File: rnn_cell.py

package info (click to toggle)
pytorch 1.13.1%2Bdfsg-4
  • links: PTS, VCS
  • area: main
  • in suites: bookworm
  • size: 139,252 kB
  • sloc: cpp: 1,100,274; python: 706,454; ansic: 83,052; asm: 7,618; java: 3,273; sh: 2,841; javascript: 612; makefile: 323; xml: 269; ruby: 185; yacc: 144; objc: 68; lex: 44
file content (1981 lines) | stat: -rw-r--r-- 68,233 bytes parent folder | download
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
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
## @package rnn_cell
# Module caffe2.python.rnn_cell





import functools
import inspect
import logging
import numpy as np
import random
from future.utils import viewkeys

from caffe2.proto import caffe2_pb2
from caffe2.python.attention import (
    apply_dot_attention,
    apply_recurrent_attention,
    apply_regular_attention,
    apply_soft_coverage_attention,
    AttentionType,
)
from caffe2.python import core, recurrent, workspace, brew, scope, utils
from caffe2.python.modeling.parameter_sharing import ParameterSharing
from caffe2.python.modeling.parameter_info import ParameterTags
from caffe2.python.modeling.initializers import Initializer
from caffe2.python.model_helper import ModelHelper


def _RectifyName(blob_reference_or_name):
    if blob_reference_or_name is None:
        return None
    if isinstance(blob_reference_or_name, str):
        return core.ScopedBlobReference(blob_reference_or_name)
    if not isinstance(blob_reference_or_name, core.BlobReference):
        raise Exception("Unknown blob reference type")
    return blob_reference_or_name


def _RectifyNames(blob_references_or_names):
    if blob_references_or_names is None:
        return None
    return [_RectifyName(i) for i in blob_references_or_names]


class RNNCell(object):
    '''
    Base class for writing recurrent / stateful operations.

    One needs to implement 2 methods: apply_override
    and get_state_names_override.

    As a result base class will provice apply_over_sequence method, which
    allows you to apply recurrent operations over a sequence of any length.

    As optional you could add input and output preparation steps by overriding
    corresponding methods.
    '''
    def __init__(self, name=None, forward_only=False, initializer=None):
        self.name = name
        self.recompute_blobs = []
        self.forward_only = forward_only
        self._initializer = initializer

    @property
    def initializer(self):
        return self._initializer

    @initializer.setter
    def initializer(self, value):
        self._initializer = value

    def scope(self, name):
        return self.name + '/' + name if self.name is not None else name

    def apply_over_sequence(
        self,
        model,
        inputs,
        seq_lengths=None,
        initial_states=None,
        outputs_with_grads=None,
    ):
        if initial_states is None:
            with scope.NameScope(self.name):
                if self.initializer is None:
                    raise Exception("Either initial states "
                                    "or initializer have to be set")
                initial_states = self.initializer.create_states(model)

        preprocessed_inputs = self.prepare_input(model, inputs)
        step_model = ModelHelper(name=self.name, param_model=model)
        input_t, timestep = step_model.net.AddScopedExternalInputs(
            'input_t',
            'timestep',
        )
        utils.raiseIfNotEqual(
            len(initial_states), len(self.get_state_names()),
            "Number of initial state values provided doesn't match the number "
            "of states"
        )
        states_prev = step_model.net.AddScopedExternalInputs(*[
            s + '_prev' for s in self.get_state_names()
        ])
        states = self._apply(
            model=step_model,
            input_t=input_t,
            seq_lengths=seq_lengths,
            states=states_prev,
            timestep=timestep,
        )

        external_outputs = set(step_model.net.Proto().external_output)
        for state in states:
            if state not in external_outputs:
                step_model.net.AddExternalOutput(state)

        if outputs_with_grads is None:
            outputs_with_grads = [self.get_output_state_index() * 2]

        # states_for_all_steps consists of combination of
        # states gather for all steps and final states. It looks like this:
        # (state_1_all, state_1_final, state_2_all, state_2_final, ...)
        states_for_all_steps = recurrent.recurrent_net(
            net=model.net,
            cell_net=step_model.net,
            inputs=[(input_t, preprocessed_inputs)],
            initial_cell_inputs=list(zip(states_prev, initial_states)),
            links=dict(zip(states_prev, states)),
            timestep=timestep,
            scope=self.name,
            forward_only=self.forward_only,
            outputs_with_grads=outputs_with_grads,
            recompute_blobs_on_backward=self.recompute_blobs,
        )

        output = self._prepare_output_sequence(
            model,
            states_for_all_steps,
        )
        return output, states_for_all_steps

    def apply(self, model, input_t, seq_lengths, states, timestep):
        input_t = self.prepare_input(model, input_t)
        states = self._apply(
            model, input_t, seq_lengths, states, timestep)
        output = self._prepare_output(model, states)
        return output, states

    def _apply(
        self,
        model, input_t, seq_lengths, states, timestep, extra_inputs=None
    ):
        '''
        This  method uses apply_override provided by a custom cell.
        On the top it takes care of applying self.scope() to all the outputs.
        While all the inputs stay within the scope this function was called
        from.
        '''
        args = self._rectify_apply_inputs(
            input_t, seq_lengths, states, timestep, extra_inputs)
        with core.NameScope(self.name):
            return self.apply_override(model, *args)

    def _rectify_apply_inputs(
            self, input_t, seq_lengths, states, timestep, extra_inputs):
        '''
        Before applying a scope we make sure that all external blob names
        are converted to blob reference. So further scoping doesn't affect them
        '''

        input_t, seq_lengths, timestep = _RectifyNames(
            [input_t, seq_lengths, timestep])
        states = _RectifyNames(states)
        if extra_inputs:
            extra_input_names, extra_input_sizes = zip(*extra_inputs)
            extra_inputs = _RectifyNames(extra_input_names)
            extra_inputs = zip(extra_input_names, extra_input_sizes)

        arg_names = inspect.getargspec(self.apply_override).args
        rectified = [input_t, seq_lengths, states, timestep]
        if 'extra_inputs' in arg_names:
            rectified.append(extra_inputs)
        return rectified


    def apply_override(
        self,
        model, input_t, seq_lengths, timestep, extra_inputs=None,
    ):
        '''
        A single step of a recurrent network to be implemented by each custom
        RNNCell.

        model: ModelHelper object new operators would be added to

        input_t: singlse input with shape (1, batch_size, input_dim)

        seq_lengths: blob containing sequence lengths which would be passed to
        LSTMUnit operator

        states: previous recurrent states

        timestep: current recurrent iteration. Could be used together with
        seq_lengths in order to determine, if some shorter sequences
        in the batch have already ended.

        extra_inputs: list of tuples (input, dim). specifies additional input
        which is not subject to prepare_input(). (useful when a cell is a
        component of a larger recurrent structure, e.g., attention)
        '''
        raise NotImplementedError('Abstract method')

    def prepare_input(self, model, input_blob):
        '''
        If some operations in _apply method depend only on the input,
        not on recurrent states, they could be computed in advance.

        model: ModelHelper object new operators would be added to

        input_blob: either the whole input sequence with shape
        (sequence_length, batch_size, input_dim) or a single input with shape
        (1, batch_size, input_dim).
        '''
        return input_blob

    def get_output_state_index(self):
        '''
        Return index into state list of the "primary" step-wise output.
        '''
        return 0

    def get_state_names(self):
        '''
        Returns recurrent state names with self.name scoping applied
        '''
        return [self.scope(name) for name in self.get_state_names_override()]

    def get_state_names_override(self):
        '''
        Override this function in your custom cell.
        It should return the names of the recurrent states.

        It's required by apply_over_sequence method in order to allocate
        recurrent states for all steps with meaningful names.
        '''
        raise NotImplementedError('Abstract method')

    def get_output_dim(self):
        '''
        Specifies the dimension (number of units) of stepwise output.
        '''
        raise NotImplementedError('Abstract method')

    def _prepare_output(self, model, states):
        '''
        Allows arbitrary post-processing of primary output.
        '''
        return states[self.get_output_state_index()]

    def _prepare_output_sequence(self, model, state_outputs):
        '''
        Allows arbitrary post-processing of primary sequence output.

        (Note that state_outputs alternates between full-sequence and final
        output for each state, thus the index multiplier 2.)
        '''
        output_sequence_index = 2 * self.get_output_state_index()
        return state_outputs[output_sequence_index]


class LSTMInitializer(object):
    def __init__(self, hidden_size):
        self.hidden_size = hidden_size

    def create_states(self, model):
        return [
            model.create_param(
                param_name='initial_hidden_state',
                initializer=Initializer(operator_name='ConstantFill',
                                        value=0.0),
                shape=[self.hidden_size],
            ),
            model.create_param(
                param_name='initial_cell_state',
                initializer=Initializer(operator_name='ConstantFill',
                                        value=0.0),
                shape=[self.hidden_size],
            )
        ]


# based on https://pytorch.org/docs/master/nn.html#torch.nn.RNNCell
class BasicRNNCell(RNNCell):
    def __init__(
        self,
        input_size,
        hidden_size,
        forget_bias,
        memory_optimization,
        drop_states=False,
        initializer=None,
        activation=None,
        **kwargs
    ):
        super(BasicRNNCell, self).__init__(**kwargs)
        self.drop_states = drop_states
        self.input_size = input_size
        self.hidden_size = hidden_size
        self.activation = activation

        if self.activation not in ['relu', 'tanh']:
            raise RuntimeError(
                'BasicRNNCell with unknown activation function (%s)'
                % self.activation)

    def apply_override(
        self,
        model,
        input_t,
        seq_lengths,
        states,
        timestep,
        extra_inputs=None,
    ):
        hidden_t_prev = states[0]

        gates_t = brew.fc(
            model,
            hidden_t_prev,
            'gates_t',
            dim_in=self.hidden_size,
            dim_out=self.hidden_size,
            axis=2,
        )

        brew.sum(model, [gates_t, input_t], gates_t)
        if self.activation == 'tanh':
            hidden_t = model.net.Tanh(gates_t, 'hidden_t')
        elif self.activation == 'relu':
            hidden_t = model.net.Relu(gates_t, 'hidden_t')
        else:
            raise RuntimeError(
                'BasicRNNCell with unknown activation function (%s)'
                % self.activation)

        if seq_lengths is not None:
            # TODO If this codepath becomes popular, it may be worth
            # taking a look at optimizing it - for now a simple
            # implementation is used to round out compatibility with
            # ONNX.
            timestep = model.net.CopyFromCPUInput(
                timestep, 'timestep_gpu')
            valid_b = model.net.GT(
                [seq_lengths, timestep], 'valid_b', broadcast=1)
            invalid_b = model.net.LE(
                [seq_lengths, timestep], 'invalid_b', broadcast=1)
            valid = model.net.Cast(valid_b, 'valid', to='float')
            invalid = model.net.Cast(invalid_b, 'invalid', to='float')

            hidden_valid = model.net.Mul(
                [hidden_t, valid],
                'hidden_valid',
                broadcast=1,
                axis=1,
            )
            if self.drop_states:
                hidden_t = hidden_valid
            else:
                hidden_invalid = model.net.Mul(
                    [hidden_t_prev, invalid],
                    'hidden_invalid',
                    broadcast=1, axis=1)
                hidden_t = model.net.Add(
                    [hidden_valid, hidden_invalid], hidden_t)
        return (hidden_t,)

    def prepare_input(self, model, input_blob):
        return brew.fc(
            model,
            input_blob,
            self.scope('i2h'),
            dim_in=self.input_size,
            dim_out=self.hidden_size,
            axis=2,
        )

    def get_state_names(self):
        return (self.scope('hidden_t'),)

    def get_output_dim(self):
        return self.hidden_size


class LSTMCell(RNNCell):

    def __init__(
        self,
        input_size,
        hidden_size,
        forget_bias,
        memory_optimization,
        drop_states=False,
        initializer=None,
        **kwargs
    ):
        super(LSTMCell, self).__init__(initializer=initializer, **kwargs)
        self.initializer = initializer or LSTMInitializer(
            hidden_size=hidden_size)

        self.input_size = input_size
        self.hidden_size = hidden_size
        self.forget_bias = float(forget_bias)
        self.memory_optimization = memory_optimization
        self.drop_states = drop_states
        self.gates_size = 4 * self.hidden_size

    def apply_override(
        self,
        model,
        input_t,
        seq_lengths,
        states,
        timestep,
        extra_inputs=None,
    ):
        hidden_t_prev, cell_t_prev = states

        fc_input = hidden_t_prev
        fc_input_dim = self.hidden_size

        if extra_inputs is not None:
            extra_input_blobs, extra_input_sizes = zip(*extra_inputs)
            fc_input = brew.concat(
                model,
                [hidden_t_prev] + list(extra_input_blobs),
                'gates_concatenated_input_t',
                axis=2,
            )
            fc_input_dim += sum(extra_input_sizes)

        gates_t = brew.fc(
            model,
            fc_input,
            'gates_t',
            dim_in=fc_input_dim,
            dim_out=self.gates_size,
            axis=2,
        )
        brew.sum(model, [gates_t, input_t], gates_t)

        if seq_lengths is not None:
            inputs = [hidden_t_prev, cell_t_prev, gates_t, seq_lengths, timestep]
        else:
            inputs = [hidden_t_prev, cell_t_prev, gates_t, timestep]

        hidden_t, cell_t = model.net.LSTMUnit(
            inputs,
            ['hidden_state', 'cell_state'],
            forget_bias=self.forget_bias,
            drop_states=self.drop_states,
            sequence_lengths=(seq_lengths is not None),
        )
        model.net.AddExternalOutputs(hidden_t, cell_t)
        if self.memory_optimization:
            self.recompute_blobs = [gates_t]

        return hidden_t, cell_t

    def get_input_params(self):
        return {
            'weights': self.scope('i2h') + '_w',
            'biases': self.scope('i2h') + '_b',
        }

    def get_recurrent_params(self):
        return {
            'weights': self.scope('gates_t') + '_w',
            'biases': self.scope('gates_t') + '_b',
        }

    def prepare_input(self, model, input_blob):
        return brew.fc(
            model,
            input_blob,
            self.scope('i2h'),
            dim_in=self.input_size,
            dim_out=self.gates_size,
            axis=2,
        )

    def get_state_names_override(self):
        return ['hidden_t', 'cell_t']

    def get_output_dim(self):
        return self.hidden_size


class LayerNormLSTMCell(RNNCell):

    def __init__(
        self,
        input_size,
        hidden_size,
        forget_bias,
        memory_optimization,
        drop_states=False,
        initializer=None,
        **kwargs
    ):
        super(LayerNormLSTMCell, self).__init__(
            initializer=initializer, **kwargs
        )
        self.initializer = initializer or LSTMInitializer(
            hidden_size=hidden_size
        )

        self.input_size = input_size
        self.hidden_size = hidden_size
        self.forget_bias = float(forget_bias)
        self.memory_optimization = memory_optimization
        self.drop_states = drop_states
        self.gates_size = 4 * self.hidden_size

    def _apply(
        self,
        model,
        input_t,
        seq_lengths,
        states,
        timestep,
        extra_inputs=None,
    ):
        hidden_t_prev, cell_t_prev = states

        fc_input = hidden_t_prev
        fc_input_dim = self.hidden_size

        if extra_inputs is not None:
            extra_input_blobs, extra_input_sizes = zip(*extra_inputs)
            fc_input = brew.concat(
                model,
                [hidden_t_prev] + list(extra_input_blobs),
                self.scope('gates_concatenated_input_t'),
                axis=2,
            )
            fc_input_dim += sum(extra_input_sizes)

        gates_t = brew.fc(
            model,
            fc_input,
            self.scope('gates_t'),
            dim_in=fc_input_dim,
            dim_out=self.gates_size,
            axis=2,
        )
        brew.sum(model, [gates_t, input_t], gates_t)

        # brew.layer_norm call is only difference from LSTMCell
        gates_t, _, _ = brew.layer_norm(
            model,
            self.scope('gates_t'),
            self.scope('gates_t_norm'),
            dim_in=self.gates_size,
            axis=-1,
        )

        hidden_t, cell_t = model.net.LSTMUnit(
            [
                hidden_t_prev,
                cell_t_prev,
                gates_t,
                seq_lengths,
                timestep,
            ],
            self.get_state_names(),
            forget_bias=self.forget_bias,
            drop_states=self.drop_states,
        )
        model.net.AddExternalOutputs(hidden_t, cell_t)
        if self.memory_optimization:
            self.recompute_blobs = [gates_t]

        return hidden_t, cell_t

    def get_input_params(self):
        return {
            'weights': self.scope('i2h') + '_w',
            'biases': self.scope('i2h') + '_b',
        }

    def prepare_input(self, model, input_blob):
        return brew.fc(
            model,
            input_blob,
            self.scope('i2h'),
            dim_in=self.input_size,
            dim_out=self.gates_size,
            axis=2,
        )

    def get_state_names(self):
        return (self.scope('hidden_t'), self.scope('cell_t'))


class MILSTMCell(LSTMCell):

    def _apply(
        self,
        model,
        input_t,
        seq_lengths,
        states,
        timestep,
        extra_inputs=None,
    ):
        hidden_t_prev, cell_t_prev = states

        fc_input = hidden_t_prev
        fc_input_dim = self.hidden_size

        if extra_inputs is not None:
            extra_input_blobs, extra_input_sizes = zip(*extra_inputs)
            fc_input = brew.concat(
                model,
                [hidden_t_prev] + list(extra_input_blobs),
                self.scope('gates_concatenated_input_t'),
                axis=2,
            )
            fc_input_dim += sum(extra_input_sizes)

        prev_t = brew.fc(
            model,
            fc_input,
            self.scope('prev_t'),
            dim_in=fc_input_dim,
            dim_out=self.gates_size,
            axis=2,
        )

        # defining initializers for MI parameters
        alpha = model.create_param(
            self.scope('alpha'),
            shape=[self.gates_size],
            initializer=Initializer('ConstantFill', value=1.0),
        )
        beta_h = model.create_param(
            self.scope('beta1'),
            shape=[self.gates_size],
            initializer=Initializer('ConstantFill', value=1.0),
        )
        beta_i = model.create_param(
            self.scope('beta2'),
            shape=[self.gates_size],
            initializer=Initializer('ConstantFill', value=1.0),
        )
        b = model.create_param(
            self.scope('b'),
            shape=[self.gates_size],
            initializer=Initializer('ConstantFill', value=0.0),
        )

        # alpha * input_t + beta_h
        # Shape: [1, batch_size, 4 * hidden_size]
        alpha_by_input_t_plus_beta_h = model.net.ElementwiseLinear(
            [input_t, alpha, beta_h],
            self.scope('alpha_by_input_t_plus_beta_h'),
            axis=2,
        )
        # (alpha * input_t + beta_h) * prev_t =
        # alpha * input_t * prev_t + beta_h * prev_t
        # Shape: [1, batch_size, 4 * hidden_size]
        alpha_by_input_t_plus_beta_h_by_prev_t = model.net.Mul(
            [alpha_by_input_t_plus_beta_h, prev_t],
            self.scope('alpha_by_input_t_plus_beta_h_by_prev_t')
        )
        # beta_i * input_t + b
        # Shape: [1, batch_size, 4 * hidden_size]
        beta_i_by_input_t_plus_b = model.net.ElementwiseLinear(
            [input_t, beta_i, b],
            self.scope('beta_i_by_input_t_plus_b'),
            axis=2,
        )
        # alpha * input_t * prev_t + beta_h * prev_t + beta_i * input_t + b
        # Shape: [1, batch_size, 4 * hidden_size]
        gates_t = brew.sum(
            model,
            [alpha_by_input_t_plus_beta_h_by_prev_t, beta_i_by_input_t_plus_b],
            self.scope('gates_t')
        )
        hidden_t, cell_t = model.net.LSTMUnit(
            [hidden_t_prev, cell_t_prev, gates_t, seq_lengths, timestep],
            [self.scope('hidden_t_intermediate'), self.scope('cell_t')],
            forget_bias=self.forget_bias,
            drop_states=self.drop_states,
        )
        model.net.AddExternalOutputs(
            cell_t,
            hidden_t,
        )
        if self.memory_optimization:
            self.recompute_blobs = [gates_t]
        return hidden_t, cell_t


class LayerNormMILSTMCell(LSTMCell):

    def _apply(
        self,
        model,
        input_t,
        seq_lengths,
        states,
        timestep,
        extra_inputs=None,
    ):
        hidden_t_prev, cell_t_prev = states

        fc_input = hidden_t_prev
        fc_input_dim = self.hidden_size

        if extra_inputs is not None:
            extra_input_blobs, extra_input_sizes = zip(*extra_inputs)
            fc_input = brew.concat(
                model,
                [hidden_t_prev] + list(extra_input_blobs),
                self.scope('gates_concatenated_input_t'),
                axis=2,
            )
            fc_input_dim += sum(extra_input_sizes)

        prev_t = brew.fc(
            model,
            fc_input,
            self.scope('prev_t'),
            dim_in=fc_input_dim,
            dim_out=self.gates_size,
            axis=2,
        )

        # defining initializers for MI parameters
        alpha = model.create_param(
            self.scope('alpha'),
            shape=[self.gates_size],
            initializer=Initializer('ConstantFill', value=1.0),
        )
        beta_h = model.create_param(
            self.scope('beta1'),
            shape=[self.gates_size],
            initializer=Initializer('ConstantFill', value=1.0),
        )
        beta_i = model.create_param(
            self.scope('beta2'),
            shape=[self.gates_size],
            initializer=Initializer('ConstantFill', value=1.0),
        )
        b = model.create_param(
            self.scope('b'),
            shape=[self.gates_size],
            initializer=Initializer('ConstantFill', value=0.0),
        )

        # alpha * input_t + beta_h
        # Shape: [1, batch_size, 4 * hidden_size]
        alpha_by_input_t_plus_beta_h = model.net.ElementwiseLinear(
            [input_t, alpha, beta_h],
            self.scope('alpha_by_input_t_plus_beta_h'),
            axis=2,
        )
        # (alpha * input_t + beta_h) * prev_t =
        # alpha * input_t * prev_t + beta_h * prev_t
        # Shape: [1, batch_size, 4 * hidden_size]
        alpha_by_input_t_plus_beta_h_by_prev_t = model.net.Mul(
            [alpha_by_input_t_plus_beta_h, prev_t],
            self.scope('alpha_by_input_t_plus_beta_h_by_prev_t')
        )
        # beta_i * input_t + b
        # Shape: [1, batch_size, 4 * hidden_size]
        beta_i_by_input_t_plus_b = model.net.ElementwiseLinear(
            [input_t, beta_i, b],
            self.scope('beta_i_by_input_t_plus_b'),
            axis=2,
        )
        # alpha * input_t * prev_t + beta_h * prev_t + beta_i * input_t + b
        # Shape: [1, batch_size, 4 * hidden_size]
        gates_t = brew.sum(
            model,
            [alpha_by_input_t_plus_beta_h_by_prev_t, beta_i_by_input_t_plus_b],
            self.scope('gates_t')
        )
        # brew.layer_norm call is only difference from MILSTMCell._apply
        gates_t, _, _ = brew.layer_norm(
            model,
            self.scope('gates_t'),
            self.scope('gates_t_norm'),
            dim_in=self.gates_size,
            axis=-1,
        )
        hidden_t, cell_t = model.net.LSTMUnit(
            [hidden_t_prev, cell_t_prev, gates_t, seq_lengths, timestep],
            [self.scope('hidden_t_intermediate'), self.scope('cell_t')],
            forget_bias=self.forget_bias,
            drop_states=self.drop_states,
        )
        model.net.AddExternalOutputs(
            cell_t,
            hidden_t,
        )
        if self.memory_optimization:
            self.recompute_blobs = [gates_t]
        return hidden_t, cell_t


class DropoutCell(RNNCell):
    '''
    Wraps arbitrary RNNCell, applying dropout to its output (but not to the
    recurrent connection for the corresponding state).
    '''

    def __init__(
        self,
        internal_cell,
        dropout_ratio=None,
        use_cudnn=False,
        **kwargs
    ):
        self.internal_cell = internal_cell
        self.dropout_ratio = dropout_ratio
        assert 'is_test' in kwargs, "Argument 'is_test' is required"
        self.is_test = kwargs.pop('is_test')
        self.use_cudnn = use_cudnn
        super(DropoutCell, self).__init__(**kwargs)

        self.prepare_input = internal_cell.prepare_input
        self.get_output_state_index = internal_cell.get_output_state_index
        self.get_state_names = internal_cell.get_state_names
        self.get_output_dim = internal_cell.get_output_dim

        self.mask = 0

    def _apply(
        self,
        model,
        input_t,
        seq_lengths,
        states,
        timestep,
        extra_inputs=None,
    ):
        return self.internal_cell._apply(
            model,
            input_t,
            seq_lengths,
            states,
            timestep,
            extra_inputs,
        )

    def _prepare_output(self, model, states):
        output = self.internal_cell._prepare_output(
            model,
            states,
        )
        if self.dropout_ratio is not None:
            output = self._apply_dropout(model, output)
        return output

    def _prepare_output_sequence(self, model, state_outputs):
        output = self.internal_cell._prepare_output_sequence(
            model,
            state_outputs,
        )
        if self.dropout_ratio is not None:
            output = self._apply_dropout(model, output)
        return output

    def _apply_dropout(self, model, output):
        if self.dropout_ratio and not self.forward_only:
            with core.NameScope(self.name or ''):
                output = brew.dropout(
                    model,
                    output,
                    str(output) + '_with_dropout_mask{}'.format(self.mask),
                    ratio=float(self.dropout_ratio),
                    is_test=self.is_test,
                    use_cudnn=self.use_cudnn,
                )
                self.mask += 1
        return output


class MultiRNNCellInitializer(object):
    def __init__(self, cells):
        self.cells = cells

    def create_states(self, model):
        states = []
        for i, cell in enumerate(self.cells):
            if cell.initializer is None:
                raise Exception("Either initial states "
                                "or initializer have to be set")

            with core.NameScope("layer_{}".format(i)),\
                    core.NameScope(cell.name):
                states.extend(cell.initializer.create_states(model))
        return states


class MultiRNNCell(RNNCell):
    '''
    Multilayer RNN via the composition of RNNCell instance.

    It is the responsibility of calling code to ensure the compatibility
    of the successive layers in terms of input/output dimensiality, etc.,
    and to ensure that their blobs do not have name conflicts, typically by
    creating the cells with names that specify layer number.

    Assumes first state (recurrent output) for each layer should be the input
    to the next layer.
    '''

    def __init__(self, cells, residual_output_layers=None, **kwargs):
        '''
        cells: list of RNNCell instances, from input to output side.

        name: string designating network component (for scoping)

        residual_output_layers: list of indices of layers whose input will
        be added elementwise to their output elementwise. (It is the
        responsibility of the client code to ensure shape compatibility.)
        Note that layer 0 (zero) cannot have residual output because of the
        timing of prepare_input().

        forward_only: used to construct inference-only network.
        '''
        super(MultiRNNCell, self).__init__(**kwargs)
        self.cells = cells

        if residual_output_layers is None:
            self.residual_output_layers = []
        else:
            self.residual_output_layers = residual_output_layers

        output_index_per_layer = []
        base_index = 0
        for cell in self.cells:
            output_index_per_layer.append(
                base_index + cell.get_output_state_index(),
            )
            base_index += len(cell.get_state_names())

        self.output_connected_layers = []
        self.output_indices = []
        for i in range(len(self.cells) - 1):
            if (i + 1) in self.residual_output_layers:
                self.output_connected_layers.append(i)
                self.output_indices.append(output_index_per_layer[i])
            else:
                self.output_connected_layers = []
                self.output_indices = []
        self.output_connected_layers.append(len(self.cells) - 1)
        self.output_indices.append(output_index_per_layer[-1])

        self.state_names = []
        for i, cell in enumerate(self.cells):
            self.state_names.extend(
                map(self.layer_scoper(i), cell.get_state_names())
            )

        self.initializer = MultiRNNCellInitializer(cells)

    def layer_scoper(self, layer_id):
        def helper(name):
            return "{}/layer_{}/{}".format(self.name, layer_id, name)
        return helper

    def prepare_input(self, model, input_blob):
        input_blob = _RectifyName(input_blob)
        with core.NameScope(self.name or ''):
            return self.cells[0].prepare_input(model, input_blob)

    def _apply(
        self,
        model,
        input_t,
        seq_lengths,
        states,
        timestep,
        extra_inputs=None,
    ):
        '''
        Because below we will do scoping across layers, we need
        to make sure that string blob names are convereted to BlobReference
        objects.
        '''

        input_t, seq_lengths, states, timestep, extra_inputs = \
            self._rectify_apply_inputs(
                input_t, seq_lengths, states, timestep, extra_inputs)

        states_per_layer = [len(cell.get_state_names()) for cell in self.cells]
        assert len(states) == sum(states_per_layer)

        next_states = []
        states_index = 0

        layer_input = input_t
        for i, layer_cell in enumerate(self.cells):
            # # If cells don't have different names we still
            # take care of scoping
            with core.NameScope(self.name), core.NameScope("layer_{}".format(i)):
                num_states = states_per_layer[i]
                layer_states = states[states_index:(states_index + num_states)]
                states_index += num_states

                if i > 0:
                    prepared_input = layer_cell.prepare_input(
                        model, layer_input)
                else:
                    prepared_input = layer_input

                layer_next_states = layer_cell._apply(
                    model,
                    prepared_input,
                    seq_lengths,
                    layer_states,
                    timestep,
                    extra_inputs=(None if i > 0 else extra_inputs),
                )
                # Since we're using here non-public method _apply,
                # instead of apply, we have to manually extract output
                # from states
                if i != len(self.cells) - 1:
                    layer_output = layer_cell._prepare_output(
                        model,
                        layer_next_states,
                    )
                    if i > 0 and i in self.residual_output_layers:
                        layer_input = brew.sum(
                            model,
                            [layer_output, layer_input],
                            self.scope('residual_output_{}'.format(i)),
                        )
                    else:
                        layer_input = layer_output

                next_states.extend(layer_next_states)
        return next_states

    def get_state_names(self):
        return self.state_names

    def get_output_state_index(self):
        index = 0
        for cell in self.cells[:-1]:
            index += len(cell.get_state_names())
        index += self.cells[-1].get_output_state_index()
        return index

    def _prepare_output(self, model, states):
        connected_outputs = []
        state_index = 0
        for i, cell in enumerate(self.cells):
            num_states = len(cell.get_state_names())
            if i in self.output_connected_layers:
                layer_states = states[state_index:state_index + num_states]
                layer_output = cell._prepare_output(
                    model,
                    layer_states
                )
                connected_outputs.append(layer_output)
            state_index += num_states
        if len(connected_outputs) > 1:
            output = brew.sum(
                model,
                connected_outputs,
                self.scope('residual_output'),
            )
        else:
            output = connected_outputs[0]
        return output

    def _prepare_output_sequence(self, model, states):
        connected_outputs = []
        state_index = 0
        for i, cell in enumerate(self.cells):
            num_states = 2 * len(cell.get_state_names())
            if i in self.output_connected_layers:
                layer_states = states[state_index:state_index + num_states]
                layer_output = cell._prepare_output_sequence(
                    model,
                    layer_states
                )
                connected_outputs.append(layer_output)
            state_index += num_states
        if len(connected_outputs) > 1:
            output = brew.sum(
                model,
                connected_outputs,
                self.scope('residual_output_sequence'),
            )
        else:
            output = connected_outputs[0]
        return output


class AttentionCell(RNNCell):

    def __init__(
        self,
        encoder_output_dim,
        encoder_outputs,
        encoder_lengths,
        decoder_cell,
        decoder_state_dim,
        attention_type,
        weighted_encoder_outputs,
        attention_memory_optimization,
        **kwargs
    ):
        super(AttentionCell, self).__init__(**kwargs)
        self.encoder_output_dim = encoder_output_dim
        self.encoder_outputs = encoder_outputs
        self.encoder_lengths = encoder_lengths
        self.decoder_cell = decoder_cell
        self.decoder_state_dim = decoder_state_dim
        self.weighted_encoder_outputs = weighted_encoder_outputs
        self.encoder_outputs_transposed = None
        assert attention_type in [
            AttentionType.Regular,
            AttentionType.Recurrent,
            AttentionType.Dot,
            AttentionType.SoftCoverage,
        ]
        self.attention_type = attention_type
        self.attention_memory_optimization = attention_memory_optimization

    def _apply(
        self,
        model,
        input_t,
        seq_lengths,
        states,
        timestep,
        extra_inputs=None,
    ):
        if self.attention_type == AttentionType.SoftCoverage:
            decoder_prev_states = states[:-2]
            attention_weighted_encoder_context_t_prev = states[-2]
            coverage_t_prev = states[-1]
        else:
            decoder_prev_states = states[:-1]
            attention_weighted_encoder_context_t_prev = states[-1]

        assert extra_inputs is None

        decoder_states = self.decoder_cell._apply(
            model,
            input_t,
            seq_lengths,
            decoder_prev_states,
            timestep,
            extra_inputs=[(
                attention_weighted_encoder_context_t_prev,
                self.encoder_output_dim,
            )],
        )

        self.hidden_t_intermediate = self.decoder_cell._prepare_output(
            model,
            decoder_states,
        )

        if self.attention_type == AttentionType.Recurrent:
            (
                attention_weighted_encoder_context_t,
                self.attention_weights_3d,
                attention_blobs,
            ) = apply_recurrent_attention(
                model=model,
                encoder_output_dim=self.encoder_output_dim,
                encoder_outputs_transposed=self.encoder_outputs_transposed,
                weighted_encoder_outputs=self.weighted_encoder_outputs,
                decoder_hidden_state_t=self.hidden_t_intermediate,
                decoder_hidden_state_dim=self.decoder_state_dim,
                scope=self.name,
                attention_weighted_encoder_context_t_prev=(
                    attention_weighted_encoder_context_t_prev
                ),
                encoder_lengths=self.encoder_lengths,
            )
        elif self.attention_type == AttentionType.Regular:
            (
                attention_weighted_encoder_context_t,
                self.attention_weights_3d,
                attention_blobs,
            ) = apply_regular_attention(
                model=model,
                encoder_output_dim=self.encoder_output_dim,
                encoder_outputs_transposed=self.encoder_outputs_transposed,
                weighted_encoder_outputs=self.weighted_encoder_outputs,
                decoder_hidden_state_t=self.hidden_t_intermediate,
                decoder_hidden_state_dim=self.decoder_state_dim,
                scope=self.name,
                encoder_lengths=self.encoder_lengths,
            )
        elif self.attention_type == AttentionType.Dot:
            (
                attention_weighted_encoder_context_t,
                self.attention_weights_3d,
                attention_blobs,
            ) = apply_dot_attention(
                model=model,
                encoder_output_dim=self.encoder_output_dim,
                encoder_outputs_transposed=self.encoder_outputs_transposed,
                decoder_hidden_state_t=self.hidden_t_intermediate,
                decoder_hidden_state_dim=self.decoder_state_dim,
                scope=self.name,
                encoder_lengths=self.encoder_lengths,
            )
        elif self.attention_type == AttentionType.SoftCoverage:
            (
                attention_weighted_encoder_context_t,
                self.attention_weights_3d,
                attention_blobs,
                coverage_t,
            ) = apply_soft_coverage_attention(
                model=model,
                encoder_output_dim=self.encoder_output_dim,
                encoder_outputs_transposed=self.encoder_outputs_transposed,
                weighted_encoder_outputs=self.weighted_encoder_outputs,
                decoder_hidden_state_t=self.hidden_t_intermediate,
                decoder_hidden_state_dim=self.decoder_state_dim,
                scope=self.name,
                encoder_lengths=self.encoder_lengths,
                coverage_t_prev=coverage_t_prev,
                coverage_weights=self.coverage_weights,
            )
        else:
            raise Exception('Attention type {} not implemented'.format(
                self.attention_type
            ))

        if self.attention_memory_optimization:
            self.recompute_blobs.extend(attention_blobs)

        output = list(decoder_states) + [attention_weighted_encoder_context_t]
        if self.attention_type == AttentionType.SoftCoverage:
            output.append(coverage_t)

        output[self.decoder_cell.get_output_state_index()] = model.Copy(
            output[self.decoder_cell.get_output_state_index()],
            self.scope('hidden_t_external'),
        )
        model.net.AddExternalOutputs(*output)

        return output

    def get_attention_weights(self):
        # [batch_size, encoder_length, 1]
        return self.attention_weights_3d

    def prepare_input(self, model, input_blob):
        if self.encoder_outputs_transposed is None:
            self.encoder_outputs_transposed = brew.transpose(
                model,
                self.encoder_outputs,
                self.scope('encoder_outputs_transposed'),
                axes=[1, 2, 0],
            )
        if (
            self.weighted_encoder_outputs is None and
            self.attention_type != AttentionType.Dot
        ):
            self.weighted_encoder_outputs = brew.fc(
                model,
                self.encoder_outputs,
                self.scope('weighted_encoder_outputs'),
                dim_in=self.encoder_output_dim,
                dim_out=self.encoder_output_dim,
                axis=2,
            )

        return self.decoder_cell.prepare_input(model, input_blob)

    def build_initial_coverage(self, model):
        """
        initial_coverage is always zeros of shape [encoder_length],
        which shape must be determined programmatically dureing network
        computation.

        This method also sets self.coverage_weights, a separate transform
        of encoder_outputs which is used to determine coverage contribution
        tp attention.
        """
        assert self.attention_type == AttentionType.SoftCoverage

        # [encoder_length, batch_size, encoder_output_dim]
        self.coverage_weights = brew.fc(
            model,
            self.encoder_outputs,
            self.scope('coverage_weights'),
            dim_in=self.encoder_output_dim,
            dim_out=self.encoder_output_dim,
            axis=2,
        )

        encoder_length = model.net.Slice(
            model.net.Shape(self.encoder_outputs),
            starts=[0],
            ends=[1],
        )
        if (
            scope.CurrentDeviceScope() is not None and
            core.IsGPUDeviceType(scope.CurrentDeviceScope().device_type)
        ):
            encoder_length = model.net.CopyGPUToCPU(
                encoder_length,
                'encoder_length_cpu',
            )
        # total attention weight applied across decoding steps_per_checkpoint
        # shape: [encoder_length]
        initial_coverage = model.net.ConstantFill(
            encoder_length,
            self.scope('initial_coverage'),
            value=0.0,
            input_as_shape=1,
        )
        return initial_coverage

    def get_state_names(self):
        state_names = list(self.decoder_cell.get_state_names())
        state_names[self.get_output_state_index()] = self.scope(
            'hidden_t_external',
        )
        state_names.append(self.scope('attention_weighted_encoder_context_t'))
        if self.attention_type == AttentionType.SoftCoverage:
            state_names.append(self.scope('coverage_t'))
        return state_names

    def get_output_dim(self):
        return self.decoder_state_dim + self.encoder_output_dim

    def get_output_state_index(self):
        return self.decoder_cell.get_output_state_index()

    def _prepare_output(self, model, states):
        if self.attention_type == AttentionType.SoftCoverage:
            attention_context = states[-2]
        else:
            attention_context = states[-1]

        with core.NameScope(self.name or ''):
            output = brew.concat(
                model,
                [self.hidden_t_intermediate, attention_context],
                'states_and_context_combination',
                axis=2,
            )

        return output

    def _prepare_output_sequence(self, model, state_outputs):
        if self.attention_type == AttentionType.SoftCoverage:
            decoder_state_outputs = state_outputs[:-4]
        else:
            decoder_state_outputs = state_outputs[:-2]

        decoder_output = self.decoder_cell._prepare_output_sequence(
            model,
            decoder_state_outputs,
        )

        if self.attention_type == AttentionType.SoftCoverage:
            attention_context_index = 2 * (len(self.get_state_names()) - 2)
        else:
            attention_context_index = 2 * (len(self.get_state_names()) - 1)

        with core.NameScope(self.name or ''):
            output = brew.concat(
                model,
                [
                    decoder_output,
                    state_outputs[attention_context_index],
                ],
                'states_and_context_combination',
                axis=2,
            )
        return output


class LSTMWithAttentionCell(AttentionCell):

    def __init__(
        self,
        encoder_output_dim,
        encoder_outputs,
        encoder_lengths,
        decoder_input_dim,
        decoder_state_dim,
        name,
        attention_type,
        weighted_encoder_outputs,
        forget_bias,
        lstm_memory_optimization,
        attention_memory_optimization,
        forward_only=False,
    ):
        decoder_cell = LSTMCell(
            input_size=decoder_input_dim,
            hidden_size=decoder_state_dim,
            forget_bias=forget_bias,
            memory_optimization=lstm_memory_optimization,
            name='{}/decoder'.format(name),
            forward_only=False,
            drop_states=False,
        )
        super(LSTMWithAttentionCell, self).__init__(
            encoder_output_dim=encoder_output_dim,
            encoder_outputs=encoder_outputs,
            encoder_lengths=encoder_lengths,
            decoder_cell=decoder_cell,
            decoder_state_dim=decoder_state_dim,
            name=name,
            attention_type=attention_type,
            weighted_encoder_outputs=weighted_encoder_outputs,
            attention_memory_optimization=attention_memory_optimization,
            forward_only=forward_only,
        )


class MILSTMWithAttentionCell(AttentionCell):

    def __init__(
        self,
        encoder_output_dim,
        encoder_outputs,
        decoder_input_dim,
        decoder_state_dim,
        name,
        attention_type,
        weighted_encoder_outputs,
        forget_bias,
        lstm_memory_optimization,
        attention_memory_optimization,
        forward_only=False,
    ):
        decoder_cell = MILSTMCell(
            input_size=decoder_input_dim,
            hidden_size=decoder_state_dim,
            forget_bias=forget_bias,
            memory_optimization=lstm_memory_optimization,
            name='{}/decoder'.format(name),
            forward_only=False,
            drop_states=False,
        )
        super(MILSTMWithAttentionCell, self).__init__(
            encoder_output_dim=encoder_output_dim,
            encoder_outputs=encoder_outputs,
            decoder_cell=decoder_cell,
            decoder_state_dim=decoder_state_dim,
            name=name,
            attention_type=attention_type,
            weighted_encoder_outputs=weighted_encoder_outputs,
            attention_memory_optimization=attention_memory_optimization,
            forward_only=forward_only,
        )


def _LSTM(
    cell_class,
    model,
    input_blob,
    seq_lengths,
    initial_states,
    dim_in,
    dim_out,
    scope=None,
    outputs_with_grads=(0,),
    return_params=False,
    memory_optimization=False,
    forget_bias=0.0,
    forward_only=False,
    drop_states=False,
    return_last_layer_only=True,
    static_rnn_unroll_size=None,
    **cell_kwargs
):
    '''
    Adds a standard LSTM recurrent network operator to a model.

    cell_class: LSTMCell or compatible subclass

    model: ModelHelper object new operators would be added to

    input_blob: the input sequence in a format T x N x D
            where T is sequence size, N - batch size and D - input dimension

    seq_lengths: blob containing sequence lengths which would be passed to
            LSTMUnit operator

    initial_states: a list of (2 * num_layers) blobs representing the initial
            hidden and cell states of each layer. If this argument is None,
            these states will be added to the model as network parameters.

    dim_in: input dimension

    dim_out: number of units per LSTM layer
            (use int for single-layer LSTM, list of ints for multi-layer)

    outputs_with_grads : position indices of output blobs for LAST LAYER which
            will receive external error gradient during backpropagation.
            These outputs are: (h_all, h_last, c_all, c_last)

    return_params: if True, will return a dictionary of parameters of the LSTM

    memory_optimization: if enabled, the LSTM step is recomputed on backward
            step so that we don't need to store forward activations for each
            timestep. Saves memory with cost of computation.

    forget_bias: forget gate bias (default 0.0)

    forward_only: whether to create a backward pass

    drop_states: drop invalid states, passed through to LSTMUnit operator

    return_last_layer_only: only return outputs from final layer
            (so that length of results does depend on number of layers)

    static_rnn_unroll_size: if not None, we will use static RNN which is
    unrolled into Caffe2 graph. The size of the unroll is the value of
    this parameter.
    '''
    if type(dim_out) is not list and type(dim_out) is not tuple:
        dim_out = [dim_out]
    num_layers = len(dim_out)

    cells = []
    for i in range(num_layers):
        cell = cell_class(
            input_size=(dim_in if i == 0 else dim_out[i - 1]),
            hidden_size=dim_out[i],
            forget_bias=forget_bias,
            memory_optimization=memory_optimization,
            name=scope if num_layers == 1 else None,
            forward_only=forward_only,
            drop_states=drop_states,
            **cell_kwargs
        )
        cells.append(cell)

    cell = MultiRNNCell(
        cells,
        name=scope,
        forward_only=forward_only,
    ) if num_layers > 1 else cells[0]

    cell = (
        cell if static_rnn_unroll_size is None
        else UnrolledCell(cell, static_rnn_unroll_size))

    # outputs_with_grads argument indexes into final layer
    outputs_with_grads = [4 * (num_layers - 1) + i for i in outputs_with_grads]
    _, result = cell.apply_over_sequence(
        model=model,
        inputs=input_blob,
        seq_lengths=seq_lengths,
        initial_states=initial_states,
        outputs_with_grads=outputs_with_grads,
    )

    if return_last_layer_only:
        result = result[4 * (num_layers - 1):]
    if return_params:
        result = list(result) + [{
            'input': cell.get_input_params(),
            'recurrent': cell.get_recurrent_params(),
        }]
    return tuple(result)


LSTM = functools.partial(_LSTM, LSTMCell)
BasicRNN = functools.partial(_LSTM, BasicRNNCell)
MILSTM = functools.partial(_LSTM, MILSTMCell)
LayerNormLSTM = functools.partial(_LSTM, LayerNormLSTMCell)
LayerNormMILSTM = functools.partial(_LSTM, LayerNormMILSTMCell)


class UnrolledCell(RNNCell):
    def __init__(self, cell, T):
        self.T = T
        self.cell = cell

    def apply_over_sequence(
        self,
        model,
        inputs,
        seq_lengths,
        initial_states,
        outputs_with_grads=None,
    ):
        inputs = self.cell.prepare_input(model, inputs)

        # Now they are blob references - outputs of splitting the input sequence
        split_inputs = model.net.Split(
            inputs,
            [str(inputs) + "_timestep_{}".format(i)
             for i in range(self.T)],
            axis=0)
        if self.T == 1:
            split_inputs = [split_inputs]

        states = initial_states
        all_states = []
        for t in range(0, self.T):
            scope_name = "timestep_{}".format(t)
            # Parameters of all timesteps are shared
            with ParameterSharing({scope_name: ''}),\
                    scope.NameScope(scope_name):
                timestep = model.param_init_net.ConstantFill(
                    [], "timestep", value=t, shape=[1],
                    dtype=core.DataType.INT32,
                    device_option=core.DeviceOption(caffe2_pb2.CPU))
                states = self.cell._apply(
                    model=model,
                    input_t=split_inputs[t],
                    seq_lengths=seq_lengths,
                    states=states,
                    timestep=timestep,
                )
            all_states.append(states)

        all_states = zip(*all_states)
        all_states = [
            model.net.Concat(
                list(full_output),
                [
                    str(full_output[0])[len("timestep_0/"):] + "_concat",
                    str(full_output[0])[len("timestep_0/"):] + "_concat_info"

                ],
                axis=0)[0]
            for full_output in all_states
        ]
        # Interleave the state values similar to
        #
        #   x = [1, 3, 5]
        #   y = [2, 4, 6]
        #   z = [val for pair in zip(x, y) for val in pair]
        #   # z is [1, 2, 3, 4, 5, 6]
        #
        # and returns it as outputs
        outputs = tuple(
            state for state_pair in zip(all_states, states) for state in state_pair
        )
        outputs_without_grad = set(range(len(outputs))) - set(
            outputs_with_grads)
        for i in outputs_without_grad:
            model.net.ZeroGradient(outputs[i], [])
        logging.debug("Added 0 gradients for blobs:",
                      [outputs[i] for i in outputs_without_grad])

        final_output = self.cell._prepare_output_sequence(model, outputs)

        return final_output, outputs


def GetLSTMParamNames():
    weight_params = ["input_gate_w", "forget_gate_w", "output_gate_w", "cell_w"]
    bias_params = ["input_gate_b", "forget_gate_b", "output_gate_b", "cell_b"]
    return {'weights': weight_params, 'biases': bias_params}


def InitFromLSTMParams(lstm_pblobs, param_values):
    '''
    Set the parameters of LSTM based on predefined values
    '''
    weight_params = GetLSTMParamNames()['weights']
    bias_params = GetLSTMParamNames()['biases']
    for input_type in viewkeys(param_values):
        weight_values = [
            param_values[input_type][w].flatten()
            for w in weight_params
        ]
        wmat = np.array([])
        for w in weight_values:
            wmat = np.append(wmat, w)
        bias_values = [
            param_values[input_type][b].flatten()
            for b in bias_params
        ]
        bm = np.array([])
        for b in bias_values:
            bm = np.append(bm, b)

        weights_blob = lstm_pblobs[input_type]['weights']
        bias_blob = lstm_pblobs[input_type]['biases']
        cur_weight = workspace.FetchBlob(weights_blob)
        cur_biases = workspace.FetchBlob(bias_blob)

        workspace.FeedBlob(
            weights_blob,
            wmat.reshape(cur_weight.shape).astype(np.float32))
        workspace.FeedBlob(
            bias_blob,
            bm.reshape(cur_biases.shape).astype(np.float32))


def cudnn_LSTM(model, input_blob, initial_states, dim_in, dim_out,
               scope, recurrent_params=None, input_params=None,
               num_layers=1, return_params=False):
    '''
    CuDNN version of LSTM for GPUs.
    input_blob          Blob containing the input. Will need to be available
                        when param_init_net is run, because the sequence lengths
                        and batch sizes will be inferred from the size of this
                        blob.
    initial_states      tuple of (hidden_init, cell_init) blobs
    dim_in              input dimensions
    dim_out             output/hidden dimension
    scope               namescope to apply
    recurrent_params    dict of blobs containing values for recurrent
                        gate weights, biases (if None, use random init values)
                        See GetLSTMParamNames() for format.
    input_params        dict of blobs containing values for input
                        gate weights, biases (if None, use random init values)
                        See GetLSTMParamNames() for format.
    num_layers          number of LSTM layers
    return_params       if True, returns (param_extract_net, param_mapping)
                        where param_extract_net is a net that when run, will
                        populate the blobs specified in param_mapping with the
                        current gate weights and biases (input/recurrent).
                        Useful for assigning the values back to non-cuDNN
                        LSTM.
    '''
    with core.NameScope(scope):
        weight_params = GetLSTMParamNames()['weights']
        bias_params = GetLSTMParamNames()['biases']

        input_weight_size = dim_out * dim_in
        upper_layer_input_weight_size = dim_out * dim_out
        recurrent_weight_size = dim_out * dim_out
        input_bias_size = dim_out
        recurrent_bias_size = dim_out

        def init(layer, pname, input_type):
            input_weight_size_for_layer = input_weight_size if layer == 0 else \
                upper_layer_input_weight_size
            if pname in weight_params:
                sz = input_weight_size_for_layer if input_type == 'input' \
                    else recurrent_weight_size
            elif pname in bias_params:
                sz = input_bias_size if input_type == 'input' \
                    else recurrent_bias_size
            else:
                assert False, "unknown parameter type {}".format(pname)
            return model.param_init_net.UniformFill(
                [],
                "lstm_init_{}_{}_{}".format(input_type, pname, layer),
                shape=[sz])

        # Multiply by 4 since we have 4 gates per LSTM unit
        first_layer_sz = input_weight_size + recurrent_weight_size + \
                         input_bias_size + recurrent_bias_size
        upper_layer_sz = upper_layer_input_weight_size + \
                         recurrent_weight_size + input_bias_size + \
                         recurrent_bias_size
        total_sz = 4 * (first_layer_sz + (num_layers - 1) * upper_layer_sz)

        weights = model.create_param(
            'lstm_weight',
            shape=[total_sz],
            initializer=Initializer('UniformFill'),
            tags=ParameterTags.WEIGHT,
        )

        lstm_args = {
            'hidden_size': dim_out,
            'rnn_mode': 'lstm',
            'bidirectional': 0,  # TODO
            'dropout': 1.0,  # TODO
            'input_mode': 'linear',  # TODO
            'num_layers': num_layers,
            'engine': 'CUDNN'
        }

        param_extract_net = core.Net("lstm_param_extractor")
        param_extract_net.AddExternalInputs([input_blob, weights])
        param_extract_mapping = {}

        # Populate the weights-blob from blobs containing parameters for
        # the individual components of the LSTM, such as forget/input gate
        # weights and bises. Also, create a special param_extract_net that
        # can be used to grab those individual params from the black-box
        # weights blob. These results can be then fed to InitFromLSTMParams()
        for input_type in ['input', 'recurrent']:
            param_extract_mapping[input_type] = {}
            p = recurrent_params if input_type == 'recurrent' else input_params
            if p is None:
                p = {}
            for pname in weight_params + bias_params:
                for j in range(0, num_layers):
                    values = p[pname] if pname in p else init(j, pname, input_type)
                    model.param_init_net.RecurrentParamSet(
                        [input_blob, weights, values],
                        weights,
                        layer=j,
                        input_type=input_type,
                        param_type=pname,
                        **lstm_args
                    )
                    if pname not in param_extract_mapping[input_type]:
                        param_extract_mapping[input_type][pname] = {}
                    b = param_extract_net.RecurrentParamGet(
                        [input_blob, weights],
                        ["lstm_{}_{}_{}".format(input_type, pname, j)],
                        layer=j,
                        input_type=input_type,
                        param_type=pname,
                        **lstm_args
                    )
                    param_extract_mapping[input_type][pname][j] = b

        (hidden_input_blob, cell_input_blob) = initial_states
        output, hidden_output, cell_output, rnn_scratch, dropout_states = \
            model.net.Recurrent(
                [input_blob, hidden_input_blob, cell_input_blob, weights],
                ["lstm_output", "lstm_hidden_output", "lstm_cell_output",
                 "lstm_rnn_scratch", "lstm_dropout_states"],
                seed=random.randint(0, 100000),  # TODO: dropout seed
                **lstm_args
            )
        model.net.AddExternalOutputs(
            hidden_output, cell_output, rnn_scratch, dropout_states)

    if return_params:
        param_extract = param_extract_net, param_extract_mapping
        return output, hidden_output, cell_output, param_extract
    else:
        return output, hidden_output, cell_output


def LSTMWithAttention(
    model,
    decoder_inputs,
    decoder_input_lengths,
    initial_decoder_hidden_state,
    initial_decoder_cell_state,
    initial_attention_weighted_encoder_context,
    encoder_output_dim,
    encoder_outputs,
    encoder_lengths,
    decoder_input_dim,
    decoder_state_dim,
    scope,
    attention_type=AttentionType.Regular,
    outputs_with_grads=(0, 4),
    weighted_encoder_outputs=None,
    lstm_memory_optimization=False,
    attention_memory_optimization=False,
    forget_bias=0.0,
    forward_only=False,
):
    '''
    Adds a LSTM with attention mechanism to a model.

    The implementation is based on https://arxiv.org/abs/1409.0473, with
    a small difference in the order
    how we compute new attention context and new hidden state, similarly to
    https://arxiv.org/abs/1508.04025.

    The model uses encoder-decoder naming conventions,
    where the decoder is the sequence the op is iterating over,
    while computing the attention context over the encoder.

    model: ModelHelper object new operators would be added to

    decoder_inputs: the input sequence in a format T x N x D
    where T is sequence size, N - batch size and D - input dimension

    decoder_input_lengths: blob containing sequence lengths
    which would be passed to LSTMUnit operator

    initial_decoder_hidden_state: initial hidden state of LSTM

    initial_decoder_cell_state: initial cell state of LSTM

    initial_attention_weighted_encoder_context: initial attention context

    encoder_output_dim: dimension of encoder outputs

    encoder_outputs: the sequence, on which we compute the attention context
    at every iteration

    encoder_lengths: a tensor with lengths of each encoder sequence in batch
    (may be None, meaning all encoder sequences are of same length)

    decoder_input_dim: input dimension (last dimension on decoder_inputs)

    decoder_state_dim: size of hidden states of LSTM

    attention_type: One of: AttentionType.Regular, AttentionType.Recurrent.
    Determines which type of attention mechanism to use.

    outputs_with_grads : position indices of output blobs which will receive
    external error gradient during backpropagation

    weighted_encoder_outputs: encoder outputs to be used to compute attention
    weights. In the basic case it's just linear transformation of
    encoder outputs (that the default, when weighted_encoder_outputs is None).
    However, it can be something more complicated - like a separate
    encoder network (for example, in case of convolutional encoder)

    lstm_memory_optimization: recompute LSTM activations on backward pass, so
                 we don't need to store their values in forward passes

    attention_memory_optimization: recompute attention for backward pass

    forward_only: whether to create only forward pass
    '''
    cell = LSTMWithAttentionCell(
        encoder_output_dim=encoder_output_dim,
        encoder_outputs=encoder_outputs,
        encoder_lengths=encoder_lengths,
        decoder_input_dim=decoder_input_dim,
        decoder_state_dim=decoder_state_dim,
        name=scope,
        attention_type=attention_type,
        weighted_encoder_outputs=weighted_encoder_outputs,
        forget_bias=forget_bias,
        lstm_memory_optimization=lstm_memory_optimization,
        attention_memory_optimization=attention_memory_optimization,
        forward_only=forward_only,
    )
    initial_states = [
        initial_decoder_hidden_state,
        initial_decoder_cell_state,
        initial_attention_weighted_encoder_context,
    ]
    if attention_type == AttentionType.SoftCoverage:
        initial_states.append(cell.build_initial_coverage(model))
    _, result = cell.apply_over_sequence(
        model=model,
        inputs=decoder_inputs,
        seq_lengths=decoder_input_lengths,
        initial_states=initial_states,
        outputs_with_grads=outputs_with_grads,
    )
    return result


def _layered_LSTM(
        model, input_blob, seq_lengths, initial_states,
        dim_in, dim_out, scope, outputs_with_grads=(0,), return_params=False,
        memory_optimization=False, forget_bias=0.0, forward_only=False,
        drop_states=False, create_lstm=None):
    params = locals()  # leave it as a first line to grab all params
    params.pop('create_lstm')
    if not isinstance(dim_out, list):
        return create_lstm(**params)
    elif len(dim_out) == 1:
        params['dim_out'] = dim_out[0]
        return create_lstm(**params)

    assert len(dim_out) != 0, "dim_out list can't be empty"
    assert return_params is False, "return_params not supported for layering"
    for i, output_dim in enumerate(dim_out):
        params.update({
            'dim_out': output_dim
        })
        output, last_output, all_states, last_state = create_lstm(**params)
        params.update({
            'input_blob': output,
            'dim_in': output_dim,
            'initial_states': (last_output, last_state),
            'scope': scope + '_layer_{}'.format(i + 1)
        })
    return output, last_output, all_states, last_state


layered_LSTM = functools.partial(_layered_LSTM, create_lstm=LSTM)