Skip to content

NestedTensor

danling.tensor.nested_tensor

NestedTensor

A container for variable-length tensors that enables efficient batch operations.

NestedTensor solves a fundamental problem in deep learning: handling sequences of different lengths in batch operations. Instead of excessive padding or complex bucketing, NestedTensor provides an elegant solution that maintains both efficiency and usability.

The class provides three main views of the data: - .tensor: A padded tensor with zeros (or other value) in place of missing elements - .mask: A boolean mask indicating which elements are real vs padding - .concat: A flattened tensor containing all elements concatenated (no padding)

When indexing a NestedTensor, the behavior depends on the index type: 1. Integer index (nt[0]): Returns a single tensor without padding 2. Slice index (nt[:]): Returns a tuple of (padded_tensor, mask) 3. Tuple index (nt[:, 1:]): Returns a new NestedTensor with the specified sliced shape

Attributes:

Name Type Description
_storage

The sequence of original tensors (internal use)

tensor Tensor

Padded tensor representing all sequences with padding

mask Tensor

Boolean mask where True indicates real elements, False indicates padding

concat Tensor

Concatenated tensor of all sequences without padding

batch_first bool

Whether the first dimension is the batch dimension (B, N, *) If False, the first dimension is the sequence dimension (N, B, *)

padding_value SupportsFloat

Value used for padding in the padded tensor

mask_value bool

Value used in the mask to indicate padding positions (usually False)

Parameters:

Name Type Description Default

*tensors

Iterable[Tensor]

Variable-length tensors or sequences to store

()

batch_first

bool

Whether to use batch-first representation.

True

padding_value

SupportsFloat

Value to use for padding.

0.0

mask_value

bool

Value to use for padding positions in mask.

False

Raises:

Type Description
ValueError

If tensors is not an iterable or is empty

Examples:

Basic usage:

Python Console Session
>>> nested_tensor = NestedTensor(torch.tensor([1, 2, 3]), torch.tensor([4, 5]))
>>> nested_tensor.shape
torch.Size([2, 3])
>>> nested_tensor.tensor  # Padded representation
tensor([[1, 2, 3],
        [4, 5, 0]])
>>> nested_tensor.mask  # Mask showing real vs padding values
tensor([[ True,  True,  True],
        [ True,  True, False]])
>>> nested_tensor.concat  # Concatenated version (no padding)
tensor([1, 2, 3, 4, 5])
Python Console Session
1
2
3
4
5
6
7
8
>>> nested_tensor[0]  # First tensor (no padding)
tensor([1, 2, 3])
>>> nested_tensor[:2]  # Padded tensor and mask
NestedTensor([[1, 2, 3],
        [4, 5, 0]])
>>> nested_tensor[:, 1:]  # Slice operations return a new NestedTensor
NestedTensor([[2, 3],
        [5, 0]])

Type conversion:

Python Console Session
1
2
3
4
5
6
>>> nested_tensor.to(torch.float).tensor
tensor([[1., 2., 3.],
        [4., 5., 0.]])
>>> nested_tensor.half().tensor
tensor([[1., 2., 3.],
        [4., 5., 0.]], dtype=torch.float16)

Conversion to Python types:

Python Console Session
>>> nested_tensor.tolist()
[[1, 2, 3], [4, 5]]

Creating from Python lists:

Python Console Session
1
2
3
>>> NestedTensor(*[[1, 2, 3], [4, 5]])
NestedTensor([[1, 2, 3],
        [4, 5, 0]])
Source code in danling/tensor/nested_tensor.py
Python
  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
class NestedTensor:
    r"""
    A container for variable-length tensors that enables efficient batch operations.

    `NestedTensor` solves a fundamental problem in deep learning: handling sequences of different lengths
    in batch operations. Instead of excessive padding or complex bucketing, `NestedTensor` provides an
    elegant solution that maintains both efficiency and usability.

    The class provides three main views of the data:
    - `.tensor`: A padded tensor with zeros (or other value) in place of missing elements
    - `.mask`: A boolean mask indicating which elements are real vs padding
    - `.concat`: A flattened tensor containing all elements concatenated (no padding)

    When indexing a `NestedTensor`, the behavior depends on the index type:
    1. Integer index (`nt[0]`): Returns a single tensor without padding
    2. Slice index (`nt[:]`): Returns a tuple of (padded_tensor, mask)
    3. Tuple index (`nt[:, 1:]`): Returns a new `NestedTensor` with the specified sliced shape

    Attributes:
        _storage: The sequence of original tensors (internal use)
        tensor: Padded tensor representing all sequences with padding
        mask: Boolean mask where True indicates real elements, False indicates padding
        concat: Concatenated tensor of all sequences without padding
        batch_first: Whether the first dimension is the batch dimension (B, N, *)
            If `False`, the first dimension is the sequence dimension (N, B, *)
        padding_value: Value used for padding in the padded tensor
        mask_value: Value used in the mask to indicate padding positions (usually False)

    Args:
        *tensors: Variable-length tensors or sequences to store
        batch_first: Whether to use batch-first representation.
        padding_value: Value to use for padding.
        mask_value: Value to use for padding positions in mask.

    Raises:
        ValueError: If `tensors` is not an iterable or is empty

    Examples:
        Basic usage:
        >>> nested_tensor = NestedTensor(torch.tensor([1, 2, 3]), torch.tensor([4, 5]))
        >>> nested_tensor.shape
        torch.Size([2, 3])
        >>> nested_tensor.tensor  # Padded representation
        tensor([[1, 2, 3],
                [4, 5, 0]])
        >>> nested_tensor.mask  # Mask showing real vs padding values
        tensor([[ True,  True,  True],
                [ True,  True, False]])
        >>> nested_tensor.concat  # Concatenated version (no padding)
        tensor([1, 2, 3, 4, 5])

        Indexing:
        >>> nested_tensor[0]  # First tensor (no padding)
        tensor([1, 2, 3])
        >>> nested_tensor[:2]  # Padded tensor and mask
        NestedTensor([[1, 2, 3],
                [4, 5, 0]])
        >>> nested_tensor[:, 1:]  # Slice operations return a new NestedTensor
        NestedTensor([[2, 3],
                [5, 0]])

        Type conversion:
        >>> nested_tensor.to(torch.float).tensor
        tensor([[1., 2., 3.],
                [4., 5., 0.]])
        >>> nested_tensor.half().tensor
        tensor([[1., 2., 3.],
                [4., 5., 0.]], dtype=torch.float16)

        Conversion to Python types:
        >>> nested_tensor.tolist()
        [[1, 2, 3], [4, 5]]

        Creating from Python lists:
        >>> NestedTensor(*[[1, 2, 3], [4, 5]])
        NestedTensor([[1, 2, 3],
                [4, 5, 0]])
    """

    __storage: Tuple[Tensor, ...]

    dtype: torch.dtype | None = None
    device: torch.device | None = None
    requires_grad: bool | None = None
    _pin_memory: bool = False

    batch_first: bool = True
    padding_value: SupportsFloat = 0.0
    mask_value: bool = False

    def __init__(
        self,
        *tensors: Iterable[Tensor],
        dtype: torch.dtype | None = None,
        device: torch.device | None = None,
        requires_grad: bool | None = None,
        pin_memory: bool = False,
        batch_first: bool = True,
        padding_value: SupportsFloat = 0.0,
        mask_value: bool = False,
    ) -> None:
        self.dtype = dtype
        self.device = device
        self.requires_grad = requires_grad
        self._pin_memory = pin_memory
        if len(tensors) == 1 and isinstance(tensors, Sequence):
            tensors = tensors[0]  # type: ignore
        self._storage = tensors
        self.batch_first = batch_first
        self.padding_value = padding_value
        self.mask_value = mask_value

    @property
    def _storage(self):
        return self.__storage

    @_storage.setter
    def _storage(self, tensors: Sequence):
        if not isinstance(tensors, Iterable):
            raise ValueError(f"tensors must be an Iterable, bug got {type(tensors)}.")
        if isinstance(tensors, Tensor) and hasattr(tensors, "unbind"):
            tensors = tensors.unbind()
        tensors_ = []
        for t in tensors:
            if not isinstance(t, Tensor):
                t = torch.tensor(
                    t,
                    dtype=self.dtype,
                    device=self.device,
                    pin_memory=self._pin_memory,
                )
            else:
                t = t.to(self.device, dtype=self.dtype)
            if self.requires_grad is not None:
                t.requires_grad_(self.requires_grad)
            tensors_.append(t)
        if len(tensors_) == 0:
            self.__storage = ()
            return
        tensors = tuple(tensors_)
        self.dtype = tensors[0].dtype
        self.device = tensors[0].device
        self.requires_grad = tensors[0].requires_grad
        # if drop_last=False, the last element is likely not a NestedTensor and has an extra batch dimension
        ndims = {t.ndim for t in tensors[:-1]}
        if len(ndims) == 1:
            (ndim,) = ndims
            if tensors[-1].ndim == ndim + 1 and tensors[-1].size(0) == 1:
                tensors = tensors[:-1] + (tensors[-1].squeeze(0),)
        self.__storage = tensors

    def storage(self):
        return self._storage

    @property
    def tensor_mask(self) -> Tuple[Tensor, Tensor]:
        r"""
        Return a tuple of padded tensor and mask tensor.

        Examples:
            >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
            >>> nested_tensor.tensor_mask
            (tensor([[1, 2, 3],
                    [4, 5, 0]]), tensor([[ True,  True,  True],
                    [ True,  True, False]]))
        """

        return self._tensor_mask(
            self._storage, self.batch_first, self.padding_value, self.mask_value, self.requires_grad
        )

    @method_cache(maxsize=1)
    def _tensor_mask(
        self, storage: tuple, batch_first: bool, padding_value: SupportsFloat, mask_value: bool, requires_grad: bool
    ) -> Tensor:
        if storage[0].dim() == 0:
            return torch.stack(storage, dim=0), torch.full(
                (len(storage),), not mask_value, dtype=torch.bool, device=self.device
            )
        return tensor_mask(
            storage,
            size=self.size(),
            batch_first=batch_first,
            padding_value=float(padding_value),
            mask_value=mask_value,
        )

    @property
    def tensor(self) -> Tensor:
        r"""
        Return a single tensor by padding all the tensors.

        Examples:
            >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
            >>> nested_tensor.tensor
            tensor([[1, 2, 3],
                    [4, 5, 0]])
        """

        return self._tensor(self._storage, self.batch_first, self.padding_value, self.requires_grad)

    @method_cache(maxsize=1)
    def _tensor(self, storage: tuple, batch_first: bool, padding_value: SupportsFloat, requires_grad: bool) -> Tensor:
        if storage[0].dim() == 0:
            return torch.stack(storage, dim=0)
        return pad_tensor(storage, size=self.size(), batch_first=batch_first, padding_value=float(padding_value))

    @property
    def mask(self) -> Tensor:
        r"""
        Padding mask of `tensor`.

        Examples:
            >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
            >>> nested_tensor.mask
            tensor([[ True,  True,  True],
                    [ True,  True, False]])
        """

        return self._mask(self._storage, self.batch_first, self.mask_value, self.requires_grad)

    @method_cache(maxsize=1)
    def _mask(self, storage: tuple, batch_first: bool, mask_value: bool, requires_grad: bool) -> Tensor:
        if storage[0].dim() == 0:
            return torch.full((len(storage),), not mask_value, dtype=torch.bool, device=self.device)
        return mask_tensor(storage, size=self.size(), batch_first=batch_first, mask_value=mask_value)

    @property
    def concat(self) -> Tensor:
        r"""
        Concat `tensor` in padding dim.

        This is particularly useful when calculating loss or passing `Linear` to avoid unnecessary computation.

        Examples:
            >>> nested_tensor = NestedTensor([torch.randn(9, 8), torch.randn(11, 8)])
            >>> nested_tensor.concat.shape
            torch.Size([20, 8])
            >>> nested_tensor = NestedTensor([torch.randn(9, 9, 8), torch.randn(11, 11, 8)])
            >>> nested_tensor.concat.shape
            torch.Size([202, 8])
            >>> nested_tensor = NestedTensor([torch.randn(9, 9, 8, 6), torch.randn(11, 11, 8, 6)])
            >>> nested_tensor.concat.shape
            torch.Size([202, 8, 6])
            >>> nested_tensor = NestedTensor([torch.randn(9, 9, 8, 7), torch.randn(11, 11, 8, 6)])
            >>> nested_tensor.concat.shape
            torch.Size([1293, 8])
            >>> nested_tensor = NestedTensor([torch.randn(1, 9, 9, 5), torch.randn(1, 11, 11, 5)])
        """
        shape = list(self.size())  # type: ignore[arg-type]
        shape = shape[1:] if self.batch_first else shape[0] + shape[2:]
        elem = self._storage[0]
        if elem.shape == shape:
            return torch.cat(self._storage, dim=1 if self.batch_first else 0)
        static_dims = set(range(len(shape)))
        for i, s in enumerate(shape):
            if not all(t.size(i) == s for t in self._storage):
                shape[i] = -1
                static_dims.remove(i)
        target_shape = [-1] + [s for s in shape if s != -1]
        storage = [i.reshape(target_shape) for i in self._storage]
        return torch.cat(storage, dim=0 if self.batch_first else 1)

    @property
    def torch(self) -> Tensor:
        r"""
        Create a `torch.nested.nested_tensor` object from `self`.

        Examples:
            >>> nested_tensor = NestedTensor([[2, 3, 5], [7, 8]])
            >>> nested_tensor.torch
            nested_tensor([
              tensor([2, 3, 5]),
              tensor([7, 8])
            ])
        """
        if nested is None:
            raise ImportError("torch.nested is not available, please install torch with nested support.")
        return nested.nested_tensor(list(self._storage))

    def unbind(self, dim: int = 0) -> Tuple[Tensor, ...]:
        r"""
        Unbind the NestedTensor.
        """
        if dim != 0:
            raise ValueError(f"NestedTensor can only be unbound along dimension 0, got dimension {dim} instead.")
        return self._storage

    @property
    def occupancy(self) -> float:
        r"""
        Occupancy of the NestedTensor.

        Examples:
            >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3, 4]), torch.tensor([5, 6])])
            >>> nested_tensor.occupancy
            0.75
        """

        return self.numel() / self.shape.numel()  # type: ignore[union-attr]

    @classmethod
    def from_tensor_mask(cls, tensor: Tensor, mask: Tensor, **kwargs):
        r"""
        Build a `NestedTensor` object from a padded `Tensor` and corresponding mask `Tensor`.

        Args:
            tensor: Padded Tensor.
            mask: Tensor Mask.

        Examples:
            >>> padded_tensor = torch.tensor([[1, 2, 3, 0, 0],
            ...                                [4, 5, 0, 0, 0],
            ...                                [6, 7, 8, 9, 0]])
            >>> mask_tensor = torch.tensor([[1, 1, 1, 0, 0],
            ...                             [1, 1, 0, 0, 0],
            ...                             [1, 1, 1, 1, 0]])
            >>> nested_tensor = NestedTensor.from_tensor_mask(padded_tensor, mask_tensor)
            >>> nested_tensor
            NestedTensor([[1, 2, 3, 0],
                    [4, 5, 0, 0],
                    [6, 7, 8, 9]])
        """

        if mask.ndim == 1:
            return cls(tensor, **kwargs)
        if mask.ndim == 2:
            return cls((t[slice(0, m.sum())] for t, m in zip(tensor, mask)), **kwargs)
        return cls(
            (
                t[[slice(0, (m.sum(dim=dim) > 0).sum().item()) for dim in reversed(range(m.dim()))]]
                for t, m in zip(tensor, mask)
            ),
            **kwargs,
        )

    def nested_like(self, tensor: Tensor, strict: bool = True) -> Self:
        r"""
        Create a new `NestedTensor` from a `Tensor`.
        The newly created `NestedTensor` will have the same shape as current `NestedTensor`.

        Args:
            tensor: The tensor to be converted to `NestedTensor`.
            strict: Check if the shape of `tensor` is the same as the current `NestedTensor`.

        Examples:
            >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
            >>> (nested_tensor == nested_tensor.nested_like(nested_tensor)).all()
            tensor(True)
            >>> tensor = nested_tensor.tensor
            >>> (nested_tensor == nested_tensor.nested_like(tensor)).all()
            tensor(True)
            >>> f = nested_tensor.nested_like(torch.randn(2, 2))
            Traceback (most recent call last):
            ValueError: The shape of NestedTensor and input tensor does not match, torch.Size([2, 3]) != torch.Size([2, 2])
            >>> p = nested_tensor.nested_like(torch.randn(2, 2), False)
            >>> p = nested_tensor.nested_like(torch.randn(3, 3), False)
            Traceback (most recent call last):
            ValueError: The batch size of NestedTensor and input tensor does not match, 2 != 3
        """  # noqa: E501

        if isinstance(tensor, NestedTensor):
            return tensor.clone()

        if strict and self.shape != tensor.shape:
            raise ValueError(
                f"The shape of NestedTensor and input tensor does not match, {self.shape} != {tensor.shape}"
            )
        if self.size(0) != tensor.size(0):
            raise ValueError(
                f"The batch size of NestedTensor and input tensor does not match, {self.size(0)} != {tensor.size(0)}"
            )
        return self.__class__([o[tuple(slice(0, dim) for dim in t.shape)] for t, o in zip(self._storage, tensor)])

    @classmethod
    def __torch_function__(cls, func, types, args=(), kwargs=None) -> Self:
        if kwargs is None:
            kwargs = {}
        if func in NestedTensorFuncRegistry:
            return NestedTensorFuncRegistry[func](*args, **kwargs)
        args = [a.tensor if hasattr(a, "tensor") else a for a in args]
        for k, v in kwargs.items():
            if hasattr(v, "tensor"):
                kwargs[k] = v.tensor
        output = func(*args, **kwargs)
        if isinstance(output, (Tensor, NestedTensor)):
            return output
        return cls(output)

    @classmethod
    def __torch_dispatch__(cls, func, types, args=(), kwargs=None) -> Self:
        args = [a.tensor if hasattr(a, "tensor") else a for a in args]
        for k, v in kwargs.items():
            if hasattr(v, "tensor"):
                kwargs[k] = v.tensor
        output = func(*args, **kwargs)
        if isinstance(output, (Tensor, NestedTensor)):
            return output
        return cls(output)

    def __getitem__(self, index: int | slice | list | tuple | Tensor | NestedTensor) -> Tensor | NestedTensor:
        if isinstance(index, int):
            return self._storage[index]
        if isinstance(index, (slice, list)):
            storage = tuple(self._storage[index] if isinstance(index, slice) else [self._storage[i] for i in index])
            return NestedTensor(storage, **self._state)
        if isinstance(index, tuple):
            return NestedTensor([t[index[0]][index[1:]] for t in self._storage])
        if isinstance(index, Tensor):
            index = self.nested_like(index, strict=False)
        if isinstance(index, NestedTensor):
            return NestedTensor([t[i] for t, i in zip(self._storage, index._storage)])
        raise ValueError(f"Unsupported index type {type(index)}")

    def __setitem__(self, index: int | slice | list | tuple, value: Tensor | NestedTensor) -> None:
        """
        Set values in the NestedTensor at the specified index.

        Args:
            index: The index to modify. Can be an integer, slice, list, or tuple.
            value: The new value to set. Can be a Tensor or NestedTensor.

        Examples:
            >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
            >>> nested_tensor[0] = torch.tensor([6, 7, 8])
            >>> nested_tensor[0]
            tensor([6, 7, 8])
            >>> nested_tensor[1] = torch.tensor([9, 10, 11, 12])
            >>> nested_tensor.shape
            torch.Size([2, 4])
        """
        if isinstance(index, int):
            if isinstance(value, NestedTensor):
                if len(value._storage) != 1:
                    raise ValueError(
                        f"When setting with an integer index, value must have a single tensor, but got {len(value)}"
                    )
                value = value._storage[0]
            if not isinstance(value, Tensor):
                value = torch.tensor(value, device=self.device)
            # Create a new list of tensors to modify
            storage_list = list(self._storage)
            storage_list[index] = value
            self._storage = tuple(storage_list)
        elif isinstance(index, (slice, list)):
            if isinstance(value, Tensor):
                # Convert tensor to NestedTensor if it's a regular tensor
                if value.dim() > 1 and value.size(0) > 1:
                    value = NestedTensor(value.unbind(0))
                else:
                    value = NestedTensor([value])

            if isinstance(index, slice):
                start, stop, step = index.indices(len(self._storage))
                indices = range(start, stop, step)
            else:
                indices = index  # type: ignore[assignment]

            if len(indices) != len(value._storage):
                raise ValueError(
                    f"Size mismatch: tried to assign {len(value._storage)} values to {len(indices)} indices"
                )

            storage_list = list(self._storage)
            for i, idx in enumerate(indices):
                storage_list[idx] = value._storage[i]
            self._storage = tuple(storage_list)
        elif isinstance(index, tuple):
            if len(index) < 2:
                raise ValueError("Tuple index must have at least two elements")

            first_idx, rest_idx = index[0], index[1:]

            if isinstance(first_idx, int):
                # Handle single tensor modification
                storage_list = list(self._storage)
                tensor = storage_list[first_idx]
                tensor[rest_idx] = value
                storage_list[first_idx] = tensor
                self._storage = tuple(storage_list)
            elif isinstance(first_idx, (slice, list)):
                # Handle multiple tensor modification
                if isinstance(first_idx, slice):
                    start, stop, step = first_idx.indices(len(self._storage))
                    indices = range(start, stop, step)
                else:
                    indices = first_idx  # type: ignore[assignment]

                storage_list = list(self._storage)
                for idx in indices:
                    tensor = storage_list[idx]
                    tensor[rest_idx] = value
                    storage_list[idx] = tensor
                self._storage = tuple(storage_list)
            else:
                raise ValueError(f"Unsupported first index type {type(first_idx)}")
        else:
            raise ValueError(f"Unsupported index type {type(index)}")

    def __getattr__(self, name: str) -> Any:
        if not self._storage:
            raise ValueError(f"Unable to get {name} from an empty {self.__class__.__name__}")
        ret = [getattr(i, name) for i in self._storage]
        elem = ret[0]
        if isinstance(elem, Tensor):
            return NestedTensor(ret, **self._state)
        if callable(elem):
            return NestedTensorFuncWrapper(ret, state=self._state)
        if elem.__hash__ is not None and len(set(ret)) == 1:
            return elem
        return ret

    def __iter__(self):
        return iter(self._storage)

    @property
    def _state(self) -> Mapping:
        return self.__state__(return_dtype=False, return_device=True, return_requires_grad=False)

    def __state__(
        self, return_dtype: bool = True, return_device: bool = True, return_requires_grad: bool = False
    ) -> Mapping:
        state = {k: v for k, v in self.__dict__.items() if not (k.startswith("_") or k.endswith("_"))}
        if not return_dtype:
            state.pop("dtype", None)
        if not return_device:
            state.pop("device", None)
        if not return_requires_grad:
            state.pop("requires_grad", None)
        return state

    def __setstate__(self, state: Mapping) -> None:
        self.__dict__.update(state)

    def __len__(self) -> int:
        return len(self._storage)

    def __repr__(self):
        if not self._storage:
            return self.__class__.__name__ + "()"
        return self.__class__.__name__ + repr(self.tensor)[len(self.tensor.__class__.__name__) :]  # noqa: E203

    def __bool__(self) -> int:
        return all(bool(x) for x in self._storage)

    def __gt__(  # type: ignore[override]
        self, other: Tensor | NestedTensor | SupportsFloat
    ) -> bool | Tensor | NestedTensor:
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor(i > j for i, j in zip(self._storage, other._storage))
        if isinstance(other, (int, float, Tensor)):
            return NestedTensor([x > other for x in self._storage], **self._state)
        raise TypeError(f"> not supported between instances of '{type(self)}' and '{type(other)}'")

    def __ge__(  # type: ignore[override]
        self, other: Tensor | NestedTensor | SupportsFloat
    ) -> bool | Tensor | NestedTensor:
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor(i >= j for i, j in zip(self._storage, other._storage))
        if isinstance(other, (int, float, Tensor)):
            return NestedTensor([x >= other for x in self._storage], **self._state)
        raise TypeError(f">= not supported between instances of '{type(self)}' and '{type(other)}'")

    def __eq__(  # type: ignore[override]
        self, other: Tensor | NestedTensor | SupportsFloat
    ) -> bool | Tensor | NestedTensor:
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor(i == j for i, j in zip(self._storage, other._storage))
        if isinstance(other, (int, float, Tensor)):
            return NestedTensor([x == other for x in self._storage], **self._state)
        return False

    def __ne__(  # type: ignore[override]
        self, other: Tensor | NestedTensor | SupportsFloat
    ) -> bool | Tensor | NestedTensor:
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor(i != j for i, j in zip(self._storage, other._storage))
        if isinstance(other, (int, float, Tensor)):
            return NestedTensor([x != other for x in self._storage], **self._state)
        return True

    def __le__(  # type: ignore[override]
        self, other: Tensor | NestedTensor | SupportsFloat
    ) -> bool | Tensor | NestedTensor:
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor(i <= j for i, j in zip(self._storage, other._storage))
        if isinstance(other, (int, float, Tensor)):
            return NestedTensor([x <= other for x in self._storage], **self._state)
        raise TypeError(f"<= not supported between instances of '{type(self)}' and '{type(other)}'")

    def __lt__(  # type: ignore[override]
        self, other: Tensor | NestedTensor | SupportsFloat
    ) -> bool | Tensor | NestedTensor:
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor(i < j for i, j in zip(self._storage, other._storage))
        if isinstance(other, (int, float, Tensor)):
            return NestedTensor([x < other for x in self._storage], **self._state)
        raise TypeError(f"< not supported between instances of '{type(self)}' and '{type(other)}'")

    def __abs__(self):
        return NestedTensor([abs(value) for value in self._storage], **self._state)

    def __add__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([x + y for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([value + other for value in self._storage], **self._state)

    def __radd__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([y + x for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([other + value for value in self._storage], **self._state)

    def __iadd__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if hasattr(other, "to"):
            other = other.to(self.dtype)
        if isinstance(other, NestedTensor):
            for x, y in zip(self._storage, other._storage):
                x += y
        else:
            for value in self._storage:
                value += other
        return self

    def __sub__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([x - y for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([value - other for value in self._storage], **self._state)

    def __rsub__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([y - x for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([other - value for value in self._storage], **self._state)

    def __isub__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if hasattr(other, "to"):
            other = other.to(self.dtype)
        if isinstance(other, NestedTensor):
            for x, y in zip(self._storage, other._storage):
                x -= y
        else:
            for value in self._storage:
                value -= other
        return self

    def __pos__(self):
        return NestedTensor([+x for x in self._storage])

    def __neg__(self):
        return NestedTensor([-x for x in self._storage])

    def __invert__(self):
        return NestedTensor([~x for x in self._storage])

    def __mul__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([x * y for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([value * other for value in self._storage], **self._state)

    def __rmul__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([y * x for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([other * value for value in self._storage], **self._state)

    def __imul__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if hasattr(other, "to"):
            other = other.to(self.dtype)
        if isinstance(other, NestedTensor):
            for x, y in zip(self._storage, other._storage):
                x *= y
        else:
            for value in self._storage:
                value *= other
        return self

    def __pow__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([x**y for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([value**other for value in self._storage], **self._state)

    def __rpow__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([y**x for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([other**value for value in self._storage], **self._state)

    def __ipow__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if hasattr(other, "to"):
            other = other.to(self.dtype)
        if isinstance(other, NestedTensor):
            for x, y in zip(self._storage, other._storage):
                x **= y
        else:
            for value in self._storage:
                value **= other
        return self

    def __matmul__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([x @ y for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([value @ other for value in self._storage], **self._state)

    def __rmatmul__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([y @ x for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([other @ value for value in self._storage], **self._state)

    def __imatmul__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if hasattr(other, "to"):
            other = other.to(self.dtype)
        if isinstance(other, NestedTensor):
            for x, y in zip(self._storage, other._storage):
                x @= y
        else:
            for value in self._storage:
                value @= other
        return self

    def __truediv__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([x / y for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([value / other for value in self._storage], **self._state)

    def __rtruediv__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([y / x for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([other / value for value in self._storage], **self._state)

    def __itruediv__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if hasattr(other, "to"):
            other = other.to(self.dtype)
        if isinstance(other, NestedTensor):
            for x, y in zip(self._storage, other._storage):
                x /= y
        else:
            for value in self._storage:
                value /= other
        return self

    def __floordiv__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([x // y for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([value // other for value in self._storage], **self._state)

    def __rfloordiv__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([y // x for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([other // value for value in self._storage], **self._state)

    def __ifloordiv__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if hasattr(other, "to"):
            other = other.to(self.dtype)
        if isinstance(other, NestedTensor):
            for x, y in zip(self._storage, other._storage):
                x //= y
        else:
            for value in self._storage:
                value //= other
        return self

    def __mod__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([x % y for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([value % other for value in self._storage], **self._state)

    def __rmod__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([y % x for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([other % value for value in self._storage], **self._state)

    def __imod__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if hasattr(other, "to"):
            other = other.to(self.dtype)
        if isinstance(other, NestedTensor):
            for x, y in zip(self._storage, other._storage):
                x %= y
        else:
            for value in self._storage:
                value %= other
        return self

    def __and__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([x & y for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([value & other for value in self._storage], **self._state)

    def __rand__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([y & x for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([other & value for value in self._storage], **self._state)

    def __iand__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if hasattr(other, "to"):
            other = other.to(self.dtype)
        if isinstance(other, NestedTensor):
            for x, y in zip(self._storage, other._storage):
                x &= y
        else:
            for value in self._storage:
                value &= other
        return self

    def __or__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([x | y for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([value | other for value in self._storage], **self._state)

    def __ror__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([y | x for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([other | value for value in self._storage], **self._state)

    def __ior__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if hasattr(other, "to"):
            other = other.to(self.dtype)
        if isinstance(other, NestedTensor):
            for x, y in zip(self._storage, other._storage):
                x |= y
        else:
            for value in self._storage:
                value |= other
        return self

    def __xor__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([x ^ y for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([value ^ other for value in self._storage], **self._state)

    def __rxor__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([y ^ x for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([other ^ value for value in self._storage], **self._state)

    def __ixor__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if hasattr(other, "to"):
            other = other.to(self.dtype)
        if isinstance(other, NestedTensor):
            for x, y in zip(self._storage, other._storage):
                x ^= y
        else:
            for value in self._storage:
                value ^= other
        return self

    def __lshift__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([x << y for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([value << other for value in self._storage], **self._state)

    def __rlshift__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([y << x for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([other << value for value in self._storage], **self._state)

    def __ilshift__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if hasattr(other, "to"):
            other = other.to(self.dtype)
        if isinstance(other, NestedTensor):
            for x, y in zip(self._storage, other._storage):
                x <<= y
        else:
            for value in self._storage:
                value <<= other
        return self

    def __rshift__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([x >> y for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([value >> other for value in self._storage], **self._state)

    def __rrshift__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(other, NestedTensor):
            return NestedTensor([y >> x for x, y in zip(self._storage, other._storage)], **self._state)
        return NestedTensor([other >> value for value in self._storage], **self._state)

    def __irshift__(self, other: Tensor | NestedTensor | SupportsFloat):
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if hasattr(other, "to"):
            other = other.to(self.dtype)
        if isinstance(other, NestedTensor):
            for x, y in zip(self._storage, other._storage):
                x >>= y
        else:
            for value in self._storage:
                value >>= other
        return self

    def all(self, dim: int | None = None, keepdim: bool = False) -> bool | Tensor | NestedTensor:
        r"""
        Tests if all elements in NestedTensor evaluate to True.

        Examples:
            >>> nested_tensor = NestedTensor([torch.ones(2, 4, dtype=torch.bool), torch.ones(3, 5, dtype=torch.bool)])
            >>> nested_tensor.all()
            tensor(True)
            >>> nested_tensor.all(dim=0)
            tensor([True, True])
            >>> nested_tensor.all(dim=0, keepdim=True)
            tensor([[True, True]])
            >>> nested_tensor.all(dim=1)
            NestedTensor([[ True,  True,  True,  True, False],
                    [ True,  True,  True,  True,  True]])
            >>> nested_tensor.all(dim=1, keepdim=True)
            NestedTensor([[[ True,  True,  True,  True, False]],
            <BLANKLINE>
                    [[ True,  True,  True,  True,  True]]])
            >>> nested_tensor.batch_first = False
            >>> nested_tensor.all(dim=1)
            tensor([True, True])
            >>> nested_tensor.all(dim=0)
            NestedTensor([[ True,  True],
                    [ True,  True],
                    [ True,  True],
                    [ True,  True],
                    [False,  True]])
            >>> nested_tensor.all(dim=-2)
            tensor([True, True])
        """

        if dim is None:
            return torch.tensor(all(i.all() for i in self._storage))

        if dim < 0:
            dim += self.dim()

        if (self.batch_first and dim == 0) or (not self.batch_first and dim == 1):
            if keepdim:
                return torch.tensor([i.all() for i in self._storage]).unsqueeze(0 if self.batch_first else 1)
            return torch.tensor([i.all() for i in self._storage])

        if self.batch_first or dim != 0:
            dim -= 1

        ret = [i.all(dim=dim, keepdim=keepdim) for i in self._storage]
        try:
            return torch.stack(ret)
        except (RuntimeError, ValueError):
            return NestedTensor(ret, **self._state)

    def dim(self) -> int:
        r"""
        Number of dimension of the NestedTensor.

        Examples:
            >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
            >>> nested_tensor.dim()
            2
        """

        return self._dim(self._storage)

    @method_cache(maxsize=1)
    def _dim(self, storage: Tuple[Tensor, ...]) -> int:  # type: ignore[name-defined]
        return max(t.dim() for t in storage) + 1

    def max(self, dim: int | None = None, keepdim: bool = False) -> Tensor | NestedTensor:
        if dim is None:
            return torch.stack([t.max() for t in self._storage]).max()

        if dim < 0:
            dim += self.dim()

        if (self.batch_first and dim == 0) or (not self.batch_first and dim == 1):
            if keepdim:
                return torch.stack([t.max() for t in self._storage]).unsqueeze(0 if self.batch_first else 1)
            return torch.stack([t.max() for t in self._storage])

        if self.batch_first or dim != 0:
            dim -= 1

        ret = [i.max(dim=dim, keepdim=keepdim) for i in self._storage]
        try:
            return torch.stack(ret)
        except (RuntimeError, ValueError):
            return NestedTensor(ret, **self._state)

    def mean(
        self,
        dim: int | None = None,
        keepdim: bool = False,
        *,
        dtype: torch.dtype | None = None,  # type: ignore[name-defined]
    ) -> Tensor | NestedTensor:
        if dim is None:
            return sum([t.sum(dtype=dtype) for t in self._storage]) / self.numel()

        if dim < 0:
            dim += self.dim()

        if (self.batch_first and dim == 0) or (not self.batch_first and dim == 1):
            if keepdim:
                return torch.stack([t.mean(dtype=dtype) for t in self._storage]).unsqueeze(0 if self.batch_first else 1)
            return torch.stack([t.mean(dtype=dtype) for t in self._storage])

        if self.batch_first or dim != 0:
            dim -= 1

        ret = [i.mean(dim=dim, keepdim=keepdim, dtype=dtype) for i in self._storage]
        try:
            return torch.stack(ret)
        except (RuntimeError, ValueError):
            return NestedTensor(ret, **self._state)

    def min(self, dim: int | None = None, keepdim: bool = False) -> Tensor | NestedTensor:
        if dim is None:
            return torch.stack([t.min() for t in self._storage]).min()

        if dim < 0:
            dim += self.dim()

        if (self.batch_first and dim == 0) or (not self.batch_first and dim == 1):
            if keepdim:
                return torch.stack([t.min() for t in self._storage]).unsqueeze(0 if self.batch_first else 1)
            return torch.stack([t.min() for t in self._storage])

        if self.batch_first or dim != 0:
            dim -= 1

        ret = [i.min(dim=dim, keepdim=keepdim) for i in self._storage]
        try:
            return torch.stack(ret)
        except (RuntimeError, ValueError):
            return NestedTensor(ret, **self._state)

    @property
    def ndim(self) -> int:
        r"""
        Alias for `dim()`.
        """

        return self.dim()

    def numel(self) -> int:
        r"""
        Number of elements in the NestedTensor.

        Examples:
            >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
            >>> nested_tensor.numel()
            5
        """

        return sum(t.numel() for t in self._storage)

    def requires_grad_(self, requires_grad: bool = True):
        self.requires_grad = requires_grad
        for t in self._storage:
            t.requires_grad = requires_grad
        return self

    def reshape(self, *shape) -> Tensor:
        r"""
        Returns a torch tensor with a different shape.

        Note:
            since NestedTensor is a collection of tensors, the reshape operation is ambiguous.

            Therefore, it is converted to a tensor and then reshaped.

        Args:
            shape: The desired size of each dimension.

        Examples:
            >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
            >>> nested_tensor.reshape(3, 2)
            tensor([[1, 2],
                    [3, 4],
                    [5, 0]])
            >>> nested_tensor.reshape(2, 3)
            tensor([[1, 2, 3],
                    [4, 5, 0]])
        """

        return self.tensor.reshape(*shape)

    @property
    def shape(self) -> torch.Size | int:  # type: ignore[name-defined]
        r"""
        Alias for `size()`.
        """

        return self.size()

    def size(self, dim: int | None = None) -> torch.Size | int:  # type: ignore[name-defined]
        r"""
        Returns the size of the self `NestedTensor`.

        Args:
            dim: If not specified, the returned value is a `torch.Size`, a subclass of `tuple`.
                If specified, returns an `int` holding the size of that dimension.
                Defaults to `None`.

        Examples:
            >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
            >>> nested_tensor.size()
            torch.Size([2, 3])
            >>> nested_tensor.size(0)
            2
            >>> nested_tensor[1] = torch.tensor([4, 5, 6, 7])
            >>> nested_tensor.shape
            torch.Size([2, 4])
            >>> nested_tensor.size(1)
            4
        """

        return self._size(self._storage, dim, self.batch_first)

    @method_cache(maxsize=1)
    def _size(
        self,
        storage: Tuple[Tensor, ...],
        dim: int | None = None,
        batch_first: bool = True,
    ) -> torch.Size | int:  # type: ignore[name-defined]
        if dim is not None:
            if dim == 0:
                return len(storage)
            return max(t.size(dim - 1) for t in storage)
        if max(t.dim() for t in storage) == 0:
            return torch.Size((len(storage),))
        ndim = max(t.dim() for t in storage)
        size = [max(t.shape[i] if i < len(t.shape) else 0 for t in storage) for i in range(ndim)]
        size.insert(0 if batch_first else 1, len(storage))
        return torch.Size(size)

    def sum(self, dim: int | None = None, keepdim: bool = False) -> Tensor | NestedTensor:
        if dim is None:
            return torch.stack([t.sum() for t in self._storage]).sum()

        if dim < 0:
            dim += self.dim()

        if (self.batch_first and dim == 0) or (not self.batch_first and dim == 1):
            if keepdim:
                return torch.stack([t.sum() for t in self._storage]).unsqueeze(0 if self.batch_first else 1)
            return torch.stack([t.sum() for t in self._storage])

        if self.batch_first or dim != 0:
            dim -= 1

        ret = [i.sum(dim=dim, keepdim=keepdim) for i in self._storage]
        try:
            return torch.stack(ret)
        except (RuntimeError, ValueError):
            return NestedTensor(ret, **self._state)

    def to(self, *args, **kwargs):
        return NestedTensor(
            tuple(t.to(*args, **kwargs) for t in self._storage),
            **self.__state__(return_dtype=False, return_device=False),
        )

    def tolist(self) -> list:
        r"""
        Convert a NestedTensor to a list of lists of values.

        Examples:
            >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
            >>> nested_tensor.tolist()
            [[1, 2, 3], [4, 5]]
        """

        return [t.tolist() for t in self._storage]

    def view(self, *shape) -> Tensor:
        r"""
        Returns a torch tensor with a different shape.

        Note:
            since NestedTensor is a collection of tensors, the view operation is ambiguous.

            Therefore, it is converted to a tensor and then reshaped.

        Args:
            shape: The desired size of each dimension.

        Examples:
            >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
            >>> nested_tensor.view(3, 2)
            tensor([[1, 2],
                    [3, 4],
                    [5, 0]])
            >>> nested_tensor.view(2, 3)
            tensor([[1, 2, 3],
                    [4, 5, 0]])
        """

        return self.tensor.view(*shape)

    def where(self, condition: Tensor | NestedTensor, other: Tensor | NestedTensor | SupportsFloat) -> Self:
        r"""
        Return a NestedTensor of elements selected from either self or other, depending on condition.

        Examples:
            >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
            >>> nested_tensor.where(nested_tensor > 2, torch.tensor([[6, 5, 4], [3, 2, 1]]))
            NestedTensor([[6, 5, 3],
                    [4, 5, 0]])
            >>> nested_tensor.where(nested_tensor > 2, NestedTensor([[6, 5, 4], [3, 2]]))
            NestedTensor([[6, 5, 3],
                    [4, 5, 0]])
            >>> nested_tensor.where(torch.tensor(True), NestedTensor([[6, 5, 4], [3, 2]]))
            NestedTensor([[1, 2, 3],
                    [4, 5, 0]])
        """

        if isinstance(condition, Tensor) and self.shape == condition.shape:
            condition = self.nested_like(condition)
        if isinstance(other, Tensor) and self.shape == other.shape:
            other = self.nested_like(other)
        if isinstance(condition, NestedTensor) and isinstance(other, NestedTensor):
            return self.__class__(
                [x.where(c, y) for x, c, y in zip(self._storage, condition._storage, other._storage)], **self._state
            )
        if isinstance(condition, NestedTensor):
            return self.__class__([x.where(c, other) for x, c in zip(self._storage, condition._storage)], **self._state)
        if isinstance(other, NestedTensor):
            return self.__class__([x.where(condition, y) for x, y in zip(self._storage, other._storage)], **self._state)
        return self.__class__(x.where(condition, other) for x in self._storage)

tensor_mask property

Python
tensor_mask: Tuple[Tensor, Tensor]

Return a tuple of padded tensor and mask tensor.

Examples:

Python Console Session
1
2
3
4
5
>>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
>>> nested_tensor.tensor_mask
(tensor([[1, 2, 3],
        [4, 5, 0]]), tensor([[ True,  True,  True],
        [ True,  True, False]]))

tensor property

Python
tensor: Tensor

Return a single tensor by padding all the tensors.

Examples:

Python Console Session
1
2
3
4
>>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
>>> nested_tensor.tensor
tensor([[1, 2, 3],
        [4, 5, 0]])

mask property

Python
mask: Tensor

Padding mask of tensor.

Examples:

Python Console Session
1
2
3
4
>>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
>>> nested_tensor.mask
tensor([[ True,  True,  True],
        [ True,  True, False]])

concat property

Python
concat: Tensor

Concat tensor in padding dim.

This is particularly useful when calculating loss or passing Linear to avoid unnecessary computation.

Examples:

Python Console Session
>>> nested_tensor = NestedTensor([torch.randn(9, 8), torch.randn(11, 8)])
>>> nested_tensor.concat.shape
torch.Size([20, 8])
>>> nested_tensor = NestedTensor([torch.randn(9, 9, 8), torch.randn(11, 11, 8)])
>>> nested_tensor.concat.shape
torch.Size([202, 8])
>>> nested_tensor = NestedTensor([torch.randn(9, 9, 8, 6), torch.randn(11, 11, 8, 6)])
>>> nested_tensor.concat.shape
torch.Size([202, 8, 6])
>>> nested_tensor = NestedTensor([torch.randn(9, 9, 8, 7), torch.randn(11, 11, 8, 6)])
>>> nested_tensor.concat.shape
torch.Size([1293, 8])
>>> nested_tensor = NestedTensor([torch.randn(1, 9, 9, 5), torch.randn(1, 11, 11, 5)])

torch property

Python
torch: Tensor

Create a torch.nested.nested_tensor object from self.

Examples:

Python Console Session
1
2
3
4
5
6
>>> nested_tensor = NestedTensor([[2, 3, 5], [7, 8]])
>>> nested_tensor.torch
nested_tensor([
  tensor([2, 3, 5]),
  tensor([7, 8])
])

occupancy property

Python
occupancy: float

Occupancy of the NestedTensor.

Examples:

Python Console Session
1
2
3
>>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3, 4]), torch.tensor([5, 6])])
>>> nested_tensor.occupancy
0.75

ndim property

Python
ndim: int

Alias for dim().

shape property

Python
shape: Size | int

Alias for size().

unbind

Python
unbind(dim: int = 0) -> Tuple[Tensor, ...]

Unbind the NestedTensor.

Source code in danling/tensor/nested_tensor.py
Python
def unbind(self, dim: int = 0) -> Tuple[Tensor, ...]:
    r"""
    Unbind the NestedTensor.
    """
    if dim != 0:
        raise ValueError(f"NestedTensor can only be unbound along dimension 0, got dimension {dim} instead.")
    return self._storage

from_tensor_mask classmethod

Python
from_tensor_mask(tensor: Tensor, mask: Tensor, **kwargs)

Build a NestedTensor object from a padded Tensor and corresponding mask Tensor.

Parameters:

Name Type Description Default
tensor
Tensor

Padded Tensor.

required
mask
Tensor

Tensor Mask.

required

Examples:

Python Console Session
>>> padded_tensor = torch.tensor([[1, 2, 3, 0, 0],
...                                [4, 5, 0, 0, 0],
...                                [6, 7, 8, 9, 0]])
>>> mask_tensor = torch.tensor([[1, 1, 1, 0, 0],
...                             [1, 1, 0, 0, 0],
...                             [1, 1, 1, 1, 0]])
>>> nested_tensor = NestedTensor.from_tensor_mask(padded_tensor, mask_tensor)
>>> nested_tensor
NestedTensor([[1, 2, 3, 0],
        [4, 5, 0, 0],
        [6, 7, 8, 9]])
Source code in danling/tensor/nested_tensor.py
Python
@classmethod
def from_tensor_mask(cls, tensor: Tensor, mask: Tensor, **kwargs):
    r"""
    Build a `NestedTensor` object from a padded `Tensor` and corresponding mask `Tensor`.

    Args:
        tensor: Padded Tensor.
        mask: Tensor Mask.

    Examples:
        >>> padded_tensor = torch.tensor([[1, 2, 3, 0, 0],
        ...                                [4, 5, 0, 0, 0],
        ...                                [6, 7, 8, 9, 0]])
        >>> mask_tensor = torch.tensor([[1, 1, 1, 0, 0],
        ...                             [1, 1, 0, 0, 0],
        ...                             [1, 1, 1, 1, 0]])
        >>> nested_tensor = NestedTensor.from_tensor_mask(padded_tensor, mask_tensor)
        >>> nested_tensor
        NestedTensor([[1, 2, 3, 0],
                [4, 5, 0, 0],
                [6, 7, 8, 9]])
    """

    if mask.ndim == 1:
        return cls(tensor, **kwargs)
    if mask.ndim == 2:
        return cls((t[slice(0, m.sum())] for t, m in zip(tensor, mask)), **kwargs)
    return cls(
        (
            t[[slice(0, (m.sum(dim=dim) > 0).sum().item()) for dim in reversed(range(m.dim()))]]
            for t, m in zip(tensor, mask)
        ),
        **kwargs,
    )

nested_like

Python
nested_like(tensor: Tensor, strict: bool = True) -> Self

Create a new NestedTensor from a Tensor. The newly created NestedTensor will have the same shape as current NestedTensor.

Parameters:

Name Type Description Default
tensor
Tensor

The tensor to be converted to NestedTensor.

required
strict
bool

Check if the shape of tensor is the same as the current NestedTensor.

True

Examples:

Python Console Session
>>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
>>> (nested_tensor == nested_tensor.nested_like(nested_tensor)).all()
tensor(True)
>>> tensor = nested_tensor.tensor
>>> (nested_tensor == nested_tensor.nested_like(tensor)).all()
tensor(True)
>>> f = nested_tensor.nested_like(torch.randn(2, 2))
Traceback (most recent call last):
ValueError: The shape of NestedTensor and input tensor does not match, torch.Size([2, 3]) != torch.Size([2, 2])
>>> p = nested_tensor.nested_like(torch.randn(2, 2), False)
>>> p = nested_tensor.nested_like(torch.randn(3, 3), False)
Traceback (most recent call last):
ValueError: The batch size of NestedTensor and input tensor does not match, 2 != 3
Source code in danling/tensor/nested_tensor.py
Python
def nested_like(self, tensor: Tensor, strict: bool = True) -> Self:
    r"""
    Create a new `NestedTensor` from a `Tensor`.
    The newly created `NestedTensor` will have the same shape as current `NestedTensor`.

    Args:
        tensor: The tensor to be converted to `NestedTensor`.
        strict: Check if the shape of `tensor` is the same as the current `NestedTensor`.

    Examples:
        >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
        >>> (nested_tensor == nested_tensor.nested_like(nested_tensor)).all()
        tensor(True)
        >>> tensor = nested_tensor.tensor
        >>> (nested_tensor == nested_tensor.nested_like(tensor)).all()
        tensor(True)
        >>> f = nested_tensor.nested_like(torch.randn(2, 2))
        Traceback (most recent call last):
        ValueError: The shape of NestedTensor and input tensor does not match, torch.Size([2, 3]) != torch.Size([2, 2])
        >>> p = nested_tensor.nested_like(torch.randn(2, 2), False)
        >>> p = nested_tensor.nested_like(torch.randn(3, 3), False)
        Traceback (most recent call last):
        ValueError: The batch size of NestedTensor and input tensor does not match, 2 != 3
    """  # noqa: E501

    if isinstance(tensor, NestedTensor):
        return tensor.clone()

    if strict and self.shape != tensor.shape:
        raise ValueError(
            f"The shape of NestedTensor and input tensor does not match, {self.shape} != {tensor.shape}"
        )
    if self.size(0) != tensor.size(0):
        raise ValueError(
            f"The batch size of NestedTensor and input tensor does not match, {self.size(0)} != {tensor.size(0)}"
        )
    return self.__class__([o[tuple(slice(0, dim) for dim in t.shape)] for t, o in zip(self._storage, tensor)])

__setitem__

Python
__setitem__(index: int | slice | list | tuple, value: Tensor | NestedTensor) -> None

Set values in the NestedTensor at the specified index.

Parameters:

Name Type Description Default
index
int | slice | list | tuple

The index to modify. Can be an integer, slice, list, or tuple.

required
value
Tensor | NestedTensor

The new value to set. Can be a Tensor or NestedTensor.

required

Examples:

Python Console Session
1
2
3
4
5
6
7
>>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
>>> nested_tensor[0] = torch.tensor([6, 7, 8])
>>> nested_tensor[0]
tensor([6, 7, 8])
>>> nested_tensor[1] = torch.tensor([9, 10, 11, 12])
>>> nested_tensor.shape
torch.Size([2, 4])
Source code in danling/tensor/nested_tensor.py
Python
def __setitem__(self, index: int | slice | list | tuple, value: Tensor | NestedTensor) -> None:
    """
    Set values in the NestedTensor at the specified index.

    Args:
        index: The index to modify. Can be an integer, slice, list, or tuple.
        value: The new value to set. Can be a Tensor or NestedTensor.

    Examples:
        >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
        >>> nested_tensor[0] = torch.tensor([6, 7, 8])
        >>> nested_tensor[0]
        tensor([6, 7, 8])
        >>> nested_tensor[1] = torch.tensor([9, 10, 11, 12])
        >>> nested_tensor.shape
        torch.Size([2, 4])
    """
    if isinstance(index, int):
        if isinstance(value, NestedTensor):
            if len(value._storage) != 1:
                raise ValueError(
                    f"When setting with an integer index, value must have a single tensor, but got {len(value)}"
                )
            value = value._storage[0]
        if not isinstance(value, Tensor):
            value = torch.tensor(value, device=self.device)
        # Create a new list of tensors to modify
        storage_list = list(self._storage)
        storage_list[index] = value
        self._storage = tuple(storage_list)
    elif isinstance(index, (slice, list)):
        if isinstance(value, Tensor):
            # Convert tensor to NestedTensor if it's a regular tensor
            if value.dim() > 1 and value.size(0) > 1:
                value = NestedTensor(value.unbind(0))
            else:
                value = NestedTensor([value])

        if isinstance(index, slice):
            start, stop, step = index.indices(len(self._storage))
            indices = range(start, stop, step)
        else:
            indices = index  # type: ignore[assignment]

        if len(indices) != len(value._storage):
            raise ValueError(
                f"Size mismatch: tried to assign {len(value._storage)} values to {len(indices)} indices"
            )

        storage_list = list(self._storage)
        for i, idx in enumerate(indices):
            storage_list[idx] = value._storage[i]
        self._storage = tuple(storage_list)
    elif isinstance(index, tuple):
        if len(index) < 2:
            raise ValueError("Tuple index must have at least two elements")

        first_idx, rest_idx = index[0], index[1:]

        if isinstance(first_idx, int):
            # Handle single tensor modification
            storage_list = list(self._storage)
            tensor = storage_list[first_idx]
            tensor[rest_idx] = value
            storage_list[first_idx] = tensor
            self._storage = tuple(storage_list)
        elif isinstance(first_idx, (slice, list)):
            # Handle multiple tensor modification
            if isinstance(first_idx, slice):
                start, stop, step = first_idx.indices(len(self._storage))
                indices = range(start, stop, step)
            else:
                indices = first_idx  # type: ignore[assignment]

            storage_list = list(self._storage)
            for idx in indices:
                tensor = storage_list[idx]
                tensor[rest_idx] = value
                storage_list[idx] = tensor
            self._storage = tuple(storage_list)
        else:
            raise ValueError(f"Unsupported first index type {type(first_idx)}")
    else:
        raise ValueError(f"Unsupported index type {type(index)}")

all

Python
all(dim: int | None = None, keepdim: bool = False) -> bool | Tensor | NestedTensor

Tests if all elements in NestedTensor evaluate to True.

Examples:

Python Console Session
>>> nested_tensor = NestedTensor([torch.ones(2, 4, dtype=torch.bool), torch.ones(3, 5, dtype=torch.bool)])
>>> nested_tensor.all()
tensor(True)
>>> nested_tensor.all(dim=0)
tensor([True, True])
>>> nested_tensor.all(dim=0, keepdim=True)
tensor([[True, True]])
>>> nested_tensor.all(dim=1)
NestedTensor([[ True,  True,  True,  True, False],
        [ True,  True,  True,  True,  True]])
>>> nested_tensor.all(dim=1, keepdim=True)
NestedTensor([[[ True,  True,  True,  True, False]],

        [[ True,  True,  True,  True,  True]]])
>>> nested_tensor.batch_first = False
>>> nested_tensor.all(dim=1)
tensor([True, True])
>>> nested_tensor.all(dim=0)
NestedTensor([[ True,  True],
        [ True,  True],
        [ True,  True],
        [ True,  True],
        [False,  True]])
>>> nested_tensor.all(dim=-2)
tensor([True, True])
Source code in danling/tensor/nested_tensor.py
Python
def all(self, dim: int | None = None, keepdim: bool = False) -> bool | Tensor | NestedTensor:
    r"""
    Tests if all elements in NestedTensor evaluate to True.

    Examples:
        >>> nested_tensor = NestedTensor([torch.ones(2, 4, dtype=torch.bool), torch.ones(3, 5, dtype=torch.bool)])
        >>> nested_tensor.all()
        tensor(True)
        >>> nested_tensor.all(dim=0)
        tensor([True, True])
        >>> nested_tensor.all(dim=0, keepdim=True)
        tensor([[True, True]])
        >>> nested_tensor.all(dim=1)
        NestedTensor([[ True,  True,  True,  True, False],
                [ True,  True,  True,  True,  True]])
        >>> nested_tensor.all(dim=1, keepdim=True)
        NestedTensor([[[ True,  True,  True,  True, False]],
        <BLANKLINE>
                [[ True,  True,  True,  True,  True]]])
        >>> nested_tensor.batch_first = False
        >>> nested_tensor.all(dim=1)
        tensor([True, True])
        >>> nested_tensor.all(dim=0)
        NestedTensor([[ True,  True],
                [ True,  True],
                [ True,  True],
                [ True,  True],
                [False,  True]])
        >>> nested_tensor.all(dim=-2)
        tensor([True, True])
    """

    if dim is None:
        return torch.tensor(all(i.all() for i in self._storage))

    if dim < 0:
        dim += self.dim()

    if (self.batch_first and dim == 0) or (not self.batch_first and dim == 1):
        if keepdim:
            return torch.tensor([i.all() for i in self._storage]).unsqueeze(0 if self.batch_first else 1)
        return torch.tensor([i.all() for i in self._storage])

    if self.batch_first or dim != 0:
        dim -= 1

    ret = [i.all(dim=dim, keepdim=keepdim) for i in self._storage]
    try:
        return torch.stack(ret)
    except (RuntimeError, ValueError):
        return NestedTensor(ret, **self._state)

dim

Python
dim() -> int

Number of dimension of the NestedTensor.

Examples:

Python Console Session
1
2
3
>>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
>>> nested_tensor.dim()
2
Source code in danling/tensor/nested_tensor.py
Python
def dim(self) -> int:
    r"""
    Number of dimension of the NestedTensor.

    Examples:
        >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
        >>> nested_tensor.dim()
        2
    """

    return self._dim(self._storage)

numel

Python
numel() -> int

Number of elements in the NestedTensor.

Examples:

Python Console Session
1
2
3
>>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
>>> nested_tensor.numel()
5
Source code in danling/tensor/nested_tensor.py
Python
def numel(self) -> int:
    r"""
    Number of elements in the NestedTensor.

    Examples:
        >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
        >>> nested_tensor.numel()
        5
    """

    return sum(t.numel() for t in self._storage)

reshape

Python
reshape(*shape) -> Tensor

Returns a torch tensor with a different shape.

Note

since NestedTensor is a collection of tensors, the reshape operation is ambiguous.

Therefore, it is converted to a tensor and then reshaped.

Parameters:

Name Type Description Default
shape

The desired size of each dimension.

()

Examples:

Python Console Session
1
2
3
4
5
6
7
8
>>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
>>> nested_tensor.reshape(3, 2)
tensor([[1, 2],
        [3, 4],
        [5, 0]])
>>> nested_tensor.reshape(2, 3)
tensor([[1, 2, 3],
        [4, 5, 0]])
Source code in danling/tensor/nested_tensor.py
Python
def reshape(self, *shape) -> Tensor:
    r"""
    Returns a torch tensor with a different shape.

    Note:
        since NestedTensor is a collection of tensors, the reshape operation is ambiguous.

        Therefore, it is converted to a tensor and then reshaped.

    Args:
        shape: The desired size of each dimension.

    Examples:
        >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
        >>> nested_tensor.reshape(3, 2)
        tensor([[1, 2],
                [3, 4],
                [5, 0]])
        >>> nested_tensor.reshape(2, 3)
        tensor([[1, 2, 3],
                [4, 5, 0]])
    """

    return self.tensor.reshape(*shape)

size

Python
size(dim: int | None = None) -> Size | int

Returns the size of the self NestedTensor.

Parameters:

Name Type Description Default
dim
int | None

If not specified, the returned value is a torch.Size, a subclass of tuple. If specified, returns an int holding the size of that dimension. Defaults to None.

None

Examples:

Python Console Session
>>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
>>> nested_tensor.size()
torch.Size([2, 3])
>>> nested_tensor.size(0)
2
>>> nested_tensor[1] = torch.tensor([4, 5, 6, 7])
>>> nested_tensor.shape
torch.Size([2, 4])
>>> nested_tensor.size(1)
4
Source code in danling/tensor/nested_tensor.py
Python
def size(self, dim: int | None = None) -> torch.Size | int:  # type: ignore[name-defined]
    r"""
    Returns the size of the self `NestedTensor`.

    Args:
        dim: If not specified, the returned value is a `torch.Size`, a subclass of `tuple`.
            If specified, returns an `int` holding the size of that dimension.
            Defaults to `None`.

    Examples:
        >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
        >>> nested_tensor.size()
        torch.Size([2, 3])
        >>> nested_tensor.size(0)
        2
        >>> nested_tensor[1] = torch.tensor([4, 5, 6, 7])
        >>> nested_tensor.shape
        torch.Size([2, 4])
        >>> nested_tensor.size(1)
        4
    """

    return self._size(self._storage, dim, self.batch_first)

tolist

Python
tolist() -> list

Convert a NestedTensor to a list of lists of values.

Examples:

Python Console Session
1
2
3
>>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
>>> nested_tensor.tolist()
[[1, 2, 3], [4, 5]]
Source code in danling/tensor/nested_tensor.py
Python
def tolist(self) -> list:
    r"""
    Convert a NestedTensor to a list of lists of values.

    Examples:
        >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
        >>> nested_tensor.tolist()
        [[1, 2, 3], [4, 5]]
    """

    return [t.tolist() for t in self._storage]

view

Python
view(*shape) -> Tensor

Returns a torch tensor with a different shape.

Note

since NestedTensor is a collection of tensors, the view operation is ambiguous.

Therefore, it is converted to a tensor and then reshaped.

Parameters:

Name Type Description Default
shape

The desired size of each dimension.

()

Examples:

Python Console Session
1
2
3
4
5
6
7
8
>>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
>>> nested_tensor.view(3, 2)
tensor([[1, 2],
        [3, 4],
        [5, 0]])
>>> nested_tensor.view(2, 3)
tensor([[1, 2, 3],
        [4, 5, 0]])
Source code in danling/tensor/nested_tensor.py
Python
def view(self, *shape) -> Tensor:
    r"""
    Returns a torch tensor with a different shape.

    Note:
        since NestedTensor is a collection of tensors, the view operation is ambiguous.

        Therefore, it is converted to a tensor and then reshaped.

    Args:
        shape: The desired size of each dimension.

    Examples:
        >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
        >>> nested_tensor.view(3, 2)
        tensor([[1, 2],
                [3, 4],
                [5, 0]])
        >>> nested_tensor.view(2, 3)
        tensor([[1, 2, 3],
                [4, 5, 0]])
    """

    return self.tensor.view(*shape)

where

Python
where(condition: Tensor | NestedTensor, other: Tensor | NestedTensor | SupportsFloat) -> Self

Return a NestedTensor of elements selected from either self or other, depending on condition.

Examples:

Python Console Session
>>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
>>> nested_tensor.where(nested_tensor > 2, torch.tensor([[6, 5, 4], [3, 2, 1]]))
NestedTensor([[6, 5, 3],
        [4, 5, 0]])
>>> nested_tensor.where(nested_tensor > 2, NestedTensor([[6, 5, 4], [3, 2]]))
NestedTensor([[6, 5, 3],
        [4, 5, 0]])
>>> nested_tensor.where(torch.tensor(True), NestedTensor([[6, 5, 4], [3, 2]]))
NestedTensor([[1, 2, 3],
        [4, 5, 0]])
Source code in danling/tensor/nested_tensor.py
Python
def where(self, condition: Tensor | NestedTensor, other: Tensor | NestedTensor | SupportsFloat) -> Self:
    r"""
    Return a NestedTensor of elements selected from either self or other, depending on condition.

    Examples:
        >>> nested_tensor = NestedTensor([torch.tensor([1, 2, 3]), torch.tensor([4, 5])])
        >>> nested_tensor.where(nested_tensor > 2, torch.tensor([[6, 5, 4], [3, 2, 1]]))
        NestedTensor([[6, 5, 3],
                [4, 5, 0]])
        >>> nested_tensor.where(nested_tensor > 2, NestedTensor([[6, 5, 4], [3, 2]]))
        NestedTensor([[6, 5, 3],
                [4, 5, 0]])
        >>> nested_tensor.where(torch.tensor(True), NestedTensor([[6, 5, 4], [3, 2]]))
        NestedTensor([[1, 2, 3],
                [4, 5, 0]])
    """

    if isinstance(condition, Tensor) and self.shape == condition.shape:
        condition = self.nested_like(condition)
    if isinstance(other, Tensor) and self.shape == other.shape:
        other = self.nested_like(other)
    if isinstance(condition, NestedTensor) and isinstance(other, NestedTensor):
        return self.__class__(
            [x.where(c, y) for x, c, y in zip(self._storage, condition._storage, other._storage)], **self._state
        )
    if isinstance(condition, NestedTensor):
        return self.__class__([x.where(c, other) for x, c in zip(self._storage, condition._storage)], **self._state)
    if isinstance(other, NestedTensor):
        return self.__class__([x.where(condition, y) for x, y in zip(self._storage, other._storage)], **self._state)
    return self.__class__(x.where(condition, other) for x in self._storage)