1# Copyright 2016 The TensorFlow Authors. All Rights Reserved. 2# 3# Licensed under the Apache License, Version 2.0 (the "License"); 4# you may not use this file except in compliance with the License. 5# You may obtain a copy of the License at 6# 7# http://www.apache.org/licenses/LICENSE-2.0 8# 9# Unless required by applicable law or agreed to in writing, software 10# distributed under the License is distributed on an "AS IS" BASIS, 11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12# See the License for the specific language governing permissions and 13# limitations under the License. 14# ============================================================================== 15 16"""## Functions for working with arbitrarily nested sequences of elements. 17 18This module can perform operations on nested structures. A nested structure is a 19Python sequence, tuple (including `namedtuple`), or dict that can contain 20further sequences, tuples, and dicts. 21 22attr.s decorated classes (http://www.attrs.org) are also supported, in the 23same way as `namedtuple`. 24 25The utilities here assume (and do not check) that the nested structures form a 26'tree', i.e., no references in the structure of the input of these functions 27should be recursive. 28 29Example structures: `((3, 4), 5, (6, 7, (9, 10), 8))`, `(np.array(0), 30 (np.array([3, 4]), tf.constant([3, 4])))` 31""" 32 33from __future__ import absolute_import 34from __future__ import division 35from __future__ import print_function 36 37import collections as _collections 38 39import six as _six 40 41from tensorflow.python import pywrap_tensorflow as _pywrap_tensorflow 42from tensorflow.python.util.tf_export import tf_export 43 44 45_SHALLOW_TREE_HAS_INVALID_KEYS = ( 46 "The shallow_tree's keys are not a subset of the input_tree's keys. The " 47 "shallow_tree has the following keys that are not in the input_tree: {}.") 48 49_STRUCTURES_HAVE_MISMATCHING_TYPES = ( 50 "The two structures don't have the same sequence type. Input structure has " 51 "type {shallow_type}, while shallow structure has type {input_type}.") 52 53_INPUT_TREE_SMALLER_THAN_SHALLOW_TREE = ( 54 "The input_tree has fewer elements than the input_tree. Input structure " 55 "has length {input_size}, while shallow structure has length " 56 "{shallow_size}.") 57 58_IF_SHALLOW_IS_SEQ_INPUT_MUST_BE_SEQ = ( 59 "If shallow structure is a sequence, input must also be a sequence. " 60 "Input has type: {}.") 61 62 63def _get_attrs_items(obj): 64 """Returns a list of (name, value) pairs from an attrs instance. 65 66 The list will be sorted by name. 67 68 Args: 69 obj: an object. 70 71 Returns: 72 A list of (attr_name, attr_value) pairs, sorted by attr_name. 73 """ 74 attrs = getattr(obj.__class__, "__attrs_attrs__") 75 attr_names = sorted([a.name for a in attrs]) 76 return [(attr_name, getattr(obj, attr_name)) for attr_name in attr_names] 77 78 79def _sorted(dict_): 80 """Returns a sorted list of the dict keys, with error if keys not sortable.""" 81 try: 82 return sorted(dict_) 83 except TypeError: 84 raise TypeError("nest only supports dicts with sortable keys.") 85 86 87def _is_namedtuple(instance, strict=False): 88 """Returns True iff `instance` is a `namedtuple`. 89 90 Args: 91 instance: An instance of a Python object. 92 strict: If True, `instance` is considered to be a `namedtuple` only if 93 it is a "plain" namedtuple. For instance, a class inheriting 94 from a `namedtuple` will be considered to be a `namedtuple` 95 iff `strict=False`. 96 97 Returns: 98 True if `instance` is a `namedtuple`. 99 """ 100 return _pywrap_tensorflow.IsNamedtuple(instance, strict) 101 102 103# See the swig file (util.i) for documentation. 104_is_mapping = _pywrap_tensorflow.IsMapping 105_is_attrs = _pywrap_tensorflow.IsAttrs 106_is_composite_tensor = _pywrap_tensorflow.IsCompositeTensor 107 108 109def _sequence_like(instance, args): 110 """Converts the sequence `args` to the same type as `instance`. 111 112 Args: 113 instance: an instance of `tuple`, `list`, `namedtuple`, `dict`, or 114 `collections.OrderedDict`. 115 args: elements to be converted to the `instance` type. 116 117 Returns: 118 `args` with the type of `instance`. 119 """ 120 if _is_mapping(instance): 121 # Pack dictionaries in a deterministic order by sorting the keys. 122 # Notice this means that we ignore the original order of `OrderedDict` 123 # instances. This is intentional, to avoid potential bugs caused by mixing 124 # ordered and plain dicts (e.g., flattening a dict but using a 125 # corresponding `OrderedDict` to pack it back). 126 result = dict(zip(_sorted(instance), args)) 127 return type(instance)((key, result[key]) for key in instance) 128 elif _is_namedtuple(instance) or _is_attrs(instance): 129 return type(instance)(*args) 130 elif _is_composite_tensor(instance): 131 return instance._from_components(args) # pylint: disable=protected-access 132 else: 133 # Not a namedtuple 134 return type(instance)(args) 135 136 137def _yield_value(iterable): 138 for _, v in _yield_sorted_items(iterable): 139 yield v 140 141 142def _yield_sorted_items(iterable): 143 """Yield (key, value) pairs for `iterable` in a deterministic order. 144 145 For Sequences, the key will be an int, the array index of a value. 146 For Mappings, the key will be the dictionary key. 147 For objects (e.g. namedtuples), the key will be the attribute name. 148 149 In all cases, the keys will be iterated in sorted order. 150 151 Args: 152 iterable: an iterable. 153 154 Yields: 155 The iterable's (key, value) pairs, in order of sorted keys. 156 """ 157 if isinstance(iterable, _collections.Mapping): 158 # Iterate through dictionaries in a deterministic order by sorting the 159 # keys. Notice this means that we ignore the original order of `OrderedDict` 160 # instances. This is intentional, to avoid potential bugs caused by mixing 161 # ordered and plain dicts (e.g., flattening a dict but using a 162 # corresponding `OrderedDict` to pack it back). 163 for key in _sorted(iterable): 164 yield key, iterable[key] 165 elif _is_attrs(iterable): 166 for item in _get_attrs_items(iterable): 167 yield item 168 elif _is_namedtuple(iterable): 169 for field in iterable._fields: 170 yield field, getattr(iterable, field) 171 elif _is_composite_tensor(iterable): 172 for item in enumerate(iterable._to_components()): # pylint: disable=protected-access 173 yield item 174 else: 175 for item in enumerate(iterable): 176 yield item 177 178 179# See the swig file (util.i) for documentation. 180is_sequence = _pywrap_tensorflow.IsSequence 181 182 183# See the swig file (util.i) for documentation. 184is_sequence_or_composite = _pywrap_tensorflow.IsSequenceOrComposite 185 186 187@tf_export("nest.is_nested") 188def is_nested(seq): 189 """Returns true if its input is a collections.Sequence (except strings). 190 191 Args: 192 seq: an input sequence. 193 194 Returns: 195 True if the sequence is a not a string and is a collections.Sequence or a 196 dict. 197 """ 198 return is_sequence(seq) 199 200 201@tf_export("nest.flatten") 202def flatten(structure, expand_composites=False): 203 """Returns a flat list from a given nested structure. 204 205 If nest is not a sequence, tuple, or dict, then returns a single-element list: 206 [nest]. 207 208 In the case of dict instances, the sequence consists of the values, sorted by 209 key to ensure deterministic behavior. This is true also for OrderedDict 210 instances: their sequence order is ignored, the sorting order of keys is used 211 instead. The same convention is followed in pack_sequence_as. This correctly 212 repacks dicts and OrderedDicts after they have been flattened, and also allows 213 flattening an OrderedDict and then repacking it back using a corresponding 214 plain dict, or vice-versa. Dictionaries with non-sortable keys cannot be 215 flattened. 216 217 Users must not modify any collections used in nest while this function is 218 running. 219 220 Args: 221 structure: an arbitrarily nested structure or a scalar object. Note, numpy 222 arrays are considered scalars. 223 expand_composites: If true, then composite tensors such as tf.SparseTensor 224 and tf.RaggedTensor are expanded into their component tensors. 225 226 Returns: 227 A Python list, the flattened version of the input. 228 229 Raises: 230 TypeError: The nest is or contains a dict with non-sortable keys. 231 """ 232 return _pywrap_tensorflow.Flatten(structure, expand_composites) 233 234 235# See the swig file (util.i) for documentation. 236_same_namedtuples = _pywrap_tensorflow.SameNamedtuples 237 238 239class _DotString(object): 240 241 def __str__(self): 242 return "." 243 244 def __repr__(self): 245 return "." 246 247 248_DOT = _DotString() 249 250 251@tf_export("nest.assert_same_structure") 252def assert_same_structure(nest1, nest2, check_types=True, 253 expand_composites=False): 254 """Asserts that two structures are nested in the same way. 255 256 Note that namedtuples with identical name and fields are always considered 257 to have the same shallow structure (even with `check_types=True`). 258 For instance, this code will print `True`: 259 260 ```python 261 def nt(a, b): 262 return collections.namedtuple('foo', 'a b')(a, b) 263 print(assert_same_structure(nt(0, 1), nt(2, 3))) 264 ``` 265 266 Args: 267 nest1: an arbitrarily nested structure. 268 nest2: an arbitrarily nested structure. 269 check_types: if `True` (default) types of sequences are checked as well, 270 including the keys of dictionaries. If set to `False`, for example a 271 list and a tuple of objects will look the same if they have the same 272 size. Note that namedtuples with identical name and fields are always 273 considered to have the same shallow structure. Two types will also be 274 considered the same if they are both list subtypes (which allows "list" 275 and "_ListWrapper" from trackable dependency tracking to compare 276 equal). 277 expand_composites: If true, then composite tensors such as `tf.SparseTensor` 278 and `tf.RaggedTensor` are expanded into their component tensors. 279 280 Raises: 281 ValueError: If the two structures do not have the same number of elements or 282 if the two structures are not nested in the same way. 283 TypeError: If the two structures differ in the type of sequence in any of 284 their substructures. Only possible if `check_types` is `True`. 285 """ 286 try: 287 _pywrap_tensorflow.AssertSameStructure(nest1, nest2, check_types, 288 expand_composites) 289 except (ValueError, TypeError) as e: 290 str1 = str(map_structure(lambda _: _DOT, nest1)) 291 str2 = str(map_structure(lambda _: _DOT, nest2)) 292 raise type(e)("%s\n" 293 "Entire first structure:\n%s\n" 294 "Entire second structure:\n%s" 295 % (str(e), str1, str2)) 296 297 298def flatten_dict_items(dictionary): 299 """Returns a dictionary with flattened keys and values. 300 301 This function flattens the keys and values of a dictionary, which can be 302 arbitrarily nested structures, and returns the flattened version of such 303 structures: 304 305 ```python 306 example_dictionary = {(4, 5, (6, 8)): ("a", "b", ("c", "d"))} 307 result = {4: "a", 5: "b", 6: "c", 8: "d"} 308 flatten_dict_items(example_dictionary) == result 309 ``` 310 311 The input dictionary must satisfy two properties: 312 313 1. Its keys and values should have the same exact nested structure. 314 2. The set of all flattened keys of the dictionary must not contain repeated 315 keys. 316 317 Args: 318 dictionary: the dictionary to zip 319 320 Returns: 321 The zipped dictionary. 322 323 Raises: 324 TypeError: If the input is not a dictionary. 325 ValueError: If any key and value do not have the same structure layout, or 326 if keys are not unique. 327 """ 328 if not isinstance(dictionary, (dict, _collections.Mapping)): 329 raise TypeError("input must be a dictionary") 330 flat_dictionary = {} 331 for i, v in _six.iteritems(dictionary): 332 if not is_sequence(i): 333 if i in flat_dictionary: 334 raise ValueError( 335 "Could not flatten dictionary: key %s is not unique." % i) 336 flat_dictionary[i] = v 337 else: 338 flat_i = flatten(i) 339 flat_v = flatten(v) 340 if len(flat_i) != len(flat_v): 341 raise ValueError( 342 "Could not flatten dictionary. Key had %d elements, but value had " 343 "%d elements. Key: %s, value: %s." 344 % (len(flat_i), len(flat_v), flat_i, flat_v)) 345 for new_i, new_v in zip(flat_i, flat_v): 346 if new_i in flat_dictionary: 347 raise ValueError( 348 "Could not flatten dictionary: key %s is not unique." 349 % (new_i)) 350 flat_dictionary[new_i] = new_v 351 return flat_dictionary 352 353 354def _packed_nest_with_indices(structure, flat, index, is_seq): 355 """Helper function for pack_sequence_as. 356 357 Args: 358 structure: Substructure (list / tuple / dict) to mimic. 359 flat: Flattened values to output substructure for. 360 index: Index at which to start reading from flat. 361 is_seq: Function used to test if a value should be treated as a sequence. 362 363 Returns: 364 The tuple (new_index, child), where: 365 * new_index - the updated index into `flat` having processed `structure`. 366 * packed - the subset of `flat` corresponding to `structure`, 367 having started at `index`, and packed into the same nested 368 format. 369 370 Raises: 371 ValueError: if `structure` contains more elements than `flat` 372 (assuming indexing starts from `index`). 373 """ 374 packed = [] 375 for s in _yield_value(structure): 376 if is_seq(s): 377 new_index, child = _packed_nest_with_indices(s, flat, index, is_seq) 378 packed.append(_sequence_like(s, child)) 379 index = new_index 380 else: 381 packed.append(flat[index]) 382 index += 1 383 return index, packed 384 385 386@tf_export("nest.pack_sequence_as") 387def pack_sequence_as(structure, flat_sequence, expand_composites=False): 388 """Returns a given flattened sequence packed into a given structure. 389 390 If `structure` is a scalar, `flat_sequence` must be a single-element list; 391 in this case the return value is `flat_sequence[0]`. 392 393 If `structure` is or contains a dict instance, the keys will be sorted to 394 pack the flat sequence in deterministic order. This is true also for 395 `OrderedDict` instances: their sequence order is ignored, the sorting order of 396 keys is used instead. The same convention is followed in `flatten`. 397 This correctly repacks dicts and `OrderedDict`s after they have been 398 flattened, and also allows flattening an `OrderedDict` and then repacking it 399 back using a corresponding plain dict, or vice-versa. 400 Dictionaries with non-sortable keys cannot be flattened. 401 402 Args: 403 structure: Nested structure, whose structure is given by nested lists, 404 tuples, and dicts. Note: numpy arrays and strings are considered 405 scalars. 406 flat_sequence: flat sequence to pack. 407 expand_composites: If true, then composite tensors such as `tf.SparseTensor` 408 and `tf.RaggedTensor` are expanded into their component tensors. 409 410 Returns: 411 packed: `flat_sequence` converted to have the same recursive structure as 412 `structure`. 413 414 Raises: 415 ValueError: If `flat_sequence` and `structure` have different 416 element counts. 417 TypeError: `structure` is or contains a dict with non-sortable keys. 418 """ 419 is_seq = is_sequence_or_composite if expand_composites else is_sequence 420 if not is_seq(flat_sequence): 421 raise TypeError("flat_sequence must be a sequence") 422 423 if not is_seq(structure): 424 if len(flat_sequence) != 1: 425 raise ValueError("Structure is a scalar but len(flat_sequence) == %d > 1" 426 % len(flat_sequence)) 427 return flat_sequence[0] 428 429 try: 430 final_index, packed = _packed_nest_with_indices(structure, flat_sequence, 431 0, is_seq) 432 if final_index < len(flat_sequence): 433 raise IndexError 434 except IndexError: 435 flat_structure = flatten(structure) 436 if len(flat_structure) != len(flat_sequence): 437 raise ValueError( 438 "Could not pack sequence. Structure had %d elements, but " 439 "flat_sequence had %d elements. Structure: %s, flat_sequence: %s." % 440 (len(flat_structure), len(flat_sequence), structure, flat_sequence)) 441 return _sequence_like(structure, packed) 442 443 444@tf_export("nest.map_structure") 445def map_structure(func, *structure, **kwargs): 446 """Applies `func` to each entry in `structure` and returns a new structure. 447 448 Applies `func(x[0], x[1], ...)` where x[i] is an entry in 449 `structure[i]`. All structures in `structure` must have the same arity, 450 and the return value will contain results with the same structure layout. 451 452 Args: 453 func: A callable that accepts as many arguments as there are structures. 454 *structure: scalar, or tuple or list of constructed scalars and/or other 455 tuples/lists, or scalars. Note: numpy arrays are considered as scalars. 456 **kwargs: Valid keyword args are: 457 458 * `check_types`: If set to `True` (default) the types of 459 iterables within the structures have to be same (e.g. 460 `map_structure(func, [1], (1,))` raises a `TypeError` 461 exception). To allow this set this argument to `False`. 462 Note that namedtuples with identical name and fields are always 463 considered to have the same shallow structure. 464 * `expand_composites`: If set to `True`, then composite tensors such 465 as `tf.SparseTensor` and `tf.RaggedTensor` are expanded into their 466 component tensors. If `False` (the default), then composite tensors 467 are not expanded. 468 469 Returns: 470 A new structure with the same arity as `structure`, whose values correspond 471 to `func(x[0], x[1], ...)` where `x[i]` is a value in the corresponding 472 location in `structure[i]`. If there are different sequence types and 473 `check_types` is `False` the sequence types of the first structure will be 474 used. 475 476 Raises: 477 TypeError: If `func` is not callable or if the structures do not match 478 each other by depth tree. 479 ValueError: If no structure is provided or if the structures do not match 480 each other by type. 481 ValueError: If wrong keyword arguments are provided. 482 """ 483 if not callable(func): 484 raise TypeError("func must be callable, got: %s" % func) 485 486 if not structure: 487 raise ValueError("Must provide at least one structure") 488 489 check_types = True 490 expand_composites = False 491 if kwargs: 492 check_types = kwargs.pop("check_types", check_types) 493 expand_composites = kwargs.pop("expand_composites", expand_composites) 494 if kwargs: 495 raise ValueError("Only valid keyword arguments are check_types " 496 "and expand_composites") 497 498 for other in structure[1:]: 499 assert_same_structure(structure[0], other, check_types=check_types, 500 expand_composites=expand_composites) 501 502 flat_structure = [flatten(s, expand_composites) for s in structure] 503 entries = zip(*flat_structure) 504 505 return pack_sequence_as( 506 structure[0], [func(*x) for x in entries], 507 expand_composites=expand_composites) 508 509 510def map_structure_with_paths(func, *structure, **kwargs): 511 """Applies `func` to each entry in `structure` and returns a new structure. 512 513 Applies `func(path, x[0], x[1], ..., **kwargs)` where x[i] is an entry in 514 `structure[i]` and `path` is the common path to x[i] in the structures. All 515 structures in `structure` must have the same arity, and the return value will 516 contain the results with the same structure layout. Special kwarg 517 `check_types` determines whether the types of iterables within the structure 518 must be the same-- see **kwargs definition below. 519 520 Args: 521 func: A callable with the signature func(path, *values, **kwargs) that is 522 evaluated on the leaves of the structure. 523 *structure: A variable number of compatible structures to process. 524 **kwargs: Optional kwargs to be passed through to func. Special kwarg 525 `check_types` is not passed to func, but instead determines whether the 526 types of iterables within the structures have to be same (e.g., 527 `map_structure(func, [1], (1,))` raises a `TypeError` exception). By 528 default, the types must match. To allow iteration over structures of 529 different types (but common arity), set this kwarg to `False`. 530 531 Returns: 532 A structure of the same form as the input structures whose leaves are the 533 result of evaluating func on corresponding leaves of the input structures. 534 535 Raises: 536 TypeError: If `func` is not callable or if the structures do not match 537 each other by depth tree. 538 TypeError: If `check_types` is not `False` and the two structures differ in 539 the type of sequence in any of their substructures. 540 ValueError: If no structures are provided. 541 """ 542 def wrapper_func(tuple_path, *inputs, **kwargs): 543 string_path = "/".join(str(s) for s in tuple_path) 544 return func(string_path, *inputs, **kwargs) 545 546 return map_structure_with_tuple_paths_up_to(structure[0], 547 wrapper_func, 548 *structure, 549 **kwargs) 550 551 552def map_structure_with_tuple_paths(func, *structure, **kwargs): 553 """Applies `func` to each entry in `structure` and returns a new structure. 554 555 Applies `func(tuple_path, x[0], x[1], ..., **kwargs)` where `x[i]` is an entry 556 in `structure[i]` and `tuple_path` is a tuple of indices and/or dictionary 557 keys (as returned by `nest.yield_flat_paths`), which uniquely specifies the 558 common path to x[i] in the structures. All structures in `structure` must have 559 the same arity, and the return value will contain the results in the same 560 structure. Special kwarg `check_types` determines whether the types of 561 iterables within the structure must be the same-- see **kwargs definition 562 below. 563 564 Args: 565 func: A callable with the signature `func(tuple_path, *values, **kwargs)` 566 that is evaluated on the leaves of the structure. 567 *structure: A variable number of compatible structures to process. 568 **kwargs: Optional kwargs to be passed through to func. Special kwarg 569 `check_types` is not passed to func, but instead determines whether the 570 types of iterables within the structures have to be same (e.g. 571 `map_structure(func, [1], (1,))` raises a `TypeError` exception). To allow 572 this set this argument to `False`. 573 574 Returns: 575 A structure of the same form as the input structures whose leaves are the 576 result of evaluating func on corresponding leaves of the input structures. 577 578 Raises: 579 TypeError: If `func` is not callable or if the structures do not match 580 each other by depth tree. 581 TypeError: If `check_types` is not `False` and the two structures differ in 582 the type of sequence in any of their substructures. 583 ValueError: If no structures are provided. 584 """ 585 return map_structure_with_tuple_paths_up_to(structure[0], 586 func, 587 *structure, 588 **kwargs) 589 590 591def _yield_flat_up_to(shallow_tree, input_tree, path=()): 592 """Yields (path, value) pairs of input_tree flattened up to shallow_tree. 593 594 Args: 595 shallow_tree: Nested structure. Traverse no further than its leaf nodes. 596 input_tree: Nested structure. Return the paths and values from this tree. 597 Must have the same upper structure as shallow_tree. 598 path: Tuple. Optional argument, only used when recursing. The path from the 599 root of the original shallow_tree, down to the root of the shallow_tree 600 arg of this recursive call. 601 602 Yields: 603 Pairs of (path, value), where path the tuple path of a leaf node in 604 shallow_tree, and value is the value of the corresponding node in 605 input_tree. 606 """ 607 if (isinstance(shallow_tree, _six.string_types) or 608 not any([isinstance(shallow_tree, _collections.Sequence), 609 isinstance(shallow_tree, _collections.Mapping), 610 _is_namedtuple(shallow_tree), 611 _is_attrs(shallow_tree)])): 612 yield (path, input_tree) 613 else: 614 input_tree = dict(_yield_sorted_items(input_tree)) 615 for shallow_key, shallow_subtree in _yield_sorted_items(shallow_tree): 616 subpath = path + (shallow_key,) 617 input_subtree = input_tree[shallow_key] 618 for leaf_path, leaf_value in _yield_flat_up_to(shallow_subtree, 619 input_subtree, 620 path=subpath): 621 yield (leaf_path, leaf_value) 622 623 624def assert_shallow_structure(shallow_tree, input_tree, check_types=True): 625 """Asserts that `shallow_tree` is a shallow structure of `input_tree`. 626 627 That is, this function tests if the `input_tree` structure can be created from 628 the `shallow_tree` structure by replacing its leaf nodes with deeper 629 tree structures. 630 631 Examples: 632 633 The following code will raise an exception: 634 ```python 635 shallow_tree = {"a": "A", "b": "B"} 636 input_tree = {"a": 1, "c": 2} 637 assert_shallow_structure(shallow_tree, input_tree) 638 ``` 639 640 The following code will not raise an exception: 641 ```python 642 shallow_tree = ["a", "b"] 643 input_tree = ["c", ["d", "e"], "f"] 644 assert_shallow_structure(shallow_tree, input_tree) 645 ``` 646 647 Args: 648 shallow_tree: an arbitrarily nested structure. 649 input_tree: an arbitrarily nested structure. 650 check_types: if `True` (default) the sequence types of `shallow_tree` and 651 `input_tree` have to be the same. Note that even with check_types==True, 652 this function will consider two different namedtuple classes with the same 653 name and _fields attribute to be the same class. 654 655 Raises: 656 TypeError: If `shallow_tree` is a sequence but `input_tree` is not. 657 TypeError: If the sequence types of `shallow_tree` are different from 658 `input_tree`. Only raised if `check_types` is `True`. 659 ValueError: If the sequence lengths of `shallow_tree` are different from 660 `input_tree`. 661 """ 662 if is_sequence(shallow_tree): 663 if not is_sequence(input_tree): 664 raise TypeError( 665 "If shallow structure is a sequence, input must also be a sequence. " 666 "Input has type: %s." % type(input_tree)) 667 668 if check_types and not isinstance(input_tree, type(shallow_tree)): 669 # Duck-typing means that nest should be fine with two different 670 # namedtuples with identical name and fields. 671 shallow_is_namedtuple = _is_namedtuple(shallow_tree, False) 672 input_is_namedtuple = _is_namedtuple(input_tree, False) 673 if shallow_is_namedtuple and input_is_namedtuple: 674 if not _same_namedtuples(shallow_tree, input_tree): 675 raise TypeError(_STRUCTURES_HAVE_MISMATCHING_TYPES.format( 676 input_type=type(input_tree), 677 shallow_type=type(shallow_tree))) 678 679 elif not (isinstance(shallow_tree, _collections.Mapping) 680 and isinstance(input_tree, _collections.Mapping)): 681 raise TypeError(_STRUCTURES_HAVE_MISMATCHING_TYPES.format( 682 input_type=type(input_tree), 683 shallow_type=type(shallow_tree))) 684 685 if len(input_tree) < len(shallow_tree): 686 raise ValueError(_INPUT_TREE_SMALLER_THAN_SHALLOW_TREE.format( 687 input_size=len(input_tree), 688 shallow_size=len(shallow_tree))) 689 690 if isinstance(shallow_tree, _collections.Mapping): 691 absent_keys = set(shallow_tree) - set(input_tree) 692 if absent_keys: 693 raise ValueError(_SHALLOW_TREE_HAS_INVALID_KEYS 694 .format(sorted(absent_keys))) 695 696 for shallow_branch, input_branch in zip(_yield_value(shallow_tree), 697 _yield_value(input_tree)): 698 assert_shallow_structure(shallow_branch, input_branch, 699 check_types=check_types) 700 701 702def flatten_up_to(shallow_tree, input_tree, check_types=True): 703 """Flattens `input_tree` up to `shallow_tree`. 704 705 Any further depth in structure in `input_tree` is retained as elements in the 706 partially flatten output. 707 708 If `shallow_tree` and `input_tree` are not sequences, this returns a 709 single-element list: `[input_tree]`. 710 711 Use Case: 712 713 Sometimes we may wish to partially flatten a nested sequence, retaining some 714 of the nested structure. We achieve this by specifying a shallow structure, 715 `shallow_tree`, we wish to flatten up to. 716 717 The input, `input_tree`, can be thought of as having the same structure layout 718 as `shallow_tree`, but with leaf nodes that are themselves tree structures. 719 720 Examples: 721 722 ```python 723 input_tree = [[[2, 2], [3, 3]], [[4, 9], [5, 5]]] 724 shallow_tree = [[True, True], [False, True]] 725 726 flattened_input_tree = flatten_up_to(shallow_tree, input_tree) 727 flattened_shallow_tree = flatten_up_to(shallow_tree, shallow_tree) 728 729 # Output is: 730 # [[2, 2], [3, 3], [4, 9], [5, 5]] 731 # [True, True, False, True] 732 ``` 733 734 ```python 735 input_tree = [[('a', 1), [('b', 2), [('c', 3), [('d', 4)]]]]] 736 shallow_tree = [['level_1', ['level_2', ['level_3', ['level_4']]]]] 737 738 input_tree_flattened_as_shallow_tree = flatten_up_to(shallow_tree, input_tree) 739 input_tree_flattened = flatten(input_tree) 740 741 # Output is: 742 # [('a', 1), ('b', 2), ('c', 3), ('d', 4)] 743 # ['a', 1, 'b', 2, 'c', 3, 'd', 4] 744 ``` 745 746 Non-Sequence Edge Cases: 747 748 ```python 749 flatten_up_to(0, 0) # Output: [0] 750 flatten_up_to(0, [0, 1, 2]) # Output: [[0, 1, 2]] 751 flatten_up_to([0, 1, 2], 0) # Output: TypeError 752 flatten_up_to([0, 1, 2], [0, 1, 2]) # Output: [0, 1, 2] 753 ``` 754 755 Args: 756 shallow_tree: a possibly pruned structure of input_tree. 757 input_tree: an arbitrarily nested structure or a scalar object. 758 Note, numpy arrays are considered scalars. 759 check_types: bool. If True, check that each node in shallow_tree has the 760 same type as the corresponding node in input_tree. 761 762 Returns: 763 A Python list, the partially flattened version of `input_tree` according to 764 the structure of `shallow_tree`. 765 766 Raises: 767 TypeError: If `shallow_tree` is a sequence but `input_tree` is not. 768 TypeError: If the sequence types of `shallow_tree` are different from 769 `input_tree`. 770 ValueError: If the sequence lengths of `shallow_tree` are different from 771 `input_tree`. 772 """ 773 assert_shallow_structure(shallow_tree, input_tree, check_types) 774 # Discard paths returned by _yield_flat_up_to. 775 return list(v for _, v in _yield_flat_up_to(shallow_tree, input_tree)) 776 777 778def flatten_with_tuple_paths_up_to(shallow_tree, input_tree, check_types=True): 779 """Flattens `input_tree` up to `shallow_tree`. 780 781 Any further depth in structure in `input_tree` is retained as elements in the 782 partially flattened output. 783 784 Returns a list of (path, value) pairs, where value a leaf node in the 785 flattened tree, and path is the tuple path of that leaf in input_tree. 786 787 If `shallow_tree` and `input_tree` are not sequences, this returns a 788 single-element list: `[((), input_tree)]`. 789 790 Use Case: 791 792 Sometimes we may wish to partially flatten a nested sequence, retaining some 793 of the nested structure. We achieve this by specifying a shallow structure, 794 `shallow_tree`, we wish to flatten up to. 795 796 The input, `input_tree`, can be thought of as having the same structure layout 797 as `shallow_tree`, but with leaf nodes that are themselves tree structures. 798 799 Examples: 800 801 ```python 802 input_tree = [[[2, 2], [3, 3]], [[4, 9], [5, 5]]] 803 shallow_tree = [[True, True], [False, True]] 804 805 flattened_input_tree = flatten_with_tuple_paths_up_to(shallow_tree, 806 input_tree) 807 flattened_shallow_tree = flatten_with_tuple_paths_up_to(shallow_tree, 808 shallow_tree) 809 810 # Output is: 811 # [((0, 0), [2, 2]), 812 # ((0, 1), [3, 3]), 813 # ((1, 0), [4, 9]), 814 # ((1, 1), [5, 5])] 815 # 816 # [((0, 0), True), 817 # ((0, 1), True), 818 # ((1, 0), False), 819 # ((1, 1), True)] 820 ``` 821 822 ```python 823 input_tree = [[('a', 1), [('b', 2), [('c', 3), [('d', 4)]]]]] 824 shallow_tree = [['level_1', ['level_2', ['level_3', ['level_4']]]]] 825 826 input_tree_flattened_as_shallow_tree = flatten_up_to(shallow_tree, input_tree) 827 input_tree_flattened = flatten(input_tree) 828 829 # Output is: 830 # [((0, 0), ('a', 1)), 831 # ((0, 1, 0), ('b', 2)), 832 # ((0, 1, 1, 0), ('c', 3)), 833 # ((0, 1, 1, 1), ('d', 4))] 834 # ['a', 1, 'b', 2, 'c', 3, 'd', 4] 835 ``` 836 837 Non-Sequence Edge Cases: 838 839 ```python 840 flatten_with_tuple_paths_up_to(0, 0) # Output: [(), 0] 841 842 flatten_with_tuple_paths_up_to(0, [0, 1, 2]) # Output: [(), [0, 1, 2]] 843 844 flatten_with_tuple_paths_up_to([0, 1, 2], 0) # Output: TypeError 845 846 flatten_with_tuple_paths_up_to([0, 1, 2], [0, 1, 2]) 847 # Output: [((0,) 0), ((1,), 1), ((2,), 2)] 848 ``` 849 850 Args: 851 shallow_tree: a possibly pruned structure of input_tree. 852 input_tree: an arbitrarily nested structure or a scalar object. 853 Note, numpy arrays are considered scalars. 854 check_types: bool. If True, check that each node in shallow_tree has the 855 same type as the corresponding node in input_tree. 856 857 Returns: 858 A Python list, the partially flattened version of `input_tree` according to 859 the structure of `shallow_tree`. 860 861 Raises: 862 TypeError: If `shallow_tree` is a sequence but `input_tree` is not. 863 TypeError: If the sequence types of `shallow_tree` are different from 864 `input_tree`. 865 ValueError: If the sequence lengths of `shallow_tree` are different from 866 `input_tree`. 867 """ 868 assert_shallow_structure(shallow_tree, input_tree, check_types=check_types) 869 return list(_yield_flat_up_to(shallow_tree, input_tree)) 870 871 872def map_structure_up_to(shallow_tree, func, *inputs, **kwargs): 873 """Applies a function or op to a number of partially flattened inputs. 874 875 The `inputs` are flattened up to `shallow_tree` before being mapped. 876 877 Use Case: 878 879 Sometimes we wish to apply a function to a partially flattened 880 sequence (for example when the function itself takes sequence inputs). We 881 achieve this by specifying a shallow structure, `shallow_tree` we wish to 882 flatten up to. 883 884 The `inputs`, can be thought of as having the same structure layout as 885 `shallow_tree`, but with leaf nodes that are themselves tree structures. 886 887 This function therefore will return something with the same base structure as 888 `shallow_tree`. 889 890 Examples: 891 892 ```python 893 shallow_tree = [None, None] 894 inp_val = [1, 2, 3] 895 out = map_structure_up_to(shallow_tree, lambda x: 2 * x, inp_val) 896 897 # Output is: [2, 4] 898 ``` 899 900 ```python 901 ab_tuple = collections.namedtuple("ab_tuple", "a, b") 902 op_tuple = collections.namedtuple("op_tuple", "add, mul") 903 inp_val = ab_tuple(a=2, b=3) 904 inp_ops = ab_tuple(a=op_tuple(add=1, mul=2), b=op_tuple(add=2, mul=3)) 905 out = map_structure_up_to(inp_val, lambda val, ops: (val + ops.add) * ops.mul, 906 inp_val, inp_ops) 907 908 # Output is: ab_tuple(a=6, b=15) 909 ``` 910 911 ```python 912 data_list = [[2, 4, 6, 8], [[1, 3, 5, 7, 9], [3, 5, 7]]] 913 name_list = ['evens', ['odds', 'primes']] 914 out = map_structure_up_to( 915 name_list, 916 lambda name, sec: "first_{}_{}".format(len(sec), name), 917 name_list, data_list) 918 919 # Output is: ['first_4_evens', ['first_5_odds', 'first_3_primes']] 920 ``` 921 922 Args: 923 shallow_tree: a shallow tree, common to all the inputs. 924 func: callable which will be applied to each input individually. 925 *inputs: arbitrarily nested combination of objects that are compatible with 926 shallow_tree. The function `func` is applied to corresponding 927 partially flattened elements of each input, so the function must support 928 arity of `len(inputs)`. 929 **kwargs: kwargs to feed to func(). Special kwarg 930 `check_types` is not passed to func, but instead determines whether the 931 types of iterables within the structures have to be same (e.g. 932 `map_structure(func, [1], (1,))` raises a `TypeError` exception). To allow 933 this set this argument to `False`. 934 935 Raises: 936 TypeError: If `shallow_tree` is a sequence but `input_tree` is not. 937 TypeError: If the sequence types of `shallow_tree` are different from 938 `input_tree`. 939 ValueError: If the sequence lengths of `shallow_tree` are different from 940 `input_tree`. 941 942 Returns: 943 result of repeatedly applying `func`, with the same structure layout as 944 `shallow_tree`. 945 """ 946 return map_structure_with_tuple_paths_up_to( 947 shallow_tree, 948 lambda _, *values: func(*values), # Discards the path arg. 949 *inputs, 950 **kwargs) 951 952 953def map_structure_with_tuple_paths_up_to(shallow_tree, func, *inputs, **kwargs): 954 """Applies a function or op to a number of partially flattened inputs. 955 956 Like map_structure_up_to(), except that the 'func' argument takes a path 957 tuple as its first argument, followed by the corresponding values from 958 *inputs. 959 960 Example: 961 962 lowercase = {'a': 'a', 'b': ('b0', 'b1')} 963 uppercase = {'a': 'A', 'b': ('B0', 'B1')} 964 965 def print_path_and_values(path, *values): 966 print("path: {}, values: {}".format(path, values)) 967 968 shallow_tree = {'a': None} 969 map_structure_with_tuple_paths_up_to(shallow_tree, 970 print_path_and_values, 971 lowercase, 972 uppercase) 973 >>> path: ('a',), values: ('a', 'A') 974 >>> path: ('b', 0), values: ('b0', 'B0') 975 >>> path: ('b', 1), values: ('b1', 'B1') 976 977 shallow_tree = {'b': None} 978 map_structure_with_tuple_paths_up_to(shallow_tree, 979 print_path_and_values, 980 lowercase, 981 uppercase, 982 check_types=False) 983 >>> path: ('b', 1), values: (('bo', 'b1'), ('B0', 'B1')) 984 985 shallow_tree = {'a': None, 'b': {1: None}} 986 map_structure_with_tuple_paths_up_to(shallow_tree, 987 print_path_and_values, 988 lowercase, 989 uppercase, 990 check_types=False) 991 >>> path: ('a',), values: ('a', 'A') 992 >>> path: ('b', 1), values: ('b1', B1') 993 994 Args: 995 shallow_tree: a shallow tree, common to all the inputs. 996 func: callable that takes args (path, inputs_0_value, ... , inputs_N_value), 997 where path is a tuple path to a leaf node in shallow_tree, and 998 inputs_i_value is the corresponding value from inputs[i]. 999 *inputs: nested structures that are all structurally compatible with 1000 shallow_tree. 1001 **kwargs: kwargs to feed to func(). Special kwarg 1002 `check_types` is not passed to func, but instead determines whether the 1003 types of iterables within the structures have to be same (e.g. 1004 `map_structure(func, [1], (1,))` raises a `TypeError` exception). To allow 1005 this set this argument to `False`. 1006 1007 Raises: 1008 TypeError: If `shallow_tree` is a sequence but one of `*inputs` is not. 1009 TypeError: If the sequence types of `shallow_tree` are different from 1010 `input_tree`. 1011 ValueError: If the sequence lengths of `shallow_tree` are different from 1012 `input_tree`. 1013 1014 Returns: 1015 Result of repeatedly applying `func`. Has the same structure layout as 1016 `shallow_tree`. 1017 """ 1018 if not inputs: 1019 raise ValueError("Cannot map over no sequences") 1020 1021 check_types = kwargs.pop("check_types", True) 1022 1023 for input_tree in inputs: 1024 assert_shallow_structure(shallow_tree, input_tree, check_types=check_types) 1025 1026 # Flatten each input separately, apply the function to corresponding elements, 1027 # then repack based on the structure of the first input. 1028 flat_value_lists = [flatten_up_to(shallow_tree, input_tree, check_types) 1029 for input_tree in inputs] 1030 flat_path_list = [path for path, _ 1031 in _yield_flat_up_to(shallow_tree, inputs[0])] 1032 results = [func(*args, **kwargs) for args in zip(flat_path_list, 1033 *flat_value_lists)] 1034 return pack_sequence_as(structure=shallow_tree, flat_sequence=results) 1035 1036 1037def get_traverse_shallow_structure(traverse_fn, structure): 1038 """Generates a shallow structure from a `traverse_fn` and `structure`. 1039 1040 `traverse_fn` must accept any possible subtree of `structure` and return 1041 a depth=1 structure containing `True` or `False` values, describing which 1042 of the top-level subtrees may be traversed. It may also 1043 return scalar `True` or `False` "traversal is OK / not OK for all subtrees." 1044 1045 Examples are available in the unit tests (nest_test.py). 1046 1047 Args: 1048 traverse_fn: Function taking a substructure and returning either a scalar 1049 `bool` (whether to traverse that substructure or not) or a depth=1 1050 shallow structure of the same type, describing which parts of the 1051 substructure to traverse. 1052 structure: The structure to traverse. 1053 1054 Returns: 1055 A shallow structure containing python bools, which can be passed to 1056 `map_structure_up_to` and `flatten_up_to`. 1057 1058 Raises: 1059 TypeError: if `traverse_fn` returns a sequence for a non-sequence input, 1060 or a structure with depth higher than 1 for a sequence input, 1061 or if any leaf values in the returned structure or scalar are not type 1062 `bool`. 1063 """ 1064 to_traverse = traverse_fn(structure) 1065 if not is_sequence(structure): 1066 if not isinstance(to_traverse, bool): 1067 raise TypeError("traverse_fn returned structure: %s for non-structure: %s" 1068 % (to_traverse, structure)) 1069 return to_traverse 1070 level_traverse = [] 1071 if isinstance(to_traverse, bool): 1072 if not to_traverse: 1073 # Do not traverse this substructure at all. Exit early. 1074 return False 1075 else: 1076 # Traverse the entire substructure. 1077 for branch in _yield_value(structure): 1078 level_traverse.append( 1079 get_traverse_shallow_structure(traverse_fn, branch)) 1080 elif not is_sequence(to_traverse): 1081 raise TypeError("traverse_fn returned a non-bool scalar: %s for input: %s" 1082 % (to_traverse, structure)) 1083 else: 1084 # Traverse some subset of this substructure. 1085 assert_shallow_structure(to_traverse, structure) 1086 for t, branch in zip(_yield_value(to_traverse), _yield_value(structure)): 1087 if not isinstance(t, bool): 1088 raise TypeError( 1089 "traverse_fn didn't return a depth=1 structure of bools. saw: %s " 1090 " for structure: %s" % (to_traverse, structure)) 1091 if t: 1092 level_traverse.append( 1093 get_traverse_shallow_structure(traverse_fn, branch)) 1094 else: 1095 level_traverse.append(False) 1096 return _sequence_like(structure, level_traverse) 1097 1098 1099def yield_flat_paths(nest): 1100 """Yields paths for some nested structure. 1101 1102 Paths are lists of objects which can be str-converted, which may include 1103 integers or other types which are used as indices in a dict. 1104 1105 The flat list will be in the corresponding order as if you called 1106 `snt.nest.flatten` on the structure. This is handy for naming Tensors such 1107 the TF scope structure matches the tuple structure. 1108 1109 E.g. if we have a tuple `value = Foo(a=3, b=Bar(c=23, d=42))` 1110 1111 ```shell 1112 >>> nest.flatten(value) 1113 [3, 23, 42] 1114 >>> list(nest.yield_flat_paths(value)) 1115 [('a',), ('b', 'c'), ('b', 'd')] 1116 ``` 1117 1118 ```shell 1119 >>> list(nest.yield_flat_paths({'a': [3]})) 1120 [('a', 0)] 1121 >>> list(nest.yield_flat_paths({'a': 3})) 1122 [('a',)] 1123 ``` 1124 1125 Args: 1126 nest: the value to produce a flattened paths list for. 1127 1128 Yields: 1129 Tuples containing index or key values which form the path to a specific 1130 leaf value in the nested structure. 1131 """ 1132 for k, _ in _yield_flat_up_to(nest, nest): 1133 yield k 1134 1135 1136def flatten_with_joined_string_paths(structure, separator="/"): 1137 """Returns a list of (string path, data element) tuples. 1138 1139 The order of tuples produced matches that of `nest.flatten`. This allows you 1140 to flatten a nested structure while keeping information about where in the 1141 structure each data element was located. See `nest.yield_flat_paths` 1142 for more information. 1143 1144 Args: 1145 structure: the nested structure to flatten. 1146 separator: string to separate levels of hierarchy in the results, defaults 1147 to '/'. 1148 1149 Returns: 1150 A list of (string, data element) tuples. 1151 """ 1152 flat_paths = yield_flat_paths(structure) 1153 def stringify_and_join(path_elements): 1154 return separator.join(str(path_element) for path_element in path_elements) 1155 flat_string_paths = [stringify_and_join(path) for path in flat_paths] 1156 return list(zip(flat_string_paths, flatten(structure))) 1157 1158 1159def flatten_with_tuple_paths(structure): 1160 """Returns a list of `(tuple_path, leaf_element)` tuples. 1161 1162 The order of pairs produced matches that of `nest.flatten`. This allows you 1163 to flatten a nested structure while keeping information about where in the 1164 structure each data element was located. See `nest.yield_flat_paths` 1165 for more information about tuple paths. 1166 1167 Args: 1168 structure: the nested structure to flatten. 1169 1170 Returns: 1171 A list of `(tuple_path, leaf_element)` tuples. Each `tuple_path` is a tuple 1172 of indices and/or dictionary keys that uniquely specify the path to 1173 `leaf_element` within `structure`. 1174 """ 1175 return list(zip(yield_flat_paths(structure), flatten(structure))) 1176 1177 1178_pywrap_tensorflow.RegisterType("Mapping", _collections.Mapping) 1179_pywrap_tensorflow.RegisterType("Sequence", _collections.Sequence) 1180