File: gradient.py

package info (click to toggle)
theano 1.0.3+dfsg-1
  • links: PTS, VCS
  • area: main
  • in suites: buster, sid
  • size: 30,752 kB
  • sloc: python: 141,182; ansic: 9,505; makefile: 259; sh: 214; pascal: 81
file content (2259 lines) | stat: -rw-r--r-- 87,073 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
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
"""Driver for gradient calculations."""
from __future__ import absolute_import, print_function, division
from collections import OrderedDict
import six.moves.builtins as builtins
import logging
import time
import warnings

import numpy as np  # for numeric_grad
from six import itervalues

import theano

from theano import gof
from theano.gof import utils, Variable
from theano.compat import izip
from six.moves import xrange, reduce
from theano.gof.null_type import NullType, null_type
from theano.gof.op import get_debug_values
from theano.compile import ViewOp, FAST_RUN, DebugMode, get_mode

__authors__ = "James Bergstra, Razvan Pascanu, Arnaud Bergeron, Ian Goodfellow"
__copyright__ = "(c) 2011, Universite de Montreal"
__license__ = "3-clause BSD License"
__contact__ = "theano-dev <theano-dev@googlegroups.com>"

__docformat__ = "restructuredtext en"
_logger = logging.getLogger('theano.gradient')

# we can't do "import theano.tensor"
# tensor depends on theano.compile
# theano.compile depends on theano.gradient (this file)
# the reason theano.compile depends on theano.gradient
# is that theano.compile.builders contains the op from graph
# functionality and it uses theano.gradient to implement
# the new op's grad method
tensor = None

_msg_retType = 'op.grad(...) returned a non-list'

grad_time = 0


def format_as(use_list, use_tuple, outputs):
    """
    Formats the outputs according to the flags `use_list` and `use_tuple`.

    If `use_list` is True, `outputs` is returned as a list (if `outputs`
    is not a list or a tuple then it is converted in a one element list).
    If `use_tuple` is True, `outputs` is returned as a tuple (if `outputs`
    is not a list or a tuple then it is converted into a one element tuple).
    Otherwise (if both flags are false), `outputs` is returned.
    """
    assert not (use_list and use_tuple), \
        "Both flags cannot be simultaneously True"
    if (use_list or use_tuple) and not isinstance(outputs, (list, tuple)):
        if use_list:
            return [outputs]
        else:
            return (outputs,)
    elif not (use_list or use_tuple) and isinstance(outputs, (list, tuple)):
        assert len(outputs) == 1, \
            "Wrong arguments. Expected a one element list"
        return outputs[0]
    elif use_list or use_tuple:
        if use_list:
            return list(outputs)
        else:
            return tuple(outputs)
    else:
        return outputs


def grad_not_implemented(op, x_pos, x, comment=""):
    """
    Return an un-computable symbolic variable of type `x.type`.

    If any call to tensor.grad results in an expression containing this
    un-computable variable, an exception (NotImplementedError) will be
    raised indicating that the gradient on the
    `x_pos`'th input of `op` has not been implemented. Likewise if
    any call to theano.function involves this variable.

    Optionally adds a comment to the exception explaining why this
    gradient is not implemented.
    """

    return (NullType((
        "This variable is Null because the grad method for "
        "input %s (%s) of the %s op is not implemented. %s"
    ) % (x_pos, x, op, comment)))()


def grad_undefined(op, x_pos, x, comment=""):
    """
    Return an un-computable symbolic variable of type `x.type`.

    If any call to tensor.grad results in an expression containing this
    un-computable variable, an exception (GradUndefinedError) will be
    raised indicating that the gradient on the
    `x_pos`'th input of `op` is mathematically undefined. Likewise if
    any call to theano.function involves this variable.

    Optionally adds a comment to the exception explaining why this
    gradient is not defined.
    """

    return (NullType(
        (
            "This variable is Null because the grad method for "
            "input %s (%s) of the %s op is mathematically undefined. %s"
        ) % (x_pos, x, op, comment)))()


class DisconnectedType(theano.gof.type.Type):

    """ A type indicating that a variable is a result
        of taking the gradient of c with respect to x
        when c is not a function of x.
        A symbolic placeholder for 0, but to convey
        the extra information that this gradient is 0
        because it is disconnected.
    """

    def filter(self, data, strict=False, allow_downcast=None):
        raise AssertionError(
            (
                "If you're assigning to a DisconnectedType you're"
                " doing something wrong. It should only be used as"
                " a symbolic placeholder."
            ))

    def fiter_variable(self, other):
        raise AssertionError(
            (
                "If you're assigning to a DisconnectedType you're"
                " doing something wrong. It should only be used as"
                " a symbolic placeholder."
            ))

    def may_share_memory(a, b):
        return False

    def value_eq(a, b, force_same_dtype=True):
        raise AssertionError(
            (
                "If you're assigning to a DisconnectedType you're"
                " doing something wrong. It should only be used as"
                " a symbolic placeholder."
            ))

    def __str__(self):
        return 'DisconnectedType'


disconnected_type = DisconnectedType()


########################
# R Operator
########################


def Rop(f, wrt, eval_points, disconnected_outputs="raise",
        return_disconnected="zero"):
    """
    Computes the R operation on `f` wrt to `wrt` at `eval_points`.

    Mathematically this stands for the jacobian of `f` wrt
    to `wrt` right muliplied by the eval points.

    Parameters
    ----------
    f : :class:`~theano.gof.graph.Variable` or list of Variables
        `f` stands for the output of the computational graph to which you
        want to apply the R operator
    wrt : :class:`~theano.gof.graph.Variable` or list of Variables
        variables for which you compute the R operator of the expression
        described by `f`
    eval_points : :class:`~theano.gof.graph.Variable` or list of Variables
        evalutation points for each of the variables in `wrt`
    disconnected_outputs : str
        Defines the behaviour if some of the variables in `f`
        have no dependency on any of the variable in `wrt` (or if
        all links are non-differentiable). The possible values are:

        - 'ignore': considers that the gradient on these parameters is zero.
        - 'warn': consider the gradient zero, and print a warning.
        - 'raise': raise DisconnectedInputError.

    return_disconnected : {'zero', 'None', 'Disconnected'}
        - 'zero' : If wrt[i] is disconnected, return value i will be
          wrt[i].zeros_like()
        - 'None' : If wrt[i] is disconnected, return value i will be
          None
        - 'Disconnected' : returns variables of type DisconnectedType

    Returns
    -------
    :class:`~theano.gof.graph.Variable` or list/tuple of Variables depending on type of f
        Symbolic expression such that
        R_op[i] = sum_j (d f[i] / d wrt[j]) eval_point[j]
        where the indices in that expression are magic multidimensional
        indices that specify both the position within a list and all
        coordinates of the tensor element in the last.
        If `wrt` is a list/tuple, then return a list/tuple with the results.
    """
    from theano.tensor import as_tensor_variable
    using_list = isinstance(f, list)
    using_tuple = isinstance(f, tuple)
    if not isinstance(wrt, (list, tuple)):
        wrt = [wrt]

    if not isinstance(eval_points, (list, tuple)):
        eval_points = [eval_points]

    if not isinstance(f, (list, tuple)):
        f = [f]

    assert len(wrt) == len(eval_points)

    # Check that each element of wrt corresponds to an element
    # of eval_points with the same dimensionality.
    for pack in enumerate(zip(wrt, eval_points)):
        i = pack[0]
        wrt_elem, eval_point = pack[1]
        if not isinstance(wrt_elem, gof.Variable):
            wrt_elem = as_tensor_variable(wrt_elem)
        if not isinstance(eval_point, gof.Variable):
            eval_point = as_tensor_variable(eval_point)

        try:

            if wrt_elem.type.ndim != eval_point.type.ndim:
                raise ValueError('Element ' +
                                 str(i) +
                                 ' of wrt/eval_point have mismatched ' +
                                 'dimensionality: ' +
                                 str(wrt_elem.type.ndim) +
                                 ' versus ' +
                                 str(eval_point.type.ndim))
        except AttributeError:
            # wrt_elem and eval_point don't always have ndim like random type
            # Tensor, Sparse and GpuArray have the ndim attribute
            pass

    seen_nodes = OrderedDict()

    def _traverse(node):
        """ TODO: writeme """

        if node is None:
            return

        op = node.op
        inputs = node.inputs

        # Compute the evaluation points corresponding to each of the
        # inputs of the node
        local_eval_points = []
        for inp in inputs:
            if inp in wrt:
                local_eval_points.append(eval_points[wrt.index(inp)])
            elif inp.owner is None:
                try:
                    local_eval_points.append(inp.zeros_like())
                except Exception:
                    # None should be used for non-differentiable
                    # arguments, like for example random states
                    local_eval_points.append(None)
            elif inp.owner in seen_nodes:

                local_eval_points.append(
                    seen_nodes[inp.owner][inp.owner.outputs.index(inp)])

            else:
                # We actually need to compute the R_op for this node

                _traverse(inp.owner)
                local_eval_points.append(
                    seen_nodes[inp.owner][inp.owner.outputs.index(inp)])
        same_type_eval_points = []
        for x, y in zip(inputs, local_eval_points):
            if y is not None:
                if not isinstance(x, gof.Variable):
                    x = as_tensor_variable(x)
                if not isinstance(y, gof.Variable):
                    y = as_tensor_variable(y)
                try:
                    y = x.type.filter_variable(y)
                except TypeError:
                    # This is a hack
                    # Originally both grad and Rop were written
                    # with the assumption that a variable and the
                    # gradient wrt that variable would have the same
                    # dtype. This was a bad assumption because the
                    # gradient wrt an integer can take on non-integer
                    # values.
                    # grad is now fixed, but Rop is not, so when grad
                    # does the right thing and violates this assumption
                    # we have to make it be wrong for Rop to keep working
                    # Rop should eventually be upgraded to handle integers
                    # correctly, the same as grad
                    y = theano.tensor.cast(y, x.type.dtype)
                    y = x.type.filter_variable(y)
                assert x.type == y.type
                same_type_eval_points.append(y)
            else:
                same_type_eval_points.append(y)

        seen_nodes[node] = op.R_op(node.inputs, same_type_eval_points)
    # end _traverse

    # Populate the dictionary
    for out in f:
        _traverse(out.owner)

    rval = []
    for out in f:
        if out in wrt:
            rval.append(eval_points[wrt.index(out)])
        elif seen_nodes.get(out.owner, None) is None or \
                seen_nodes[out.owner][out.owner.outputs.index(out)] is None:
            message = ("Rop method was asked to compute the gradient "
                       "with respect to a variable that is not part of "
                       "the computational graph of variables in wrt, or is "
                       "used only by a non-differentiable operator: %s" % out)
            if disconnected_outputs == 'ignore':
                pass
            elif disconnected_outputs == 'warn':
                warnings.warn(message, stacklevel=2)
            elif disconnected_outputs == 'raise':
                message = utils.get_variable_trace_string(out)
                raise DisconnectedInputError(message)
            else:
                raise ValueError("Invalid value for keyword "
                                 "'disconnected_inputs', valid values are "
                                 "'ignore', 'warn' and 'raise'.")
            if return_disconnected.lower() == "zero":
                rval.append(tensor.zeros_like(out))
            elif return_disconnected.lower() == "none":
                rval.append(None)
            elif return_disconnected.lower() == "disconnected":
                rval.append(disconnected_type())
            else:
                raise ValueError("Invalid value for keyword "
                                 "'return_disconnected', valid values are "
                                 "'zero', 'None' and 'Disconnected'.")
        else:
            rval.append(seen_nodes[out.owner][out.owner.outputs.index(out)])

    return format_as(using_list, using_tuple, rval)


def Lop(f, wrt, eval_points, consider_constant=None,
        disconnected_inputs='raise'):
    """
    Computes the L operation on `f` wrt to `wrt` at `eval_points`.

    Mathematically this stands for the jacobian of `f` wrt
    to `wrt` left muliplied by the eval points.

    Parameters
    ----------
    f : :class:`~theano.gof.graph.Variable` or list of Variables
        `f` stands for the output of the computational graph to which you
        want to apply the L operator
    wrt : :class:`~theano.gof.graph.Variable` or list of Variables
        variables for which you compute the L operator of the expression
        described by `f`
    eval_points : :class:`~theano.gof.graph.Variable` or list of Variables
        evalutation points for each of the variables in `f`

    Returns
    -------
    :class:`~theano.gof.Variable` or list/tuple of Variables depending on type of f
        Symbolic expression such that
        L_op[i] = sum_i (d f[i] / d wrt[j]) eval_point[i]
        where the indices in that expression are magic multidimensional
        indices that specify both the position within a list and all
        coordinates of the tensor element in the last
        If `f` is a list/tuple, then return a list/tuple with the results.
    """
    if type(eval_points) not in (list, tuple):
        eval_points = [eval_points]

    using_list = isinstance(wrt, list)
    using_tuple = isinstance(wrt, tuple)

    if not isinstance(f, (list, tuple)):
        f = [f]

    # make copies of f and grads so we don't modify the client's copy
    f = list(f)
    grads = list(eval_points)

    if not isinstance(wrt, (list, tuple)):
        wrt = [wrt]

    assert len(f) == len(grads)
    known = OrderedDict(izip(f, grads))

    ret = grad(cost=None, known_grads=known,
               consider_constant=consider_constant, wrt=wrt,
               disconnected_inputs=disconnected_inputs)

    return format_as(using_list, using_tuple, ret)


#########################
# Gradient
#########################

def grad(cost, wrt, consider_constant=None,
         disconnected_inputs='raise', add_names=True,
         known_grads=None, return_disconnected='zero',
         null_gradients='raise'):
    """
    Return symbolic gradients of one cost with respect to one or more variables.

    For more information about how automatic differentiation works in Theano,
    see :mod:`gradient`. For information on how to implement the gradient of
    a certain Op, see :func:`grad`.

    Parameters
    ----------
    cost : :class:`~theano.gof.graph.Variable` scalar (0-dimensional) tensor variable or ``None``
        Value that we are differentiating (that we want the gradient of).
        May be `None` if `known_grads` is provided.
    wrt : :class:`~theano.gof.graph.Variable` or list of Variables
        Term[s] with respect to which we want gradients
    consider_constant : list of variables
        Expressions not to backpropagate through
    disconnected_inputs : {'ignore', 'warn', 'raise'}
        Defines the behaviour if some of the variables in `wrt` are
        not part of the computational graph computing `cost` (or if
        all links are non-differentiable). The possible values are:

        - 'ignore': considers that the gradient on these parameters is zero.
        - 'warn': consider the gradient zero, and print a warning.
        - 'raise': raise DisconnectedInputError.
    add_names : bool
        If True, variables generated by grad will be named
        (d<cost.name>/d<wrt.name>) provided that both cost and wrt
        have names
    known_grads : OrderedDict, optional
        A ordered dictionary mapping variables to their gradients. This is
        useful in the case where you know the gradient on some
        variables but do not know the original cost.
    return_disconnected : {'zero', 'None', 'Disconnected'}
        - 'zero' : If wrt[i] is disconnected, return value i will be
          wrt[i].zeros_like()
        - 'None' : If wrt[i] is disconnected, return value i will be
          None
        - 'Disconnected' : returns variables of type DisconnectedType
    null_gradients : {'raise', 'return'}
        Defines the behaviour if some of the variables in `wrt` have a
        null gradient. The possibles values are:

        - 'raise' : raise a NullTypeGradError exception
        - 'return' : return the null gradients

    Returns
    -------
    variable or list/tuple of variables (matches `wrt`)
        Symbolic expression of gradient of `cost` with respect to each
        of the `wrt` terms.  If an element of `wrt` is not
        differentiable with respect to the output, then a zero
        variable is returned.

    """
    t0 = time.time()
    global tensor
    if tensor is None:
        from theano import tensor

    if cost is None:
        if known_grads is None:
            raise AssertionError("cost and known_grads can't both be None.")

    if cost is not None and isinstance(cost.type, NullType):
        raise ValueError("Can't differentiate a NaN cost."
                         "cost is NaN because " +
                         cost.type.why_null)

    if cost is not None and cost.ndim != 0:
        raise TypeError("cost must be a scalar.")

    if isinstance(wrt, set):
        raise TypeError("wrt must not be a set. sets have no defined "
                        "iteration order, so we can't return gradients in a"
                        "  matching order.")

    using_list = isinstance(wrt, list)
    using_tuple = isinstance(wrt, tuple)
    if not using_list and not using_tuple:
        wrt = [wrt]

    for elem in wrt:
        if not isinstance(elem, Variable):
            raise TypeError("Expected Variable, got " + str(elem) +
                            " of type " + str(type(elem)))

    outputs = []
    if cost is not None:
        outputs.append(cost)
    if known_grads is not None:
        outputs.extend(list(known_grads.keys()))

    var_to_app_to_idx = _populate_var_to_app_to_idx(
        outputs, wrt, consider_constant)

    # build a dict mapping var to the gradient of cost with respect to var
    grad_dict = OrderedDict()

    if known_grads is None:
        known_grads = OrderedDict()
    else:
        m = "known_grads must be an OrderedDict. "
        assert isinstance(known_grads, OrderedDict) or len(known_grads) <= 1, m

    # The gradient of the cost is 1 unless specified otherwise by known_grads.
    if cost is not None:
        if cost in known_grads:
            g_cost = known_grads[cost]
        else:
            g_cost = _float_ones_like(cost)
        # g_cost may be Disconnected or NullType. A creative use of the
        # function, sure, but nonetheless one we can and should support.
        # So before we try to cast it make sure it even has a dtype
        if (hasattr(g_cost.type, 'dtype') and
                cost.type.dtype in tensor.continuous_dtypes):
                # Here we enforce the constraint that floating point variables
                # have the same dtype as their gradient.
                g_cost = g_cost.astype(cost.type.dtype)
        # DO NOT enforce g_cost to be 0 if cost is an integer.
        # This is to be enforced by the Op.grad method for the
        # Op that outputs cost.
        if hasattr(g_cost.type, 'dtype'):
            assert g_cost.type.dtype in tensor.continuous_dtypes

        grad_dict[cost] = g_cost

    for var in known_grads:
        g_var = known_grads[var]

        if not hasattr(g_var, 'type'):
            raise TypeError('output grads must be theano variables.'
                            'Ambiguous whether %s should be made into tensor'
                            ' or sparse theano variable' % str(type(g_var)))

        if (not isinstance(g_var.type, (NullType, DisconnectedType)) and
                'float' not in str(g_var.type.dtype)):
            raise TypeError("Gradients must always be NullType, "
                            "DisconnectedType, or continuous, but grad was "
                            "given a known_grad of type " + str(g_var.type))

        # DO NOT check that these gradients are equal to 0 if var is int
        # The gradient is allowed to be non-zero on var in that case
        # Ops outputing var should not backpropagate its gradient further
        # but that is enforced elsewhere (grep for only_connected_to_int)

        grad_dict[var] = g_var

    def handle_disconnected(var):
            message = ("grad method was asked to compute the gradient "
                       "with respect to a variable that is not part of "
                       "the computational graph of the cost, or is used "
                       "only by a non-differentiable operator: %s" % var)
            if disconnected_inputs == 'ignore':
                pass
            elif disconnected_inputs == 'warn':
                warnings.warn(message, stacklevel=2)
            elif disconnected_inputs == 'raise':
                message = utils.get_variable_trace_string(var)
                raise DisconnectedInputError(message)
            else:
                raise ValueError("Invalid value for keyword "
                                 "'disconnected_inputs', valid values are "
                                 "'ignore', 'warn' and 'raise'.")

    # variables that do not influence the cost have zero gradient.
    # if wrt is such a variable, populate the grad_dict with this info
    # so that wrt not being in var_to_app_to_idx won't cause an error below
    # according to the flag, possibly raise an error if wrt is disconnected
    for elem in wrt:
        if elem not in var_to_app_to_idx and elem is not cost \
                and elem not in grad_dict:
            handle_disconnected(elem)
            grad_dict[elem] = disconnected_type()

    cost_name = None
    if add_names and cost is not None:
        cost_name = cost.name

    # Make sure we didn't initialize the grad_dict with any ints
    # The gradient may NEVER be an int, even if the variable is an int.
    # Read the Op contract and talk to Ian Goodfellow before changing this!
    for var in grad_dict:
        g = grad_dict[var]
        if hasattr(g.type, 'dtype'):
            assert g.type.dtype in tensor.float_dtypes

    rval = _populate_grad_dict(var_to_app_to_idx,
                               grad_dict, wrt, cost_name)

    for i in xrange(len(rval)):
        if isinstance(rval[i].type, NullType):
            if null_gradients == 'raise':
                raise NullTypeGradError("tensor.grad encountered a NaN. " +
                                        rval[i].type.why_null)
            else:
                assert null_gradients == 'return'
        if isinstance(rval[i].type, DisconnectedType):
            handle_disconnected(rval[i])
            if return_disconnected == 'zero':
                rval[i] = _float_zeros_like(wrt[i])
            elif return_disconnected == 'None':
                rval[i] = None
            else:
                assert return_disconnected == 'Disconnected'

    if using_tuple:
        rval = tuple(rval)
    elif not using_list:
        rval, = rval
    t1 = time.time()
    global grad_time
    grad_time += t1 - t0
    return rval


def subgraph_grad(wrt, end, start=None, cost=None, details=False):
    '''
    With respect to `wrt`, computes gradients of cost and/or from
    existing `start` gradients, up to the `end` variables of a
    symbolic digraph.  In other words, computes gradients for a
    subgraph of the symbolic theano function. Ignores all disconnected
    inputs.

    This can be useful when one needs to perform the gradient descent
    iteratively (e.g. one layer at a time in an MLP), or when a
    particular operation is not differentiable in theano
    (e.g. stochastic sampling from a multinomial). In the latter case,
    the gradient of the non-differentiable process could be
    approximated by user-defined formula, which could be calculated
    using the gradients of a cost with respect to samples (0s and
    1s). These gradients are obtained by performing a subgraph_grad
    from the `cost` or previously known gradients (`start`) up to the
    outputs of the stochastic process (`end`).  A dictionary mapping
    gradients obtained from the user-defined differentiation of the
    process, to variables, could then be fed into another
    subgraph_grad as `start` with any other `cost` (e.g. weight
    decay).

    In an MLP, we could use subgraph_grad to iteratively backpropagate:

    .. code-block:: python

        x, t = theano.tensor.fvector('x'), theano.tensor.fvector('t')
        w1 = theano.shared(np.random.randn(3,4))
        w2 = theano.shared(np.random.randn(4,2))
        a1 = theano.tensor.tanh(theano.tensor.dot(x,w1))
        a2 = theano.tensor.tanh(theano.tensor.dot(a1,w2))
        cost2 = theano.tensor.sqr(a2 - t).sum()
        cost2 += theano.tensor.sqr(w2.sum())
        cost1 = theano.tensor.sqr(w1.sum())

        params = [[w2],[w1]]
        costs = [cost2,cost1]
        grad_ends = [[a1], [x]]

        next_grad = None
        param_grads = []
        for i in xrange(2):
            param_grad, next_grad = theano.subgraph_grad(
                wrt=params[i], end=grad_ends[i],
                start=next_grad, cost=costs[i]
            )
            next_grad = dict(zip(grad_ends[i], next_grad))
            param_grads.extend(param_grad)

    Parameters
    ----------

    wrt : list of variables
        Gradients are computed with respect to `wrt`.

    end : list of variables
        Theano variables at which to end gradient descent (they are
        considered constant in theano.grad).  For convenience, the
        gradients with respect to these variables are also returned.

    start : dictionary of variables
        If not None, a dictionary mapping variables to their
        gradients. This is useful when the gradient on some variables
        are known. These are used to compute the gradients backwards up
        to the variables in `end` (they are used as known_grad in
        theano.grad).

    cost : :class:`~theano.gof.Variable` scalar (0-dimensional) variable
        Additional costs for which to compute the gradients.  For
        example, these could be weight decay, an l1 constraint, MSE,
        NLL, etc. May optionally be None if start is provided.

        .. warning::

            If the gradients of `cost` with respect to any of the `start`
            variables is already part of the `start` dictionary, then it
            may be counted twice with respect to `wrt` and `end`.

    details : bool
        When True, additionally returns the list of gradients from
        `start` and of `cost`, respectively, with respect to `wrt` (not
        `end`).

    Returns
    -------
    Tuple of 2 or 4 Lists of Variables
        Returns lists of gradients with respect to `wrt` and `end`,
        respectively.


    .. versionadded:: 0.7
    '''
    assert ((cost is not None) or (start is not None))
    assert isinstance(end, list)
    assert isinstance(wrt, list)
    if start is not None:
        assert isinstance(start, dict)

    params = list(set(wrt + end))

    start_grads = None
    cost_grads = None
    if start is not None:
        start_grads = list(
            theano.grad(
                cost=None, wrt=params, known_grads=start,
                consider_constant=end,
                disconnected_inputs='ignore'
            )
        )

    if cost is not None:
        cost_grads = list(
            theano.grad(
                cost=cost, wrt=params,
                consider_constant=end,
                disconnected_inputs='ignore'
            )
        )

    grads = None
    if start is None:
        grads = cost_grads
    else:
        grads = start_grads
        if cost_grads is not None:
            for i in range(len(grads)):
                grads[i] += cost_grads[i]

    pgrads = OrderedDict(izip(params, grads))
    # separate wrt from end grads:
    wrt_grads = list(pgrads[k] for k in wrt)
    end_grads = list(pgrads[k] for k in end)

    if details:
        return wrt_grads, end_grads, start_grads, cost_grads

    return wrt_grads, end_grads


def _node_to_pattern(node):
    """ given an apply node, obtain its connection pattern
     this is just a wrapper around Op.connection_pattern
     that does type checking and supplies the default value
     if the method is not implemented
    """

    if hasattr(node.op, 'connection_pattern'):
        connection_pattern = node.op.connection_pattern(node)

        if not isinstance(connection_pattern, list):
            raise TypeError(
                "Op.connection_pattern should return " +
                ("list of list of bool, but for Op=%s" % node.op) +
                "got %s with type %s." % (connection_pattern,
                                          type(connection_pattern)))
        if len(connection_pattern) != len(node.inputs):
            raise ValueError(
                '%s.connection_pattern should have %d' %
                (node.op, len(node.inputs)) + ' rows but has %d.' %
                len(connection_pattern))
        for ii, output_pattern in enumerate(connection_pattern):
            if not isinstance(output_pattern, list):
                raise TypeError(
                    '%s.connection_pattern should return' %
                    node.op + ' a list of lists, but element %d' % ii +
                    'is %s of type %s.' % (output_pattern,
                                           type(output_pattern)))
    else:
        connection_pattern = [[True for output in node.outputs]
                              for ipt in node.inputs]
    assert isinstance(connection_pattern, list)
    assert len(connection_pattern) == len(node.inputs)
    for ii in xrange(len(node.inputs)):
        assert isinstance(connection_pattern[ii], list)
        assert len(connection_pattern[ii]) == len(node.outputs)
    return connection_pattern


def _populate_var_to_app_to_idx(outputs, wrt, consider_constant):
    """
    Helper function for grad function.

    Parameters
    ----------
    outputs
        a list of variables we want to take gradients of

    wrt
        a list of variables we want to take the gradient with
        respect to.

    consider_constant
        a list of variables not to backpropagate through.

    Returns
    -------
    var_to_app_to_idx:
        A dictionary mapping a variable to a second dictionary.
        The second dictionary maps apply nodes acting on this
        variable to the variable's index in the apply node's
        input list.

        This dictionary will only contain variables that
        meet two criteria:

        1) The elements of at least one output are a
           function of the elements of the variable

        2) The elements of the variable are a function of the
           elements of at least one member of wrt.

    This set is exactly the set of variables that connect
    the variables in wrt to the cost being differentiated.

    (A variable in consider_constant is not a function of
    anything)

    """

    # Validate and format consider_constant
    if consider_constant is None:
        consider_constant = []
    else:
        # error checking on consider_constant: verify that it is a collection
        # of theano variables
        # this is important, if someone accidentally passes a nested data
        # structure with theano variables at the leaves, only the root will
        # be properly considered constant
        try:
            iter(consider_constant)
        except TypeError:
            raise TypeError('consider_constant must be an iterable collection,'
                            ' got ' + str(type(consider_constant)))
        for elem in consider_constant:
            if not isinstance(elem, gof.Variable):
                raise TypeError('Elements of consider_constant must be '
                                'variables, but got ' + str(type(elem)))

    # var_to_app_to_idx[var][node] = [i,j] means node has
    # var as input at positions i and j
    var_to_app_to_idx = OrderedDict()

    # Set of variables that have been added to their true parents
    # ('true' here means that the elements of the variable are a function
    #  of the elements of the parent, according to the op's
    #  connection_pattern)
    # Note: we need to revisit the apply nodes repeatedly, because
    #       different outputs of the apply node are connected to
    #       different subsets of the inputs.
    accounted_for = set([])

    def account_for(var):
        # Don't visit the same variable twice
        if var in accounted_for:
            return
        accounted_for.add(var)

        # Constants are not a function of anything
        if var in consider_constant:
            return

        # Recursively add the variables that this variable is
        # a function of.
        if var.owner is not None:
            app = var.owner

            connection_pattern = _node_to_pattern(app)

            var_idx = app.outputs.index(var)

            for i, ipt in enumerate(app.inputs):

                # don't process ipt if it is not a true
                # parent of var
                if not connection_pattern[i][var_idx]:
                    continue

                if ipt not in var_to_app_to_idx:
                    # This object here *must* be an OrderedDict, because
                    # we iterate over its keys when adding up the terms of the
                    # gradient on ipt. If it is a regular dict, the grad method
                    # will return something that is analytically correct, but
                    # whose order of doing additions depends on the memory
                    # location of the apply nodes.
                    var_to_app_to_idx[ipt] = OrderedDict()
                app_to_idx = var_to_app_to_idx[ipt]
                if app not in app_to_idx:
                    app_to_idx[app] = []
                idx = app_to_idx[app]
                if i not in idx:
                    idx.append(i)
                account_for(ipt)

    # add all variables that are true ancestors of the cost
    for output in outputs:
        account_for(output)

    # determine which variables have elements of wrt as a true
    # ancestor. Do this with an upward pass starting from wrt,
    # following only true connections
    visited = set([])

    def visit(var):
        if var in visited:
            return
        if var not in var_to_app_to_idx:
            return
        visited.add(var)
        nodes = var_to_app_to_idx[var]
        for node in nodes:
            connection_pattern = _node_to_pattern(node)
            for idx in nodes[node]:
                for ii, output in enumerate(node.outputs):
                    if connection_pattern[idx][ii]:
                        visit(output)

    for elem in wrt:
        visit(elem)

    # Remove variables that don't have wrt as a true ancestor
    orig_vars = list(var_to_app_to_idx.keys())
    for var in orig_vars:
        if var not in visited:
            del var_to_app_to_idx[var]

    return var_to_app_to_idx


class NullTypeGradError(TypeError):
    """
    Raised when grad encounters a NullType.
    """


class DisconnectedInputError(ValueError):
    """
    Raised when grad is asked to compute the gradient
    with respect to a disconnected input and
    disconnected_inputs='raise'.
    """


def _populate_grad_dict(var_to_app_to_idx,
                        grad_dict, wrt, cost_name=None):
    """Helper function for grad function.

    Parameters
    ----------
    var_to_app_to_idx : dict
        a dictionary mapping a variable to a second dictionary.
        the second dictionary maps apply nodes acting on
        this variable to the variable's index in the apply
        node's input list
    grad_dict : dict
        A dictionary mapping variables to their gradients.
        Should be populated by grad function, which should:

        - Set the gradient with respect to the cost to 1
        - Load all gradients from known_grads, possibly
          overriding the cost
        - Set the gradient for disconnected
          inputs to a variable with type DisconnectedType()

    wrt : list of Variables
        the minimal set of variables that must be included in `grad_dict`
    cost_name: string
        The name of the cost being differentiated, optional.
        Used to name the grad with respect to x as (d<cost_name>/dx)

    Returns
    -------
    list of Variables
        A list of gradients corresponding to `wrt`

    """
    # build a dict mapping node to the terms node contributes to each of
    # its inputs' gradients
    term_dict = OrderedDict()

    def access_term_cache(node):
        """ Populates term_dict[node] and returns it """

        if node not in term_dict:

            inputs = node.inputs

            output_grads = [access_grad_cache(var) for var in node.outputs]

            # list of bools indicating if each output is connected to the cost
            outputs_connected = [not isinstance(g.type, DisconnectedType)
                                 for g in output_grads]

            connection_pattern = _node_to_pattern(node)

            # list of bools indicating if each input is connected to the cost
            inputs_connected = [
                (True in [input_to_output and output_to_cost for
                          input_to_output, output_to_cost in
                          zip(input_to_outputs, outputs_connected)]) for
                input_to_outputs in connection_pattern
            ]

            # List of bools indicating if each output is an integer dtype
            output_is_int = [hasattr(output.type, 'dtype') and
                             output.type.dtype in theano.tensor.discrete_dtypes
                             for output in node.outputs]

            # List of bools indicating if each output is NullType
            ograd_is_nan = [isinstance(output.type, NullType)
                            for output in output_grads]

            # List of bools indicating if each input only has NullType outputs
            only_connected_to_nan = [
                (True not in
                 [in_to_out and out_to_cost and not out_nan
                  for in_to_out, out_to_cost, out_nan in
                  zip(in_to_outs, outputs_connected, ograd_is_nan)])
                for in_to_outs in connection_pattern]

            if True not in inputs_connected:
                # All outputs of this op are disconnected so we can skip
                # Calling the op's grad method and report that the inputs
                # are disconnected
                # (The op's grad method could do this too, but this saves the
                # implementer the trouble of worrying about this case)
                input_grads = [disconnected_type() for ipt in inputs]
            elif False not in only_connected_to_nan:
                # All inputs are only connected to nan gradients, so we don't
                # need to bother calling the grad method. We know the gradient
                # with respect to all connected inputs is nan.
                input_grads = []
                for connected in inputs_connected:
                    if connected:
                        input_grads.append(null_type())
                    else:
                        input_grads.append(disconnected_type())
            else:
                # At least one input of this op is connected to the cost so and
                # not all output gradients are undefined so we must
                # call the op's grad method

                # Each Op's grad function requires inputs and output_grads
                # If the Op destroys any input, but the grad expression uses
                # it, then chances are the resulting graph will have a
                # dependency cycle. We avoid this cycle by passing (symbolic)
                # copies of each destroyed input.
                try:
                    dinputs = [node.inputs[x[0]] for x in
                               itervalues(node.op.destroy_map)]
                except AttributeError:
                    dinputs = []

                def try_to_copy_if_needed(var):
                    if var in dinputs and hasattr(var, 'copy'):
                        return var.copy()
                    return var

                inputs = [try_to_copy_if_needed(ipt) for ipt in inputs]

                # Build a list of output gradients with the same dtype as
                # the corresponding output variable.
                # If an output is of a float dtype, we want to cast the
                # output gradient into the same dtype, to avoid having a
                # gradient graph with double precision (taking more memory,
                # and more computation).
                # If an output is of an integer dtype, then we just leave it
                # alone.
                # DO NOT force integer variables to have zero grad. This causes
                # bugs where we fail to detect disconnected or undefined
                # gradients.
                # DO NOT force integer variables to have integer dtype.
                # This is a violation of the op contract.
                new_output_grads = []
                for o, og in zip(node.outputs, output_grads):
                    o_dt = getattr(o.type, 'dtype', None)
                    og_dt = getattr(og.type, 'dtype', None)
                    if (o_dt not in theano.tensor.discrete_dtypes and
                            og_dt and o_dt != og_dt):
                        new_output_grads.append(og.astype(o_dt))
                    else:
                        new_output_grads.append(og)

                # Make sure that, if new_output_grads[i] has a floating point
                # dtype, it is the same dtype as outputs[i]
                for o, ng in zip(node.outputs, new_output_grads):
                    o_dt = getattr(o.type, 'dtype', None)
                    ng_dt = getattr(ng.type, 'dtype', None)
                    if (ng_dt is not None and
                            o_dt not in theano.tensor.discrete_dtypes):
                        assert ng_dt == o_dt

                # Someone who had obviously not read the Op contract tried
                # to modify this part of the function.
                # If you ever think it is a good idea to make an integer
                # valued gradient, please
                # 1) Read the Op contract again
                # 2) Talk to Ian Goodfellow
                # (Both of these sources will tell you not to do it)
                for ng in new_output_grads:
                    assert (getattr(ng.type, 'dtype', None)
                            not in theano.tensor.discrete_dtypes)

                # If config.compute_test_value is turned on, check that the
                # gradients on the outputs of this node have the right shape.
                # We also check the gradient on the inputs later--both checks
                # are needed, because some gradients are only ever specified
                # by the user, not computed by Op.grad, and some gradients are
                # only computed and returned, but never passed as another
                # node's output grads.
                for idx, packed in enumerate(izip(node.outputs,
                                             new_output_grads)):
                    orig_output, new_output_grad = packed
                    if not hasattr(orig_output, 'shape'):
                        continue
                    if isinstance(new_output_grad.type, DisconnectedType):
                        continue
                    for orig_output_v, new_output_grad_v in get_debug_values(
                            *packed):
                        o_shape = orig_output_v.shape
                        g_shape = new_output_grad_v.shape
                        if o_shape != g_shape:
                            raise ValueError(
                                "Got a gradient of shape " +
                                str(o_shape) + " on an output of shape " +
                                str(g_shape))

                input_grads = node.op.L_op(inputs, node.outputs,
                                           new_output_grads)

                if input_grads is None:
                    raise TypeError("%s.grad returned NoneType, "
                                    "expected iterable." % str(node.op))

                if len(input_grads) != len(inputs):
                    raise ValueError(("%s returned the wrong number of" +
                                      " gradient terms.") % str(node.op))
# We can not enforce this, as AdvancedSubtensor1 has an option to
# return the sparse grad for optimization reason.

                    #            for ig, i in zip(input_grads, inputs):
#                if (not isinstance(ig.type, (DisconnectedType, NullType)) and
#                    type(ig.type) != type(i.type)):
#                    raise ValueError(
#                        "%s returned the wrong type for gradient terms."
#                        " Sparse inputs must have sparse grads and dense"
#                        " inputs must have dense grad. Got %s, expected %s" %(
#                            str(node.op), ig.type, i.type))

            # must convert to list in case the op returns a tuple
            # we won't be able to post-process out the Nones if it does that
            input_grads = list(input_grads)

            # Need to propagate the NullType gradients; if an input grad is
            # not disconnected and the corresponding input is connected
            # to at least one output whose gradient is NullType then the input
            # grad should be NullType.
            for inp_idx in range(len(input_grads)):
                for out_idx in range(len(ograd_is_nan)):
                    if (ograd_is_nan[out_idx] and
                            connection_pattern[inp_idx][out_idx] and
                            not isinstance(input_grads[inp_idx].type,
                                           DisconnectedType)):
                        input_grads[inp_idx] = output_grads[out_idx]

            # Do type checking on the result

            # List of bools indicating if each input only has integer outputs
            only_connected_to_int = [
                (True not in
                 [in_to_out and out_to_cost and not out_int
                  for in_to_out, out_to_cost, out_int in
                  zip(in_to_outs, outputs_connected, output_is_int)])
                for in_to_outs in connection_pattern]

            for i, term in enumerate(input_grads):

                # Disallow Nones
                if term is None:
                    # We don't know what None means. in the past it has been
                    # used to mean undefined, zero, or disconnected.
                    # We therefore don't allow it because its usage has become
                    # so muddied.
                    raise TypeError(
                        ('%s.grad returned None for' +
                         ' a gradient term, '
                         'this is prohibited. Instead of None,'
                         'return zeros_like(input), disconnected_type(),'
                         ' or a NullType variable such as those made with '
                         'the grad_undefined or grad_unimplemented helper '
                         'functions.') % node.op)

                # Check that the gradient term for this input
                # has the right shape
                if hasattr(term, 'shape'):
                    orig_ipt = inputs[i]
                    for orig_ipt_v, term_v in get_debug_values(orig_ipt, term):
                        i_shape = orig_ipt_v.shape
                        t_shape = term_v.shape
                        if i_shape != t_shape:
                            raise ValueError(
                                "%s.grad returned object of "
                                "shape %s as gradient term on input %d "
                                "of shape %s" % (node.op, t_shape, i, i_shape))

                if not isinstance(term.type,
                                  (NullType, DisconnectedType)):
                    if term.type.dtype not in theano.tensor.float_dtypes:
                        raise TypeError(str(node.op) + '.grad illegally '
                                        ' returned an integer-valued variable.'
                                        ' (Input index %d, dtype %s)' % (
                                            i, term.type.dtype))

                    if only_connected_to_nan[i]:
                        assert isinstance(term.type, NullType)

                    if only_connected_to_int[i]:
                        # This term has only integer outputs and we know
                        # it's not undefined or disconnected
                        # The only other valid thing it can be is 0

                        is_zero = _is_zero(term)
                        assert is_zero in ['yes', 'no', 'maybe']
                        if is_zero == 'maybe':
                            msg = ("%s.grad returned %s of type %s for input"
                                   " %d. This input's only connections to "
                                   "the cost through this op are via "
                                   "integer-valued outputs so it should be "
                                   "NullType, DisconnectedType, or some form "
                                   "of zeros. It is not NullType or "
                                   "DisconnectedType and theano can't "
                                   "simplify it to a constant, so it's not "
                                   "verifiably zeros.")

                            msg %= (node.op, term, type(term), i)

                        elif is_zero == 'no':
                            msg = ("%s.grad returned %s of type %s for input"
                                   " %d. Since this input is only connected "
                                   "to integer-valued outputs, it should "
                                   "evaluate to zeros, but it evaluates to"
                                   "%s.")

                            msg %= (node.op, term, type(term), i,
                                    theano.get_scalar_constant_value(term))

                            raise ValueError(msg)

            # Check that op.connection_pattern matches the connectivity
            # logic driving the op.grad method
            for i, (ipt, ig, connected) in enumerate(
                zip(inputs, input_grads, inputs_connected)
            ):
                actually_connected = \
                    not isinstance(ig.type, DisconnectedType)

                if actually_connected and not connected:
                    msg = ("%s.grad returned %s of type %s for input %d."
                           " Expected DisconnectedType instance based on "
                           " the output of the op's connection_pattern "
                           "method.")
                    msg %= (str(node.op), str(ig), str(ig.type), i)
                    raise TypeError(msg)

                elif connected and not actually_connected:
                    msg = "%s.grad returned DisconnectedType for input %d."
                    msg %= (str(node.op), i)
                    if hasattr(node.op, 'connection_pattern'):
                        msg += (' Its connection_pattern method does not'
                                ' allow this.')
                        raise TypeError(msg)
                    else:
                        msg += (' You may want to implement a '
                                'connection_pattern method for it.')
                        warnings.warn(msg)

            # cache the result
            term_dict[node] = input_grads

        return term_dict[node]

    # populate grad_dict[var] and return it
    def access_grad_cache(var):
        if var not in grad_dict:
            # If var is not in grad_dict already, we must compute it
            if var in var_to_app_to_idx:
                null_terms = []
                terms = []
                node_to_idx = var_to_app_to_idx[var]
                for node in node_to_idx:
                    for idx in node_to_idx[node]:

                        term = access_term_cache(node)[idx]

                        if not isinstance(term, gof.Variable):
                            raise TypeError(
                                "%s.grad returned %s, expected"
                                " Variable instance." % (str(node.op),
                                                         type(term)))

                        if isinstance(term.type, NullType):
                            null_terms.append(term)
                            continue

                        # Don't try to sum up DisconnectedType placeholders
                        if isinstance(term.type, DisconnectedType):
                            continue

                        if hasattr(var, 'ndim') and term.ndim != var.ndim:
                            raise ValueError(
                                ("%s.grad returned a term with"
                                 " %d dimensions, but %d are required.") % (
                                     str(node.op), term.ndim, var.ndim))

                        terms.append(term)

                # Add up the terms to get the total gradient on this variable
                if len(null_terms) > 0:
                    # At least one term is a NullType : the total gradient
                    # will also be a NullType
                    grad_dict[var] = null_terms[0]
                elif len(terms) > 0:
                    # the next line is like sum(terms) but doesn't add an
                    # extraneous TensorConstant(0)
                    grad_dict[var] = reduce(lambda x, y: x + y, terms)
                else:
                    grad_dict[var] = disconnected_type()

                if cost_name is not None and var.name is not None:
                    grad_dict[var].name = '(d%s/d%s)' % (cost_name, var.name)
            else:
                # this variable isn't connected to the cost in the
                # computational graph
                grad_dict[var] = disconnected_type()
        # end if cache miss
        return grad_dict[var]

    rval = [access_grad_cache(elem) for elem in wrt]

    return rval


def _float_zeros_like(x):
    """ Like zeros_like, but forces the object to have a
    a floating point dtype """

    rval = x.zeros_like()

    if rval.type.dtype.find('float') != -1:
        return rval

    return rval.astype(theano.config.floatX)


def _float_ones_like(x):
    """ Like ones_like, but forces the object to have a
    floating point dtype """

    dtype = x.type.dtype
    if dtype not in tensor.float_dtypes:
        dtype = theano.config.floatX

    return x.ones_like(dtype=dtype)


class numeric_grad(object):
    """
    Compute the numeric derivative of a scalar-valued function at a particular
    point.
    """

    # Note on step sizes and tolerances:
    #
    # There is a relationship between the step size and the function value and
    # the measurement error that is incurred due to rounding.  The finite
    # difference we measure is
    # delta = f(x0) - f(x0+eps)
    #
    # For maximum precision, f should be close to zero.
    # For every power of 2 that f departs from zero, we lose a bit of precision
    # in delta.
    #
    # Even in this case of maximum accuracy, there is a tradeoff between
    # stepsize and measurement error.
    # Taking small steps allows us to measure large derivatives accuractly,
    # but longer steps are required to measure small derivatives accurately.
    # However longer steps introduce bias into our measurement in general
    # for non-linear functions.
    #
    # It would be interesting to have a version of numeric grad that used an
    # adaptive stepsize.
    #
    # For now, we use a heuristic that catches very bad gradients, but is not
    # perfectly accurate.
    type_eps = {'float64': 1e-7,
                'float32': 3e-4,
                'float16': 1e-1,
                np.dtype('float64'): 1e-7,
                np.dtype('float32'): 3e-4,
                np.dtype('float16'): 1e-1}

    def __init__(self, f, pt, eps=None, out_type=None):
        """Return the gradient of f at pt.

        This function computes the gradient by a one-sided finite
        differences of a fixed step size (eps).

        Parameters
        ----------
        f : a differentiable function such that f(*pt) is a scalar
            The function to compute the gradient of.
            It is assumed that f(...) will return a scalar.
            It is assumed that all f's inputs are numpy.ndarray objects.
        pt : an ndarray, a list of ndarrays or tuple of ndarrays
            The point where to evaluate the gradient
        out_type: float
            dtype of output, if complex (i.e. 'complex32' or 'complex64')
        eps : float, optional
            The stepsize for the finite differencing.  None means
            input dtype-dependent. See `type_eps`.
        """

        def prod(inputs):
            rval = 1
            for i in inputs:
                rval *= i
            return rval

        packed_pt = False
        if not isinstance(pt, (list, tuple)):
            pt = [pt]
            packed_pt = True

        apt = [np.array(p) for p in pt]

        shapes = [p.shape for p in apt]
        dtypes = [str(p.dtype) for p in apt]

        # TODO: remove this eventually (why was this here in the first place ?)
        # In the case of CSM, the arguments are a mixture of floats and
        # integers...
        # if not dtypes == [dtypes[0]] * len(apt):
        #      raise TypeError('All function arguments must have same dtype')

        total_size = builtins.sum(prod(sh) for sh in shapes)

        working_dtype = builtins.min(
            (self.type_eps[dt], dt) for dt in dtypes)[1]

        # create un-initialized memory
        x = np.ndarray((total_size,), dtype=working_dtype)
        # (not out_type is None) --> (out_type is not None) ???
        if (out_type is not None) and (out_type.startswith('complex')):
            gx = np.ndarray((total_size,), dtype=out_type)
        else:
            gx = np.ndarray((total_size,), dtype=working_dtype)

        if eps is None:
            eps = builtins.max(self.type_eps[dt] for dt in dtypes)

        # set up aliases so that apt[i] is backed by memory in x
        # and self.gf is backed by memory in gx
        cur_pos = 0
        self.gf = []
        for i, p in enumerate(apt):
            p_size = prod(p.shape)
            # set up alias
            apt[i] = x[cur_pos: cur_pos + p_size].reshape(p.shape)
            self.gf.append(gx[cur_pos: cur_pos + p_size].reshape(p.shape))
            # initialize with p's value
            apt[i][...] = p
            cur_pos += p_size

        f_x = f(*[p.copy() for p in apt])

        # now iterate over the elements of x, and call f on apt.
        x_copy = x.copy()
        for i in xrange(total_size):
            x[:] = x_copy

            x[i] += eps
            f_eps = f(*apt)

            # TODO: remove this when it is clear that the next
            # replacemement does not pose problems of its own.  It was replaced
            # for its inability to handle complex variables.
            # gx[i] = numpy.asarray((f_eps - f_x) / eps)

            gx[i] = ((f_eps - f_x) / eps)

        if packed_pt:
            self.gf = self.gf[0]

    @staticmethod
    def abs_rel_err(a, b):
        """Return absolute and relative error between a and b.

        The relative error is a small number when a and b are close, relative
        to how big they are.

        Formulas used:
            abs_err = abs(a - b)

            rel_err = abs_err / max(abs(a) + abs(b), 1e-8)

        The denominator is clipped at 1e-8 to avoid dividing by 0 when a and b
        are both close to 0.

        The tuple (abs_err, rel_err) is returned
        """
        abs_err = abs(a - b)
        # 1e-8 is to prevent division by zeros.
        # [] is to make sure that if a and b are float16, 1e-8 don't get
        # dowcasted to float16 as that give 0! This would add back the
        # division by zero
        rel_err = abs_err / np.maximum(abs(a) + abs(b), [1e-8])
        # The numpy.asarray are needed as if a or b is a sparse matrix
        # this would result in a numpy.matrix and not a numpy.ndarray
        # and the behave differently causing problem later.
        # In particular a_npy_matrix.flatten().shape == (1, n_element)
        abs_err = np.asarray(abs_err)
        rel_err = np.asarray(rel_err)
        return (abs_err, rel_err)

    def abs_rel_errors(self, g_pt):
        """Return the abs and rel error of gradient estimate `g_pt`

        `g_pt` must be a list of ndarrays of the same length as self.gf,
        otherwise a ValueError is raised.

        Corresponding ndarrays in `g_pt` and `self.gf` must have the same
        shape or ValueError is raised.

        """
        if len(g_pt) != len(self.gf):
            raise ValueError('argument has wrong number of elements',
                             len(g_pt))
        errs = []
        for i, (a, b) in enumerate(zip(g_pt, self.gf)):
            if a.shape != b.shape:
                raise ValueError('argument element %i has wrong shape %s' % (
                    i, str((a.shape, b.shape))))
            errs.append(numeric_grad.abs_rel_err(a, b))
        return errs

    def max_err(self, g_pt, abs_tol, rel_tol):
        """Find the biggest error between g_pt and self.gf.

        What is measured is the violation of relative and absolute errors,
        wrt the provided tolerances (abs_tol, rel_tol).
        A value > 1 means both tolerances are exceeded.

        Return the argmax of min(abs_err / abs_tol, rel_err / rel_tol) over
        g_pt, as well as abs_err and rel_err at this point.
        """
        pos = []
        errs = []
        abs_errs = []
        rel_errs = []

        abs_rel_errs = self.abs_rel_errors(g_pt)
        for abs_err, rel_err in abs_rel_errs:
            if not np.all(np.isfinite(abs_err)):
                raise ValueError('abs_err not finite', repr(abs_err))
            if not np.all(np.isfinite(rel_err)):
                raise ValueError('rel_err not finite', repr(rel_err))
            scaled_err = np.minimum(abs_err / abs_tol, rel_err / rel_tol)
            max_i = scaled_err.argmax()

            pos.append(max_i)
            errs.append(scaled_err.flatten()[max_i])
            abs_errs.append(abs_err.flatten()[max_i])
            rel_errs.append(rel_err.flatten()[max_i])

        # max over the arrays in g_pt
        max_arg = np.argmax(errs)
        max_pos = pos[max_arg]
        return (max_arg, max_pos, abs_errs[max_arg], rel_errs[max_arg])


def mode_not_slow(mode):
    if mode == 'FAST_COMPILE':
        return FAST_RUN
    mode = get_mode(mode)
    if isinstance(mode, DebugMode):
        opt = mode.optimizer
        return FAST_RUN.clone(optimizer=opt)
    else:
        return mode


def verify_grad(fun, pt, n_tests=2, rng=None, eps=None,
                out_type=None, abs_tol=None,
                rel_tol=None, mode=None, cast_to_output_type=False,
                no_debug_ref=True):
    """Test a gradient by Finite Difference Method. Raise error on failure.

    Raises an Exception if the difference between the analytic gradient and
    numerical gradient (computed through the Finite Difference Method) of a
    random projection of the fun's output to a scalar exceeds the given
    tolerance.

    Examples
    --------
    >>> verify_grad(theano.tensor.tanh,
    ...             (np.asarray([[2, 3, 4], [-1, 3.3, 9.9]]),),
    ...             rng=np.random)

    Parameters
    ----------
    fun : a Python function
        `fun` takes Theano variables as inputs, and returns a Theano variable.
        For instance, an Op instance with  a single output.
    pt : list of numpy.ndarrays
        Input values, points where the gradient is estimated.
        These arrays must be either float16, float32, or float64 arrays.
    n_tests : int
        number of times to run the test
    rng : numpy.random.RandomState, optional
        random number generator used to sample the output random projection `u`,
        we test gradient of sum(u * fun) at `pt`
    eps : float, optional
        stepsize used in the Finite Difference Method (Default
        None is type-dependent).
        Raising the value of eps can raise or lower the absolute
        and relative errors of the verification depending on the
        Op. Raising eps does not lower the verification quality for
        linear operations. It is better to raise `eps` than raising
        `abs_tol` or `rel_tol`.
    out_type : string
        dtype of output, if complex (i.e., 'complex32' or 'complex64')
    abs_tol : float
        absolute tolerance used as threshold for gradient comparison
    rel_tol : float
        relative tolerance used as threshold for gradient comparison
    cast_to_output_type : bool
        if the output is float32 and cast_to_output_type is True, cast
        the random projection to float32. Otherwise it is float64.
        float16 is not handled here.
    no_debug_ref : bool
        Don't use DebugMode for the numerical gradient function.

    Note
    ----
    This function does not support multiple outputs. In
    tests/test_scan.py there is an experimental verify_grad that
    covers that case as well by using random projections.

    """
    # The import is here to prevent circular import.
    from theano import compile, shared
    import theano.tensor
    from theano.tensor import as_tensor_variable, TensorType
    assert isinstance(pt, (list, tuple))
    pt = [np.array(p) for p in pt]

    for i, p in enumerate(pt):
        if p.dtype not in ('float16', 'float32', 'float64'):
            raise TypeError(
                ('verify_grad can work only with floating point '
                 'inputs, but input %i has dtype "%s".') % (i, p.dtype))

    _type_tol = dict(  # relative error tolerances for different types
        float16=5e-2,
        float32=1e-2,
        float64=1e-4)

    if abs_tol is None:
        abs_tol = builtins.max(_type_tol[str(p.dtype)] for p in pt)
    if rel_tol is None:
        rel_tol = builtins.max(_type_tol[str(p.dtype)] for p in pt)

    if rng is None:
        raise TypeError(('rng should be a valid instance of '
                        'numpy.random.RandomState. You may '
                         'want to use theano.tests.unittest'
                         '_tools.verify_grad instead of '
                         'theano.gradient.verify_grad.'))

    # We allow input downcast in function, because numeric_grad works in the
    # most precise dtype used among the inputs, so we may need to cast some.
    def function(inputs, output, name, mode=mode):
        f = compile.function(inputs, output, accept_inplace=True,
                             allow_input_downcast=True, mode=mode,
                             on_unused_input='ignore', name=name)
        return f

    tensor_pt = [
        TensorType(
            as_tensor_variable(p).dtype,
            as_tensor_variable(p).broadcastable)(name='input %i' % i)
        for i, p in enumerate(pt)]

    # fun can be either a function or an actual Op instance
    o_output = fun(*tensor_pt)

    if isinstance(o_output, list):
        raise NotImplementedError(('cant (yet) autotest gradient of fun '
                                   'with multiple outputs'))
        # we could make loop over outputs making random projections R for each,
        # but this doesn't handle the case where not all the outputs are
        # differentiable... so I leave this as TODO for now -JB.

    o_fn = function(tensor_pt, o_output, name='gradient.py fwd')
    o_fn_out = o_fn(*[p.copy() for p in pt])

    if isinstance(o_fn_out, tuple) or isinstance(o_fn_out, list):
        raise TypeError(
            'It seems like you are trying to use verify_grad '
            'on an op or a function which outputs a list: there should'
            ' be a single (array-like) output instead')

    # random_projection should not have elements too small,
    # otherwise too much precision is lost in numerical gradient
    def random_projection():
        plain = rng.rand(*o_fn_out.shape) + 0.5
        if cast_to_output_type and o_output.dtype == "float32":
            return np.array(plain, o_output.dtype)
        return plain

    t_r = shared(random_projection(), borrow=True)
    t_r.name = 'random_projection'

    # random projection of o onto t_r
    # This sum() is defined above, it's not the builtin sum.
    cost = theano.tensor.sum(t_r * o_output)

    if no_debug_ref:
        mode_for_cost = mode_not_slow(mode)
    else:
        mode_for_cost = mode

    cost_fn = function(tensor_pt, cost, name='gradient.py cost',
                       mode=mode_for_cost)

    symbolic_grad = grad(cost, tensor_pt,
                         disconnected_inputs='ignore')

    grad_fn = function(tensor_pt, symbolic_grad,
                       name='gradient.py symbolic grad')

    for test_num in xrange(n_tests):
        try:
            num_grad = numeric_grad(cost_fn, [p.copy() for p in pt],
                                    eps, out_type)

            analytic_grad = grad_fn(*[p.copy() for p in pt])

            # Since `tensor_pt` is a list, `analytic_grad` should be one too.
            assert isinstance(analytic_grad, list)

            max_arg, max_err_pos, max_abs_err, max_rel_err = num_grad.max_err(
                analytic_grad, abs_tol, rel_tol)

            if max_abs_err > abs_tol and max_rel_err > rel_tol:

                raise verify_grad.E_grad(max_arg, max_err_pos,
                                         analytic_grad[max_arg].shape,
                                         analytic_grad[max_arg].flatten()[max_err_pos],
                                         num_grad.gf[max_arg].flatten()[max_err_pos],
                                         max_abs_err, max_rel_err,
                                         abs_tol, rel_tol)

            # get new random projection for next test
            if test_num < n_tests - 1:
                t_r.set_value(random_projection(), borrow=True)
        except Exception as e:
            e.args += ("\nThe error happened with the following inputs:", pt,
                       "\nThe value of eps is:", eps,
                       "\nThe out_type is:", out_type)
            raise


class GradientError(Exception):
    """This error is raised when a gradient is calculated, but incorrect."""
    def __init__(self, arg, err_pos, shape, val1, val2,
                 abs_err, rel_err, abs_tol, rel_tol):
        Exception.__init__(self)  # to be compatible with python2.4
        self.arg = arg
        self.err_pos = err_pos
        self.shape = shape
        self.val1 = val1
        self.val2 = val2
        self.abs_err = abs_err
        self.rel_err = rel_err
        self.abs_tol = abs_tol
        self.rel_tol = rel_tol

    def __str__(self):
        # args may have been inserted by e.g. makeTester
        args_msg = ", ".join(str(a) for a in self.args)
        return """\
GradientError: numeric gradient and analytic gradient exceed tolerance:
        At position %i of argument %i with shape %s,
            val1 = %f      ,  val2 = %f
            abs. error = %f,  abs. tolerance = %f
            rel. error = %f,  rel. tolerance = %f
Exception args: %s""" % (self.err_pos, self.arg,
                         self.shape,
                         self.val1, self.val2,
                         self.abs_err, self.abs_tol,
                         self.rel_err, self.rel_tol,
                         args_msg)


verify_grad.E_grad = GradientError


def jacobian(expression, wrt, consider_constant=None,
             disconnected_inputs='raise'):
    """
    Compute the full Jacobian, row by row.

    Parameters
    ----------
    expression : Vector (1-dimensional) :class:`~theano.gof.graph.Variable`
        Values that we are differentiating (that we want the Jacobian of)
    wrt : :class:`~theano.gof.graph.Variable` or list of Variables
        Term[s] with respect to which we compute the Jacobian
    consider_constant : list of variables
        Expressions not to backpropagate through

    disconnected_inputs: string
        Defines the behaviour if some of the variables
        in `wrt` are not part of the computational graph computing `cost`
        (or if all links are non-differentiable). The possible values are:

        - 'ignore': considers that the gradient on these parameters is zero.
        - 'warn': consider the gradient zero, and print a warning.
        - 'raise': raise an exception.

    Returns
    -------
    :class:`~theano.gof.graph.Variable` or list/tuple of Variables (depending upon `wrt`)
        The Jacobian of `expression` with respect to (elements of) `wrt`.
        If an element of `wrt` is not differentiable with respect to the
        output, then a zero variable is returned. The return value is
        of same type as `wrt`: a list/tuple or TensorVariable in all cases.
    """
    from theano.tensor import arange
    # Check inputs have the right format
    assert isinstance(expression, Variable), \
        "tensor.jacobian expects a Variable as `expression`"
    assert expression.ndim < 2, \
        ("tensor.jacobian expects a 1 dimensional variable as "
         "`expression`. If not use flatten to make it a vector")

    using_list = isinstance(wrt, list)
    using_tuple = isinstance(wrt, tuple)

    if isinstance(wrt, (list, tuple)):
        wrt = list(wrt)
    else:
        wrt = [wrt]

    if expression.ndim == 0:
        # expression is just a scalar, use grad
        return format_as(using_list, using_tuple,
                         grad(expression,
                              wrt,
                              consider_constant=consider_constant,
                              disconnected_inputs=disconnected_inputs))

    def inner_function(*args):
        idx = args[0]
        expr = args[1]
        rvals = []
        for inp in args[2:]:
            rval = grad(expr[idx],
                        inp,
                        consider_constant=consider_constant,
                        disconnected_inputs=disconnected_inputs)
            rvals.append(rval)
        return rvals
    # Computing the gradients does not affect the random seeds on any random
    # generator used n expression (because during computing gradients we are
    # just backtracking over old values. (rp Jan 2012 - if anyone has a
    # counter example please show me)
    jacobs, updates = theano.scan(inner_function,
                                  sequences=arange(expression.shape[0]),
                                  non_sequences=[expression] + wrt)
    assert not updates, \
        ("Scan has returned a list of updates. This should not "
         "happen! Report this to theano-users (also include the "
         "script that generated the error)")
    return format_as(using_list, using_tuple, jacobs)


def hessian(cost, wrt, consider_constant=None,
            disconnected_inputs='raise'):
    """
    Parameters
    ----------
    cost: Scalar (0-dimensional) variable.
    wrt: Vector (1-dimensional tensor) 'Variable' or list of
    vectors (1-dimensional tensors) Variables
    consider_constant:
        a list of expressions not to backpropagate through
    disconnected_inputs: string
        Defines the behaviour if some of the variables
        in ``wrt`` are not part of the computational graph computing ``cost``
        (or if all links are non-differentiable). The possible values are:

        - 'ignore': considers that the gradient on these parameters is zero.
        - 'warn': consider the gradient zero, and print a warning.
        - 'raise': raise an exception.

    Returns
    -------
    :class:`~theano.gof.graph.Variable` or list/tuple of Variables
        The Hessian of the `cost` with respect to (elements of) `wrt`.
        If an element of `wrt` is not differentiable with respect to the
        output, then a zero variable is returned. The return value is
        of same type as `wrt`: a list/tuple or TensorVariable in all cases.
    """
    from theano.tensor import arange
    # Check inputs have the right format
    assert isinstance(cost, Variable), \
        "tensor.hessian expects a Variable as `cost`"
    assert cost.ndim == 0, \
        "tensor.hessian expects a 0 dimensional variable as `cost`"

    using_list = isinstance(wrt, list)
    using_tuple = isinstance(wrt, tuple)

    if isinstance(wrt, (list, tuple)):
        wrt = list(wrt)
    else:
        wrt = [wrt]

    hessians = []
    for input in wrt:
        assert isinstance(input, Variable), \
            "tensor.hessian expects a (list of) Variable as `wrt`"
        assert input.ndim == 1, \
            "tensor.hessian expects a (list of) 1 dimensional variable "\
            "as `wrt`"
        expr = grad(cost, input, consider_constant=consider_constant,
                    disconnected_inputs=disconnected_inputs)

        # It is possible that the inputs are disconnected from expr,
        # even if they are connected to cost.
        # This should not be an error.
        hess, updates = theano.scan(lambda i, y, x: grad(
            y[i],
            x,
            consider_constant=consider_constant,
            disconnected_inputs='ignore'),
            sequences=arange(expr.shape[0]),
            non_sequences=[expr, input])
        assert not updates, \
            ("Scan has returned a list of updates. This should not "
             "happen! Report this to theano-users (also include the "
             "script that generated the error)")
        hessians.append(hess)
    return format_as(using_list, using_tuple, hessians)


def _is_zero(x):
    """
    Returns 'yes', 'no', or 'maybe' indicating whether x
    is always 0.
    'maybe' means that x is an expression that is complicated enough
    that we can't tell that it simplifies to 0.
    """
    if not hasattr(x, 'type'):
        return np.all(x == 0.)
    if isinstance(x.type, NullType):
        return 'no'
    if isinstance(x.type, DisconnectedType):
        return 'yes'

    no_constant_value = True
    try:
        constant_value = theano.get_scalar_constant_value(x)
        no_constant_value = False
    except theano.tensor.basic.NotScalarConstantError:
        pass

    if no_constant_value:
        return 'maybe'

    if constant_value != 0.:
        return 'no'

    return 'yes'


class ConsiderConstant(ViewOp):
    def grad(self, args, g_outs):
        return [g_out.zeros_like(g_out) for g_out in g_outs]


consider_constant_ = ConsiderConstant()


# I create a function only to have the doc show well.
def consider_constant(x):
    """
    DEPRECATED: use zero_grad() or disconnected_grad() instead.

    Consider an expression constant when computing gradients.

    The expression itself is unaffected, but when its gradient is
    computed, or the gradient of another expression that this
    expression is a subexpression of, it will not be backpropagated
    through. In other words, the gradient of the expression is
    truncated to 0.

    :param x: A Theano expression whose gradient should be truncated.

    :return: The expression is returned unmodified, but its gradient
        is now truncated to 0.

    .. versionadded:: 0.7
    """
    warnings.warn((
        "consider_constant() is deprecated, use zero_grad() or "
        "disconnected_grad() instead."), stacklevel=3)

    return consider_constant_(x)


class ZeroGrad(ViewOp):
    def grad(self, args, g_outs):
        return [g_out.zeros_like(g_out) for g_out in g_outs]

    def R_op(self, inputs, eval_points):
        if eval_points[0] is None:
            return [None]

        return theano.tensor.zeros(1)


zero_grad_ = ZeroGrad()


def zero_grad(x):
    """
    Consider an expression constant when computing gradients.

    The expression itself is unaffected, but when its gradient is
    computed, or the gradient of another expression that this
    expression is a subexpression of, it will be backpropagated
    through with a value of zero. In other words, the gradient of
    the expression is truncated to 0.

    Parameters
    ----------
    x: :class:`~theano.gof.graph.Variable`
        A Theano expression whose gradient should be truncated.

    Returns
    -------
    :class:`~theano.gof.graph.Variable`
        An expression equivalent to ``x``, with its gradient
        truncated to 0.
    """
    return zero_grad_(x)


class UndefinedGrad(ViewOp):
    def grad(self, args, g_outs):
        return [grad_undefined(self, i, arg) for i, arg in enumerate(args)]

    def R_op(self, inputs, eval_points):
        return [None]

    def connection_pattern(self, node):
        return [[True]]


undefined_grad_ = UndefinedGrad()


def undefined_grad(x):
    """
    Consider the gradient of this variable undefined.

    This will generate an error message if its gradient is taken.

    The expression itself is unaffected, but when its gradient is
    computed, or the gradient of another expression that this
    expression is a subexpression of, an error message will be generated
    specifying such gradient is not defined.

    Parameters
    ----------
    x: :class:`~theano.gof.graph.Variable`
        A Theano expression whose gradient should be undefined.

    Returns
    -------
    :class:`~theano.gof.graph.Variable`
        An expression equivalent to ``x``, with its gradient undefined.
    """
    return undefined_grad_(x)


class DisconnectedGrad(ViewOp):
    def grad(self, args, g_outs):
        return [disconnected_type() for g_out in g_outs]

    def R_op(self, inputs, eval_points):
        return [None]

    def connection_pattern(self, node):
        return [[False]]


disconnected_grad_ = DisconnectedGrad()


def disconnected_grad(x):
    """
    Consider an expression constant when computing gradients.

    It will effectively not backpropagating through it.

    The expression itself is unaffected, but when its gradient is
    computed, or the gradient of another expression that this
    expression is a subexpression of, it will not be backpropagated
    through. This is effectively equivalent to truncating the gradient
    expression to 0, but is executed faster than zero_grad(), which stilll
    has to go through the underlying computational graph related to the
    expression.

    Parameters
    ----------
    x: :class:`~theano.gof.graph.Variable`
        A Theano expression whose gradient should not be
        backpropagated through.

    Returns
    -------
    :class:`~theano.gof.graph.Variable`
        An expression equivalent to ``x``, with its gradient
        now effectively truncated to 0.
    """
    return disconnected_grad_(x)


class GradClip(ViewOp):
    # See doc in user fct grad_clip
    __props__ = ()

    def __init__(self, clip_lower_bound, clip_upper_bound):
        # We do not put those member in __eq__ or __hash__
        # as they do not influence the perform of this op.
        self.clip_lower_bound = clip_lower_bound
        self.clip_upper_bound = clip_upper_bound
        assert(self.clip_upper_bound >= self.clip_lower_bound)

    def grad(self, args, g_outs):
        return [theano.tensor.clip(g_out, self.clip_lower_bound,
                                   self.clip_upper_bound)
                for g_out in g_outs]


def grad_clip(x, lower_bound, upper_bound):
    """
    This op do a view in the forward, but clip the gradient.

    This is an elemwise operation.

    Parameters
    ----------
    x:
        The variable we want its gradient inputs clipped
    lower_bound:
        The lower bound of the gradient value
    upper_bound:
        The upper bound of the gradient value.

    Examples
    --------
    >>> x = theano.tensor.scalar()
    >>> z = theano.tensor.grad(grad_clip(x, -1, 1)**2, x)
    >>> z2 = theano.tensor.grad(x**2, x)
    >>> f = theano.function([x], outputs = [z, z2])
    >>> print(f(2.0))
    [array(1.0), array(4.0)]

    Note
    ----
    We register an opt in tensor/opt.py that remove the GradClip.
    So it have 0 cost in the forward and only do work in the grad.

    """
    return GradClip(lower_bound, upper_bound)(x)


class GradScale(ViewOp):
    def __init__(self, multiplier):
        self.multiplier = multiplier

    def grad(self, args, g_outs):
        return [self.multiplier * g_out for g_out in g_outs]


def grad_scale(x, multiplier):
    """
    This op scale or inverse the gradient in the backpropagation.

    Parameters
    ----------
    x:
        The variable we want its gradient inputs scale
    multiplier:
        Scale of the gradient

    Examples
    --------
    >>> x = theano.tensor.fscalar()
    >>> fx = theano.tensor.sin(x)
    >>> fp = theano.tensor.grad(fx, wrt=x)
    >>> fprime = theano.function([x], fp)
    >>> print(fprime(2))  # doctest: +ELLIPSIS
    -0.416...
    >>> f_inverse=grad_scale(fx, -1.)
    >>> fpp = theano.tensor.grad(f_inverse, wrt=x)
    >>> fpprime = theano.function([x], fpp)
    >>> print(fpprime(2))  # doctest: +ELLIPSIS
    0.416...
    """
    return GradScale(multiplier)(x)