of shape `[batch_size, dim_2]`.
name: `string`, Default(None)
Returns:
A tensor of shape `[max_samples, dim_2]`.
Raises:
ValueError: If array's is not matrix-shaped.
ValueError: If array's batch_size cannot be inferred.
"""
with tf_ops.name_scope(name, "subsample", [array]):
array = tf_ops.convert_to_tensor(array)
if len(array.shape) != 2:
raise ValueError("Input param array must be a matrix.")
batch_size = array.shape.as_list()[0]
if batch_size is None:
raise ValueError("Unable to get batch_size from input param array.")
num_cov_rows = array.shape.as_list()[-1]
max_batch_size = int(_MAX_NUM_OUTER_PRODUCTS_PER_COV_ROW * num_cov_rows)
if batch_size <= max_batch_size:
return array
return _random_tensor_gather(array, max_batch_size)
def _random_tensor_gather(array, max_size):
"""Generates a random set of indices and gathers the value at the indcices.
Args:
array: Tensor, of shape `[batch_size, dim_2]`.
max_size: int, Number of indices to sample.
Returns:
A tensor of shape `[max_size, ...]`.
"""
batch_size = array.shape.as_list()[0]
indices = random_ops.random_shuffle(math_ops.range(0, batch_size))[:max_size]
return array_ops.gather(array, indices)
@six.add_metaclass(abc.ABCMeta)
class FisherFactor(object):
"""Base class for objects modeling factors of approximate Fisher blocks.
A FisherFactor represents part of an approximate Fisher Information matrix.
For example, one approximation to the Fisher uses the Kronecker product of two
FisherFactors A and B, F = kron(A, B). FisherFactors are composed with
FisherBlocks to construct a block-diagonal approximation to the full Fisher.
FisherFactors are backed by a single, non-trainable variable that is updated
by running FisherFactor.make_covariance_update_op(). The shape and type of
this variable is implementation specific.
Note that for blocks that aren't based on approximations, a 'factor' can
be the entire block itself, as is the case for the diagonal and full
representations.
"""
def __init__(self):
self._cov = None
@abc.abstractproperty
def _var_scope(self):
"""Variable scope for this FisherFactor instance.
Returns:
string that unique identifies this FisherFactor instance.
"""
pass
@property
def name(self):
return self._var_scope
@abc.abstractproperty
def _cov_shape(self):
"""The shape of the variable backing this FisherFactor."""
pass
@abc.abstractproperty
def _num_sources(self):
"""The number of things to sum over when updating covariance variable.
The default make_covariance_update_op function will call _compute_new_cov
with indices ranging from 0 to _num_sources-1. The typical situation is
where the factor wants to sum the statistics it computes over multiple
backpropped "gradients" (typically passed in via "tensors" or
"outputs_grads" arguments).
"""
pass
@abc.abstractproperty
def _num_towers(self):
pass
@abc.abstractproperty
def _dtype(self):
"""dtype for variable backing this factor."""
pass
@property
def _cov_initializer(self):
"""Function for initializing covariance variable."""
return covariance_initializer
def instantiate_cov_variables(self):
"""Makes the internal cov variable(s)."""
assert self._cov is None
with variable_scope.variable_scope(self._var_scope):
self._cov = variable_scope.get_variable(
"cov",
initializer=self._cov_initializer,
shape=self._cov_shape,
trainable=False,
dtype=self._dtype)
@abc.abstractmethod
def _compute_new_cov(self, source, tower):
"""Computes minibatch-estimated covariance for a single source.
Args:
source: int in [0, self._num_sources). Which source to use when computing
the cov update.
tower: int in [0, self._num_towers). Which tower to use when computing
the cov update.
Returns:
Tensor of same shape as self.get_cov().
"""
pass
def make_covariance_update_op(self, ema_decay):
"""Constructs and returns the covariance update Op.
Args:
ema_decay: The exponential moving average decay (float or Tensor).
Returns:
An Op for updating the covariance Variable referenced by _cov.
"""
new_cov_contribs = []
for source in range(self._num_sources):
for tower in range(self._num_towers):
device = (self._get_data_device(tower)
if TOWER_STRATEGY == "separate" else None)
with place_on_device(device):
new_cov_contribs.append(self._compute_new_cov(source, tower))
new_cov = math_ops.add_n(new_cov_contribs) / float(self._num_towers)
# Compute average of 'new_cov' across all TPU cores. On a TPU, each
# instance of 'new_cov' will be based on a different minibatch. This ensures
# that by the end of assign_moving_average(), all TPU cores see the same
# value for self._cov.
#
# Other implementations of make_covariance_update_op() that accumulate
# statistics in other variables should mimic this behavior.
if utils.on_tpu():
new_cov = utils.cross_replica_mean(new_cov)
return moving_averages.assign_moving_average(
self._cov, new_cov, ema_decay, zero_debias=ZERO_DEBIAS)
@abc.abstractmethod
def _get_data_device(self, tower):
pass
@abc.abstractmethod
def instantiate_inv_variables(self):
"""Makes the internal "inverse" variable(s)."""
pass
@abc.abstractmethod
def make_inverse_update_ops(self):
"""Create and return update ops corresponding to registered computations."""
pass
def get_cov(self):
return self._cov
@abc.abstractmethod
def get_cov_as_linear_operator(self):
pass
@abc.abstractmethod
def register_matpower(self, exp, damping_func):
pass
@abc.abstractmethod
def register_cholesky(self, damping_func):
pass
@abc.abstractmethod
def register_cholesky_inverse(self, damping_func):
pass
@abc.abstractmethod
def get_matpower(self, exp, damping_func):
pass
@abc.abstractmethod
def get_cholesky(self, damping_func):
pass
@abc.abstractmethod
def get_cholesky_inverse(self, damping_func):
pass
class DenseSquareMatrixFactor(FisherFactor):
"""Base class for FisherFactors that are stored as dense square matrices.
This class explicitly calculates and stores inverses of their `cov` matrices,
which must be square dense matrices.
Subclasses must implement the _compute_new_cov method, and the _var_scope and
_cov_shape properties.
"""
# TODO(b/69108481): This class (and its subclasses) should be refactored to
# serve the matrix quantities it computes as both (potentially stale)
# variables, updated by the inverse update ops, and fresh values stored in
# tensors that recomputed once every session.run() call. Currently matpower
# and damp_inverse have the former behavior, while eigendecomposition has
# the latter.
def __init__(self):
self._matpower_by_exp_and_damping = {} # { (float, hashable): variable }
self._matpower_registrations = set() # { (float, hashable) }
self._eigendecomp = None
self._damping_funcs_by_id = {} # {hashable: lambda}
self._cholesky_registrations = set() # { hashable }
self._cholesky_inverse_registrations = set() # { hashable }
self._cholesky_by_damping = {} # { hashable: variable }
self._cholesky_inverse_by_damping = {} # { hashable: variable }
super(DenseSquareMatrixFactor, self).__init__()
def get_cov_as_linear_operator(self):
assert self.get_cov().shape.ndims == 2
return lo.LinearOperatorFullMatrix(self.get_cov(),
is_self_adjoint=True,
is_square=True)
def _register_damping(self, damping_func):
damping_id = graph_func_to_id(damping_func)
if damping_id not in self._damping_funcs_by_id:
self._damping_funcs_by_id[damping_id] = damping_func
return damping_id
def register_inverse(self, damping_func):
# Just for backwards compatibility of some old code and tests
self.register_matpower(-1, damping_func)
def register_matpower(self, exp, damping_func):
"""Registers a matrix power to be maintained and served on demand.
This creates a variable and signals make_inverse_update_ops to make the
corresponding update op. The variable can be read via the method
get_matpower.
Args:
exp: float. The exponent to use in the matrix power.
damping_func: A function that computes a 0-D Tensor or a float which will
be the damping value used. i.e. damping = damping_func().
"""
if exp == 1.0:
return
damping_id = self._register_damping(damping_func)
if (exp, damping_id) not in self._matpower_registrations:
self._matpower_registrations.add((exp, damping_id))
def register_cholesky(self, damping_func):
"""Registers a Cholesky factor to be maintained and served on demand.
This creates a variable and signals make_inverse_update_ops to make the
corresponding update op. The variable can be read via the method
get_cholesky.
Args:
damping_func: A function that computes a 0-D Tensor or a float which will
be the damping value used. i.e. damping = damping_func().
"""
damping_id = self._register_damping(damping_func)
if damping_id not in self._cholesky_registrations:
self._cholesky_registrations.add(damping_id)
def register_cholesky_inverse(self, damping_func):
"""Registers an inverse Cholesky factor to be maintained/served on demand.
This creates a variable and signals make_inverse_update_ops to make the
corresponding update op. The variable can be read via the method
get_cholesky_inverse.
Args:
damping_func: A function that computes a 0-D Tensor or a float which will
be the damping value used. i.e. damping = damping_func().
"""
damping_id = self._register_damping(damping_func)
if damping_id not in self._cholesky_inverse_registrations:
self._cholesky_inverse_registrations.add(damping_id)
def instantiate_inv_variables(self):
"""Makes the internal "inverse" variable(s)."""
for (exp, damping_id) in self._matpower_registrations:
exp_string = scalar_or_tensor_to_string(exp)
damping_func = self._damping_funcs_by_id[damping_id]
damping_string = graph_func_to_string(damping_func)
with variable_scope.variable_scope(self._var_scope):
matpower = variable_scope.get_variable(
"matpower_exp{}_damp{}".format(exp_string, damping_string),
initializer=inverse_initializer,
shape=self._cov_shape,
trainable=False,
dtype=self._dtype)
assert (exp, damping_id) not in self._matpower_by_exp_and_damping
self._matpower_by_exp_and_damping[(exp, damping_id)] = matpower
for damping_id in self._cholesky_registrations:
damping_func = self._damping_funcs_by_id[damping_id]
damping_string = graph_func_to_string(damping_func)
with variable_scope.variable_scope(self._var_scope):
chol = variable_scope.get_variable(
"cholesky_damp{}".format(damping_string),
initializer=inverse_initializer,
shape=self._cov_shape,
trainable=False,
dtype=self._dtype)
assert damping_id not in self._cholesky_by_damping
self._cholesky_by_damping[damping_id] = chol
for damping_id in self._cholesky_inverse_registrations:
damping_func = self._damping_funcs_by_id[damping_id]
damping_string = graph_func_to_string(damping_func)
with variable_scope.variable_scope(self._var_scope):
cholinv = variable_scope.get_variable(
"cholesky_inverse_damp{}".format(damping_string),
initializer=inverse_initializer,
shape=self._cov_shape,
trainable=False,
dtype=self._dtype)
assert damping_id not in self._cholesky_inverse_by_damping
self._cholesky_inverse_by_damping[damping_id] = cholinv
def make_inverse_update_ops(self):
"""Create and return update ops corresponding to registered computations."""
ops = []
num_inverses = sum(1 for (exp, _) in self._matpower_by_exp_and_damping
if exp == -1)
num_other_matpower = len(self._matpower_by_exp_and_damping) - num_inverses
other_matrix_power_registered = num_other_matpower >= 1
use_eig = (
self._eigendecomp or other_matrix_power_registered or
num_inverses >= EIGENVALUE_DECOMPOSITION_THRESHOLD)
# We precompute these so we don't need to evaluate them multiple times (for
# each matrix power that uses them)
damping_value_by_id = {damping_id: math_ops.cast(
self._damping_funcs_by_id[damping_id](), self._dtype)
for damping_id in self._damping_funcs_by_id}
if use_eig:
eigenvalues, eigenvectors = self.get_eigendecomp() # pylint: disable=unpacking-non-sequence
for (exp, damping_id), matpower in (
self._matpower_by_exp_and_damping.items()):
damping = damping_value_by_id[damping_id]
ops.append(
matpower.assign(
math_ops.matmul(eigenvectors *
(eigenvalues + damping)**exp,
array_ops.transpose(eigenvectors))))
# These ops share computation and should be run on a single device.
ops = [control_flow_ops.group(*ops)]
else:
for (exp, damping_id), matpower in (
self._matpower_by_exp_and_damping.items()):
assert exp == -1
damping = damping_value_by_id[damping_id]
ops.append(matpower.assign(utils.posdef_inv(self.get_cov(), damping)))
# TODO(b/77902055): If inverses are being computed with Cholesky's
# we can share the work. Instead this code currently just computes the
# Cholesky a second time. It does at least share work between requests for
# Cholesky's and Cholesky inverses with the same damping id.
for damping_id, cholesky_inv in self._cholesky_inverse_by_damping.items():
cholesky_ops = []
damping = damping_value_by_id[damping_id]
cholesky_value = utils.cholesky(self.get_cov(), damping)
if damping_id in self._cholesky_by_damping:
cholesky = self._cholesky_by_damping[damping_id]
cholesky_ops.append(cholesky.assign(cholesky_value))
identity = linalg_ops.eye(cholesky_value.shape.as_list()[0],
dtype=cholesky_value.dtype)
cholesky_inv_value = linalg_ops.matrix_triangular_solve(cholesky_value,
identity)
cholesky_ops.append(cholesky_inv.assign(cholesky_inv_value))
ops.append(control_flow_ops.group(*cholesky_ops))
for damping_id, cholesky in self._cholesky_by_damping.items():
if damping_id not in self._cholesky_inverse_by_damping:
damping = damping_value_by_id[damping_id]
cholesky_value = utils.cholesky(self.get_cov(), damping)
ops.append(cholesky.assign(cholesky_value))
self._eigendecomp = False
return ops
def get_inverse(self, damping_func):
# Just for backwards compatibility of some old code and tests
return self.get_matpower(-1, damping_func)
def get_matpower(self, exp, damping_func):
# Note that this function returns a variable which gets updated by the
# inverse ops. It may be stale / inconsistent with the latest value of
# get_cov().
if exp != 1:
damping_id = graph_func_to_id(damping_func)
matpower = self._matpower_by_exp_and_damping[(exp, damping_id)]
else:
matpower = self.get_cov()
identity = linalg_ops.eye(matpower.shape.as_list()[0],
dtype=matpower.dtype)
matpower += math_ops.cast(damping_func(), dtype=matpower.dtype)*identity
assert matpower.shape.ndims == 2
return lo.LinearOperatorFullMatrix(matpower,
is_non_singular=True,
is_self_adjoint=True,
is_positive_definite=True,
is_square=True)
def get_cholesky(self, damping_func):
# Note that this function returns a variable which gets updated by the
# inverse ops. It may be stale / inconsistent with the latest value of
# get_cov().
damping_id = graph_func_to_id(damping_func)
cholesky = self._cholesky_by_damping[damping_id]
assert cholesky.shape.ndims == 2
return lo.LinearOperatorFullMatrix(cholesky,
is_non_singular=True,
is_square=True)
def get_cholesky_inverse(self, damping_func):
# Note that this function returns a variable which gets updated by the
# inverse ops. It may be stale / inconsistent with the latest value of
# get_cov().
damping_id = graph_func_to_id(damping_func)
cholesky_inv = self._cholesky_inverse_by_damping[damping_id]
assert cholesky_inv.shape.ndims == 2
return lo.LinearOperatorFullMatrix(cholesky_inv,
is_non_singular=True,
is_square=True)
def get_eigendecomp(self):
"""Creates or retrieves eigendecomposition of self._cov."""
# Unlike get_matpower this doesn't retrieve a stored variable, but instead
# always computes a fresh version from the current value of get_cov().
if not self._eigendecomp:
eigenvalues, eigenvectors = linalg_ops.self_adjoint_eig(self.get_cov())
# The matrix self._cov is positive semidefinite by construction, but the
# numerical eigenvalues could be negative due to numerical errors, so here
# we clip them to be at least FLAGS.eigenvalue_clipping_threshold
clipped_eigenvalues = math_ops.maximum(eigenvalues,
EIGENVALUE_CLIPPING_THRESHOLD)
self._eigendecomp = (clipped_eigenvalues, eigenvectors)
return self._eigendecomp
class FullFactor(DenseSquareMatrixFactor):
"""FisherFactor for a full matrix representation of the Fisher of a parameter.
Note that this uses the naive "square the sum estimator", and so is applicable
to any type of parameter in principle, but has very high variance.
"""
def __init__(self,
params_grads,
batch_size):
self._batch_size = batch_size
self._params_grads = tuple(utils.ensure_sequence(params_grad)
for params_grad in params_grads)
super(FullFactor, self).__init__()
@property
def _var_scope(self):
return "ff_full_" + scope_string_from_params(
[self._params_grads, self._batch_size])
@property
def _cov_shape(self):
size = sum(param_grad.shape.num_elements()
for param_grad in self._params_grads[0])
return (size, size)
@property
def _num_sources(self):
return len(self._params_grads)
@property
def _num_towers(self):
return 1
@property
def _dtype(self):
return self._params_grads[0][0].dtype
def _compute_new_cov(self, source, tower):
assert tower == 0
# This will be a very basic rank 1 estimate
params_grads_flat = utils.tensors_to_column(self._params_grads[source])
return ((params_grads_flat * array_ops.transpose(
params_grads_flat)) / math_ops.cast(self._batch_size,
params_grads_flat.dtype))
def _get_data_device(self, tower):
return None
class DiagonalFactor(FisherFactor):
"""A base class for FisherFactors that use diagonal approximations.
A DiagonalFactor's covariance variable can be of any shape, but must contain
exactly one entry per parameter.
"""
def __init__(self):
super(DiagonalFactor, self).__init__()
def get_cov_as_linear_operator(self):
assert self._matrix_diagonal.shape.ndims == 1
return lo.LinearOperatorDiag(self._matrix_diagonal,
is_self_adjoint=True,
is_square=True)
@property
def _cov_initializer(self):
return diagonal_covariance_initializer
@property
def _matrix_diagonal(self):
return array_ops.reshape(self.get_cov(), [-1])
def make_inverse_update_ops(self):
return []
def instantiate_inv_variables(self):
pass
def register_matpower(self, exp, damping_func):
pass
def register_cholesky(self, damping_func):
pass
def register_cholesky_inverse(self, damping_func):
pass
def get_matpower(self, exp, damping_func):
matpower_diagonal = (self._matrix_diagonal
+ math_ops.cast(damping_func(), self._dtype))**exp
return lo.LinearOperatorDiag(matpower_diagonal,
is_non_singular=True,
is_self_adjoint=True,
is_positive_definite=True,
is_square=True)
def get_cholesky(self, damping_func):
return self.get_matpower(0.5, damping_func)
def get_cholesky_inverse(self, damping_func):
return self.get_matpower(-0.5, damping_func)
class NaiveDiagonalFactor(DiagonalFactor):
"""FisherFactor for a diagonal approximation of any type of param's Fisher.
Note that this uses the naive "square the sum estimator", and so is applicable
to any type of parameter in principle, but has very high variance.
"""
def __init__(self,
params_grads,
batch_size):
"""Initializes NaiveDiagonalFactor instance.
Args:
params_grads: Sequence of Tensors, each with same shape as parameters this
FisherFactor corresponds to. For example, the gradient of the loss with
respect to parameters.
batch_size: int or 0-D Tensor. Size
"""
self._params_grads = tuple(utils.ensure_sequence(params_grad)
for params_grad in params_grads)
self._batch_size = batch_size
super(NaiveDiagonalFactor, self).__init__()
@property
def _var_scope(self):
return "ff_naivediag_" + scope_string_from_params(
[self._params_grads, self._batch_size])
@property
def _cov_shape(self):
size = sum(param_grad.shape.num_elements()
for param_grad in self._params_grads[0])
return [size, 1]
@property
def _num_sources(self):
return len(self._params_grads)
@property
def _num_towers(self):
return 1
@property
def _dtype(self):
return self._params_grads[0][0].dtype
def _compute_new_cov(self, source, tower):
assert tower == 0
params_grads_flat = utils.tensors_to_column(self._params_grads[source])
return (math_ops.square(params_grads_flat) / math_ops.cast(
self._batch_size, params_grads_flat.dtype))
def _get_data_device(self, tower):
return None
class EmbeddingInputKroneckerFactor(DiagonalFactor):
r"""FisherFactor for input to an embedding layer.
Given input_ids = [batch_size, input_size] representing indices into an
[vocab_size, embedding_size] embedding matrix, approximate input covariance by
a diagonal matrix,
Cov(input_ids, input_ids) =
(1/batch_size) sum_{i} diag(n_hot(input[i]) ** 2).
where n_hot() constructs an n-hot binary vector and diag() constructs a
diagonal matrix of size [vocab_size, vocab_size].
"""
def __init__(self, input_ids, vocab_size, dtype=None):
"""Instantiate EmbeddingInputKroneckerFactor.
Args:
input_ids: List of Tensors of shape [batch_size, input_size] and dtype
int32. Indices into embedding matrix. List index is tower.
vocab_size: int or 0-D Tensor. Maximum value for entries in 'input_ids'.
dtype: dtype for covariance statistics. Must be a floating point type.
Defaults to float32.
"""
self._input_ids = input_ids
self._vocab_size = vocab_size
self._cov_dtype = dtype or dtypes.float32
super(EmbeddingInputKroneckerFactor, self).__init__()
@property
def _var_scope(self):
return "ff_diag_embedding_" + scope_string_from_params(self._input_ids)
@property
def _cov_shape(self):
return [self._vocab_size]
@property
def _num_sources(self):
return 1
@property
def _num_towers(self):
return len(self._input_ids)
@property
def _dtype(self):
return self._cov_dtype
def _compute_new_cov(self, source, tower):
assert source == 0
input_ids = self._input_ids[tower]
if len(input_ids.shape) > 2:
raise ValueError(
"Input to embeddings must have rank <= 2. Found rank %d." % len(
input_ids.shape))
batch_size = array_ops.shape(input_ids)[0]
# Transform indices into one-hot vectors.
#
# TODO(b/72714822): There must be a faster way to construct the diagonal
# covariance matrix! This operation is O(batch_size * vocab_size), where
# it should be O(batch_size * input_size).