File size: 32,359 Bytes
18652d8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
import abc
import dataclasses
import functools
import os
from os import environ
from typing import Mapping, Optional, Sequence, List
from absl import logging
import clu
import gin
from pathlib import Path

import seqio
from seqio import utils
from seqio.feature_converters import _check_exact_match, _check_lengths

import tensorflow as tf
from tensorflow.python.ops import control_flow_ops
from tensorflow.python.ops.image_ops_impl import _ImageDimensions, _CheckAtLeast3DImage, _assert, _is_tensor

from tensorflow.python.framework import ops
from tensorflow.python.ops import array_ops
from transformers import PreTrainedTokenizerFast

from . import seqio_tokenizer as vocab
from .constants import *
from .utils import pop_metadata
from .util import is_url

DEFAULT_EXTRA_IDS = 0
OutputFeaturesType = Mapping[str, utils.Feature]


def build_tokenizer(
    tokenizer_type, has_extra_token=True,
    adds_space=False,
    olmo_bos_token_id=1, olmo_eos_token_id=2,
    tokenizer_dir="gs://mm-olmo/tokenizer",
    pad_tokenizer_to=None, cache={},
):
  cache_key = (tokenizer_type, has_extra_token, adds_space, olmo_bos_token_id,
               olmo_eos_token_id, pad_tokenizer_to)
  if cache_key in cache:
    return cache[cache_key]

  if tokenizer_type == 'llama':
    tok = vocab.SentencePieceVocabulary(
      os.path.join(tokenizer_dir, "llama_tokenizer.model"),
      extra_ids=DEFAULT_EXTRA_IDS,
      reverse_extra_ids=True,
      extra_tokens=EXTRA_TOKENS if has_extra_token else None,
    )
  elif tokenizer_type == 'yi':
      tok = vocab.SentencePieceVocabulary(
          os.path.join(tokenizer_dir, "yi_tokenizer.model"),
          extra_ids=DEFAULT_EXTRA_IDS,
          reverse_extra_ids=True,
          extra_tokens=EXTRA_TOKENS if has_extra_token else None,
      )
  elif tokenizer_type == 'mistral':
      tok = vocab.SentencePieceVocabulary(
      os.path.join(tokenizer_dir, "mistral_tokenizer.model"),
      extra_ids=DEFAULT_EXTRA_IDS,
      reverse_extra_ids=True,
      extra_tokens=EXTRA_TOKENS if has_extra_token else None,
    )

  elif tokenizer_type == "mistral0.3":
      tok = vocab.SentencePieceVocabulary(
          os.path.join(tokenizer_dir, "mistral0.3_tokenizer.model.v3"),
          extra_ids=DEFAULT_EXTRA_IDS,
          reverse_extra_ids=True,
          extra_tokens=EXTRA_TOKENS if has_extra_token else None,
      )
  elif tokenizer_type == 'gemma':
      tok = vocab.SentencePieceVocabulary(
      os.path.join(tokenizer_dir, "gemma_tokenizer.model"),
      extra_ids=DEFAULT_EXTRA_IDS,
      reverse_extra_ids=True,
      extra_tokens=EXTRA_TOKENS if has_extra_token else None,
    )
  elif tokenizer_type.startswith("hf-"):
      # FIXME When using the beaker image "sanghol/mm-olmo" for hosting endpoints,
      # we should set the cache_dir, otherwise FileNotFound errors will be raised
      cache_dir = None if tokenizer_dir is None or is_url(tokenizer_dir) else tokenizer_dir
      from transformers import AutoTokenizer

      extra_tokens = list(EXTRA_TOKENS)
      if pad_tokenizer_to is not None:
          tokenizer = AutoTokenizer.from_pretrained(tokenizer_type[3:], token=environ.get("HF_ACCESS_TOKEN"), cache_dir=cache_dir)
          n_extra_tokens = pad_tokenizer_to - len(tokenizer)
          # This handles a case where the LLM embedding matrix is larger than the vocab size
          # We need the extra tokens in `EXTRA_TOKENS` to be assigned id's higher than the embedding
          # matrix size, not the vocab size, since we will concat the embedding and matrix with
          # the special token embedding matrix, so we pad the vocab with additional special tokens
          if n_extra_tokens > 0:
              logging.info(f"Padding tokenizer with {n_extra_tokens} tokens")
              extra_tokens = [f"|<EXTRA_TOKENS_{i}>|" for i in range(n_extra_tokens)] + extra_tokens

      bos_token_id = None

      tokenizer = AutoTokenizer.from_pretrained(
          tokenizer_type[3:], additional_special_tokens=extra_tokens,
          token=environ.get("HF_ACCESS_TOKEN"),
          cache_dir=cache_dir,
      )
      if ("qwen2" in tokenizer_type.lower()) or ("olmo" in tokenizer_type.lower()):
          # These tokenizers do not have a BOS, and instead use EOS as a generic seperator token.
          # In this case we will use EOS as BOS
          assert tokenizer.bos_token_id is None
          bos_token_id = tokenizer.eos_token_id

      if pad_tokenizer_to is not None:
          for ix, tok in enumerate(EXTRA_TOKENS):
              ids = tokenizer.encode(tok, add_special_tokens=False)
              assert ids == [pad_tokenizer_to + ix]

      tok = vocab.HfTokenizerWrapper(tokenizer, bos_token_id=bos_token_id, adds_space=adds_space)
  elif tokenizer_type.startswith("olmo-"):
      from olmo.tokenizer import Tokenizer
      assert Path(tokenizer_type[5:]).is_file()
      tokenizer = Tokenizer.from_file(
          tokenizer_type[5:],
          eos_token_id=olmo_eos_token_id,
          pad_token_id=-1,
      )
      tok = vocab.OLMoTokenizerWrapper(tokenizer, bos_token_id=olmo_bos_token_id, adds_space=adds_space)
  else:
    raise NotImplementedError(tokenizer_type)
  cache[cache_key] = tok
  return tok


def get_special_token_ids(tokenizer):
  if isinstance(tokenizer, (vocab.HfTokenizerWrapper, vocab.OLMoTokenizerWrapper)):
      ids = tokenizer.encode("".join(EXTRA_TOKENS))
      if len(ids) == len(EXTRA_TOKENS) + 1:
          ids = ids[1:]
  elif ("gemma_tokenizer" in tokenizer._sentencepiece_model_file or
        "yi_tokenizer" in tokenizer._sentencepiece_model_file
  ):
      # Not sure why ATM, but the LLaMa tokenizer will add an extra space token
      # if this string starts with a space, while the gemma one needs the leading space
      ids = tokenizer.encode(" " + " ".join(EXTRA_TOKENS))
  else:
      ids = tokenizer.encode(" ".join(EXTRA_TOKENS))

  assert len(ids) == len(EXTRA_TOKENS)
  return {k: i for k, i in zip(EXTRA_TOKENS, ids)}


def _append_to_innermost_axis(
    tensor: tf.Tensor, scalar: tf.Tensor,
) -> tf.Tensor:
  """Appends `scalar` to each slice in the innermost axis of `tensor`.

  >>> _append_to_innermost_axis([1, 2, 3], -1)
  [1, 2, 3, -1]
  >>> _append_to_innermost_axis([[1, 2], [3, 4]], -1)
  [[1, 2, -1], [3, 4, -1]]
  >>> _append_to_innermost_axis(tf.ragged.constant([[1, 2], [3]]), -1)
  [[1, 2, -1], [3, -1]]

  Args:
    tensor: The tensor that should have a value appended.
    scalar: The value to append.

  Returns:
    A copy of `tensor` with `scalar` appended to each slice along
    the innermost axis.
  """
  if isinstance(tensor, tf.RaggedTensor):
    if tensor.shape.rank > 2:
      return tensor.with_values(
          _append_to_innermost_axis(tensor.values, scalar)
      )
    else:
      return tf.concat([tensor, tf.fill([tensor.nrows(), 1], scalar)], axis=1)
  else:
    ndims = tf.rank(tensor)
    paddings = tf.concat(
        [tf.zeros((ndims - 1, 2), dtype=tf.int32), tf.constant([[0, 1]])],
        axis=0,
    )
    return tf.pad(tensor, paddings=paddings, constant_values=scalar)


def _shift_right_by_one(tensor: tf.Tensor, bos_id: int = 0) -> tf.Tensor:
  """Shift the input tensor to the right by one position without wrapping."""

  if not (tensor.dtype.is_integer or tensor.dtype.is_floating):
    raise ValueError(f"Only numeric types are supported. Got: {tensor.dtype}")
  # tf.roll wraps around the axis.
  rolled = tf.roll(tensor, shift=1, axis=0)

  # Zero out the first position by multiplying with [0, 1, 1, ..., 1].
  depth = tf.shape(tensor)[0]
  mask = tf.one_hot(0, depth=depth, on_value=0, off_value=1, dtype=tensor.dtype)

  # Expand dims of mask to broadcast to rolled.
  dim_expansion = [slice(None, None)] + [None] * (len(rolled.shape) - 1)
  mask = mask[dim_expansion]
  return rolled * mask + (1 - mask) * bos_id


def make_autoregressive_inputs(
    targets: tf.Tensor,
    sequence_id: tf.Tensor = None,
    output_dtype: Optional[tf.dtypes.DType] = None,
    bos_id: int = 0,
) -> tf.Tensor:
  """Generate inputs for an autoregressive model, by shifting the targets.

  Modified from mesh_tensorflow.transformer.transformer.autoregressive_inputs.

  For the first element of each sequence, the returned input id is 0.

  For a "packed" dataset, also pass the sequence_id tensor, which aligns
  with the targets tensor and contains different values for different
  concatenated examples.

  Example for a packed dataset:

  ```
        targets = [3, 8, 2, 9, 2, 5, 4, 2, -1, -1]
    sequence_id = [1, 1, 1, 2, 2, 3, 3, 3, 0, 0]
         inputs = [1, 3, 8, 1, 9, 1, 5, 4, -1, -1]
                            |     |        |
                            These positions are set to 0 if sequence_id is not
                            None.
  ```

  Args:
    targets: a tf.int32 tensor with shape [length].
    sequence_id: an optional tensor with the same shape as targets.
    output_dtype: an optional output data type.
    bos_id: bos id.

  Returns:
    a tensor with dtype tf.int32 and the same shape as targets.
  """
  output_dtype = output_dtype or targets.dtype
  if sequence_id is not None and not sequence_id.dtype.is_integer:
    raise ValueError(
        "The sequence_id should be integer-valued tensors for a packed dataset."
    )
  if sequence_id is not None and len(targets.shape) > 1:
    raise ValueError(
        "Only 1-D sequences are supported with packing. Got a "
        f"packed {len(targets.shape)}-D sequence."
    )

  inputs = _shift_right_by_one(targets, bos_id)
  if inputs.dtype != output_dtype:
    inputs = tf.cast(inputs, output_dtype)

  # We should have a 0 at the beginning of each sequence rather than the
  # shifted EOS (e.g. 1) from the previous sequence.
  if sequence_id is not None:
    not_first_in_sequence = tf.equal(
        sequence_id, _shift_right_by_one(sequence_id)
    )
    not_first_in_sequence = tf.cast(not_first_in_sequence, output_dtype)
    first_ids = tf.cast((1 - not_first_in_sequence) * bos_id, output_dtype)
    inputs = inputs * not_first_in_sequence + first_ids
  return inputs


@tf.function
def sum_except_first_axis(tensor):
    # Compute the sum along all axes except the first
    axes_to_sum = tuple(range(1, len(tensor.shape)))
    return tf.reduce_sum(tensor, axis=axes_to_sum)


@seqio.map_over_dataset()
def add_segment_ids(ex):
    ex["subsegment_ids"] = tf.zeros_like(ex["target_tokens"], dtype=tf.int32)
    return ex


def trim_and_pad_dataset(
    dataset: tf.data.Dataset, feature_lengths: Mapping[str, int]
) -> tf.data.Dataset:
  """Trim and pad first dimension of features to `feature_lengths`.

  Args:
    dataset: tf.data.Dataset, the dataset to trim/pad examples in.
    feature_lengths: map from feature key to final length. Other features will
      be returned unchanged.

  Returns:
    Trimmed/padded tf.data.Dataset.
  """

  def _trim_and_pad(k: str, t: tf.Tensor) -> tf.Tensor:
    """Trim/pad to the first axis of `t` to be of size `length`."""
    if k not in feature_lengths:
      return t
    if isinstance(t, tf.RaggedTensor):
      t = t.to_tensor()

    constant_values = -1
    length_k = feature_lengths[k]
    if isinstance(length_k, int):
      t = t[:length_k]
      pad_amt = length_k - tf.shape(t)[0]
      padded_t = tf.pad(t, [(0, pad_amt)] + [(0, 0)] * (len(t.shape) - 1), constant_values=constant_values)
      padded_t.set_shape([length_k] + t.shape.as_list()[1:])
      return padded_t

    slices = tuple((slice(0, limit) for limit in length_k))
    t = t[slices]
    pad_amt = tf.pad((length_k - tf.shape(t))[..., None], ((0, 0), (1, 0)), constant_values=constant_values)
    padded_t = tf.pad(t, pad_amt, constant_values=constant_values)
    padded_t.set_shape(length_k)
    return padded_t

  return dataset.map(
      lambda x: {k: _trim_and_pad(k, t) for k, t in x.items()},
      num_parallel_calls=tf.data.experimental.AUTOTUNE,
  )


def get_3d_subsegments(segmented_suffix):
    q_lens, text_lens = segmented_suffix.nested_row_lengths()
    text_segments = tf.range(0, tf.shape(text_lens)[0], dtype=tf.int32)
    question_repeat = tf.reshape(tf.stack([tf.ones_like(q_lens), q_lens-1], 1), [-1])
    question_offset = tf.range(1, tf.shape(q_lens)[0]+1, dtype=tf.int32)*200
    question_offset = tf.reshape(tf.stack([question_offset, question_offset-100], 1), [-1])
    text_segments = text_segments + tf.repeat(question_offset, question_repeat)
    segment_ids = tf.cast(tf.repeat(text_segments, text_lens), tf.int32)
    return segment_ids


def assert_not_truncated(ds, keys, max_val):
    def _check(ex):
        for k in keys:
            tf.assert_less(tf.shape(ex[k])[0], max_val+1,
                           message=f"Field {k} was unexpectedly truncated max_len={max_val}")
        return ex
    return ds.map(_check)


def apply_with_random_selector(x, func, num_cases):
    """Computes func(x, sel), with sel sampled from [0...num_cases-1].
    Args:
      x: input Tensor.
      func: Python function to apply.
      num_cases: Python int32, number of cases to sample sel from.
    Returns:
      The result of func(x, sel), where func receives the value of the
      selector as a python integer, but sel is sampled dynamically.
    """
    sel = tf.random.uniform([], maxval=num_cases, dtype=tf.int32)
    # Pass the real x only to one of the func calls.
    return control_flow_ops.merge([
        func(control_flow_ops.switch(x, tf.equal(sel, case))[1], case)
        for case in range(num_cases)])[0]


def denormalize_boxes(boxes, image_shape):
    """Converts boxes normalized by [height, width] to pixel coordinates.
    Args:
      boxes: a tensor whose last dimension is 4 representing the coordinates of
        boxes in ymin, xmin, ymax, xmax order.
      image_shape: a list of two integers, a two-element vector or a tensor such
        that all but the last dimensions are `broadcastable` to `boxes`. The last
        dimension is 2, which represents [height, width].
    Returns:
      denormalized_boxes: a tensor whose shape is the same as `boxes` representing
        the denormalized boxes.
    Raises:
      ValueError: If the last dimension of boxes is not 4.
    """
    with tf.name_scope('denormalize_boxes'):
      if isinstance(image_shape, list) or isinstance(image_shape, tuple):
        height, width = image_shape
        height = tf.cast(height, dtype=boxes.dtype)
        width = tf.cast(width, dtype=boxes.dtype)
      else:
        image_shape = tf.cast(image_shape, dtype=boxes.dtype)
        height, width = tf.split(image_shape, 2, axis=-1)

      ymin, xmin, ymax, xmax = tf.split(boxes, 4, axis=-1)
      ymin = ymin * height
      xmin = xmin * width
      ymax = ymax * height
      xmax = xmax * width

      denormalized_boxes = tf.concat([ymin, xmin, ymax, xmax], axis=-1)
      return denormalized_boxes

def pad_to_bounding_box(image, offset_height, offset_width, target_height,
                        target_width, value=0):

  return pad_to_bounding_box_internal(
      image,
      offset_height,
      offset_width,
      target_height,
      target_width,
      check_dims=True, 
      value=value)

def pad_to_bounding_box_internal(image, offset_height, offset_width,
                                 target_height, target_width, check_dims, value):

  with ops.name_scope(None, 'pad_to_bounding_box_with_one_internal', [image]):
    image = ops.convert_to_tensor(image, name='image')

    is_batch = True
    image_shape = image.get_shape()
    if image_shape.ndims == 3:
      is_batch = False
      image = array_ops.expand_dims(image, 0)
    elif image_shape.ndims is None:
      is_batch = False
      image = array_ops.expand_dims(image, 0)
      image.set_shape([None] * 4)
    elif image_shape.ndims != 4:
      raise ValueError(
          '\'image\' (shape %s) must have either 3 or 4 dimensions.' %
          image_shape)

    batch, height, width, depth = _ImageDimensions(image, rank=4)

    after_padding_width = target_width - offset_width - width

    after_padding_height = target_height - offset_height - height

    if check_dims:
      assert_ops = _CheckAtLeast3DImage(image, require_static=False)
      assert_ops += _assert(offset_height >= 0, ValueError,
                            'offset_height must be >= 0')
      assert_ops += _assert(offset_width >= 0, ValueError,
                            'offset_width must be >= 0')
      assert_ops += _assert(after_padding_width >= 0, ValueError,
                            'width must be <= target - offset')
      assert_ops += _assert(after_padding_height >= 0, ValueError,
                            'height must be <= target - offset')
      image = control_flow_ops.with_dependencies(assert_ops, image)

    # Do not pad on the depth dimensions.
    paddings = array_ops.reshape(
        tf.stack([
            0, 0, offset_height, after_padding_height, offset_width,
            after_padding_width, 0, 0
        ]), [4, 2])
    padded = array_ops.pad(image, paddings, constant_values=value)

    padded_shape = [
        None if _is_tensor(i) else i
        for i in [batch, target_height, target_width, depth]
    ]
    padded.set_shape(padded_shape)

    if not is_batch:
      padded = array_ops.squeeze(padded, axis=[0])

    return padded

def resize_and_crop_boxes(boxes, image_scale, output_size, offset, paddings):
    """Resizes boxes to output size with scale and offset.
    Args:
      boxes: `Tensor` of shape [N, 4] representing ground truth boxes.
      image_scale: 2D float `Tensor` representing scale factors that apply to
        [height, width] of input image.
      output_size: 2D `Tensor` or `int` representing [height, width] of target
        output image size.
      offset: 2D `Tensor` representing top-left corner [y0, x0] to crop scaled
        boxes.
      paddings: 2D `Tensor` representing top/left paddings.
    Returns:
      boxes: `Tensor` of shape [N, 4] representing the scaled boxes.
    """
    # Adjusts box coordinates based on image_scale, offset and paddings.
    boxes *= tf.tile(tf.expand_dims(image_scale, axis=0), [1, 2])
    boxes -= tf.tile(tf.expand_dims(offset, axis=0), [1, 2])
    boxes += tf.tile(tf.expand_dims(paddings, axis=0), [1, 2])
    # Clips the boxes.
    boxes = clip_boxes(boxes, output_size)
    return boxes

def clip_boxes(boxes, image_shape):
  """Clips boxes to image boundaries.
  Args:
    boxes: a tensor whose last dimension is 4 representing the coordinates of
      boxes in ymin, xmin, ymax, xmax order.
    image_shape: a list of two integers, a two-element vector or a tensor such
      that all but the last dimensions are `broadcastable` to `boxes`. The last
      dimension is 2, which represents [height, width].
  Returns:
    clipped_boxes: a tensor whose shape is the same as `boxes` representing the
      clipped boxes.
  Raises:
    ValueError: If the last dimension of boxes is not 4.
  """
  if boxes.shape[-1] != 4:
    raise ValueError('boxes.shape[-1] is {:d}, but must be 4.'.format(
        boxes.shape[-1]))

  with tf.name_scope('clip_boxes'):
    if isinstance(image_shape, list) or isinstance(image_shape, tuple):
      height, width = image_shape
      max_length = [height, width, height, width]
    else:
      image_shape = tf.cast(image_shape, dtype=boxes.dtype)
      height, width = tf.unstack(image_shape, axis=-1)
      max_length = tf.stack(
          [height, width, height, width], axis=-1)

    clipped_boxes = tf.math.maximum(tf.math.minimum(boxes, max_length), 0.0)
    return clipped_boxes
  
  
def get_non_empty_box_indices(boxes):
    """Get indices for non-empty boxes."""
    # Selects indices if box height or width is 0.
    height = boxes[:, 2] - boxes[:, 0]
    width = boxes[:, 3] - boxes[:, 1]
    indices = tf.where(
        tf.logical_and(tf.greater(height, 0), tf.greater(width, 0)))
    return indices[:, 0]


def resize_and_pad(image, desired_output_size, masks=None, boxes=None, labels=None,
                   random_scale_min=0.1, random_scale_max=2.0, do_random_scale=False,
                   shrink_both_sides=True, boxes1=None, filter_box=True,
                   desired_target_size=None, random_scale_ratio=0.0,
                   resize_method=tf.image.ResizeMethod.BILINEAR, return_outputs=True,
                   pad_value=0, normalize=True):
    desired_height, desired_width = desired_output_size
    desired_height_f = tf.cast(desired_height, dtype=tf.float32)
    desired_width_f = tf.cast(desired_width, dtype=tf.float32)

    height = tf.cast(tf.shape(image)[0], tf.float32)
    width = tf.cast(tf.shape(image)[1], tf.float32)

    if boxes is not None:
        # Converts boxes from normalized coordinates to pixel coordinates.
        # Now the coordinates of boxes are w.r.t. the original image.
        boxes = denormalize_boxes(boxes, [height, width])

    if boxes1 is not None:
        boxes1 = denormalize_boxes(boxes1, [height, width])

    if do_random_scale:
        random_scale_factor = tf.random.uniform([], random_scale_min, random_scale_max)
        if not shrink_both_sides:
            # Max random is where scale * W > W_desired
            #                     scale * H > H_desired
            rsf_max = tf.maximum(desired_width_f / width, desired_height_f / height)
            random_scale_factor = tf.minimum(rsf_max, random_scale_factor)

        scaled_y = tf.cast(random_scale_factor * desired_height_f, tf.int32)
        scaled_x = tf.cast(random_scale_factor * desired_width_f, tf.int32)

        # Recompute the accurate scale_factor using rounded scaled image size.
        image_scale_y = tf.cast(scaled_y, tf.float32) / height
        image_scale_x = tf.cast(scaled_x, tf.float32) / width

        image_scale = tf.cond(tf.less(
            tf.random.uniform([], minval=0, maxval=1, dtype=tf.float32),
            tf.cast(random_scale_ratio, tf.float32)),
            lambda: tf.maximum(image_scale_x, image_scale_y),
            lambda: tf.minimum(image_scale_x, image_scale_y))

        # image_scale = tf.minimum(image_scale_x, image_scale_y)

        # Conceptual captions has some REALLY WIDE images I believe
        # this ensures that we won't scale any side lower than to 64
        image_scale = tf.maximum(image_scale, 64.0 / tf.minimum(height, width))

        # Select non-zero random offset (x, y) if scaled image is larger than
        # self._output_size.
        scaled_height = tf.cast(height * image_scale, tf.int32)
        scaled_width = tf.cast(width * image_scale, tf.int32)
        offset_y = tf.cast(scaled_height - desired_height, tf.float32)
        offset_x = tf.cast(scaled_width - desired_width, tf.float32)
        offset_y = tf.maximum(0.0, offset_y) * tf.random.uniform([], 0, 1)
        offset_x = tf.maximum(0.0, offset_x) * tf.random.uniform([], 0, 1)
        offset_y = tf.cast(offset_y, tf.int32)
        offset_x = tf.cast(offset_x, tf.int32)
    else:
        image_scale_y = desired_height_f / height
        image_scale_x = desired_width_f / width
        image_scale = tf.minimum(image_scale_x, image_scale_y)
        scaled_height = tf.cast(height * image_scale, tf.int32)
        scaled_width = tf.cast(width * image_scale, tf.int32)
        offset_y = tf.constant(0)
        offset_x = tf.constant(0)

    # Now resize and crop
    if resize_method == 'random' and do_random_scale:
        resize_methods = sorted([k for k in tf.image.ResizeMethod.__dict__.keys() if k.isupper()])
        image = apply_with_random_selector(
            image,
            lambda x, method_idx: tf.image.resize(x, [scaled_height, scaled_width],
                                                  tf.image.ResizeMethod.__dict__[resize_methods[method_idx]],
                                                  antialias=True),
            num_cases=len(resize_methods))

    elif resize_method != 'random':
        image = tf.image.resize(image, [scaled_height, scaled_width], method=resize_method, antialias=True)
    else:
        image = tf.image.resize(image, [scaled_height, scaled_width],
                                method=tf.image.ResizeMethod.BILINEAR, antialias=True)

    image = tf.clip_by_value(image, 0.0, 1.0)

    # H x W x C
    image = image[offset_y:offset_y + desired_height, offset_x:offset_x + desired_width, :]

    H = tf.shape(image)[0]
    W = tf.shape(image)[1]

    top_pad = (desired_height - H) // 2
    left_pad = (desired_width - W) // 2

    image_mask = pad_to_bounding_box(
        tf.ones_like(image, dtype=tf.bool), top_pad, left_pad, desired_height, desired_width)[:,:,0]

    image = pad_to_bounding_box(image, top_pad, left_pad, desired_height, desired_width, value=pad_value)

    if isinstance(desired_height, int) and isinstance(desired_width, int):
        image.set_shape([desired_height, desired_width, 3])

    if masks is not None and tf.size(masks) != 0:
        masks = tf.image.resize(masks, [scaled_height, scaled_width],
                                method=tf.image.ResizeMethod.NEAREST_NEIGHBOR)

        if len(masks.shape) == 3:
            masks = masks[offset_y:offset_y + desired_height, offset_x:offset_x + desired_width]
        else:
            masks = masks[:, offset_y:offset_y + desired_height, offset_x:offset_x + desired_width]

        masks = pad_to_bounding_box(masks, top_pad, left_pad, desired_height, desired_width)
        masks = tf.image.resize(masks, desired_target_size,
                                method=tf.image.ResizeMethod.NEAREST_NEIGHBOR)

    indices = None
    if boxes is not None:
        # assert ValueError("the box need to be shift which is not tested yet.")
        boxes = resize_and_crop_boxes(
            boxes,
            tf.stack([image_scale, image_scale]),
            [desired_height, desired_width],
            tf.cast(tf.stack([offset_y, offset_x]), dtype=tf.float32),
            tf.cast(tf.stack([top_pad, left_pad]), dtype=tf.float32))

        if filter_box:
            indices = get_non_empty_box_indices(boxes)
        else:
            indices = tf.range(tf.shape(boxes)[0])
        boxes = tf.gather(boxes, indices)

        if labels is not None:
            labels = tf.gather(labels, indices)

        if boxes1 is not None:
            boxes1 = resize_and_crop_boxes(
                boxes1,
                tf.stack([image_scale, image_scale]),
                [desired_height, desired_width],
                tf.cast(tf.stack([offset_y, offset_x]), dtype=tf.float32),
                tf.cast(tf.stack([top_pad, left_pad]), dtype=tf.float32))

    image_info = tf.stack([
        tf.cast(top_pad, tf.float32),
        tf.cast(left_pad, tf.float32),
        1.0 / image_scale,
        height,
        width,
        tf.cast(offset_y, dtype=tf.float32) / height,
        tf.cast(offset_x, dtype=tf.float32) / width,
        tf.cast(offset_y, dtype=tf.float32),
        tf.cast(offset_x, dtype=tf.float32),
        tf.cast(scaled_height, dtype=tf.float32),
        tf.cast(scaled_width, dtype=tf.float32),
        ])

    if boxes1 is not None:
        outputs = (image_info, masks, boxes, labels, indices, boxes1)
    else:
        outputs = (image_info, masks, boxes, labels, indices)

    if normalize:
        image = normalize_image(image)

    if return_outputs:
        return image, image_mask, outputs
    else:
        return image, image_mask


def _remove_bars_from_frames(frames, black_bar=True, threshold=32, max_perc_to_trim=0.3):
    """
    :param frames: [num_frames, height, width, 3]
    :param blackbar_threshold: Pixels must be this intense for us to not trim
    :param max_perc_to_prim: Will trim x% by default of the image at most in each dimension
    :return:
    """
    # Detect black bars####################
    frames_shape = tf.shape(frames)
    h, w = frames_shape[1], frames_shape[2]
    if black_bar:
      has_content = tf.reduce_max(frames, axis=(0, -1)) >= threshold
    else:
      has_content = tf.reduce_min(frames, axis=(0, -1)) <= threshold

    y_frames = tf.cast(tf.reshape(tf.where(tf.reduce_any(has_content, axis=1)), [-1]), tf.int32)
    nhbars = tf.shape(y_frames)[0]
    y_frames = tf.cond(nhbars > 0, lambda: y_frames, lambda: tf.expand_dims(tf.cast(h // 2, tf.int32), axis=0))

    y1 = tf.minimum(y_frames[0], tf.cast(tf.cast(h, tf.float32) * max_perc_to_trim, tf.int32))
    y2 = tf.maximum(y_frames[-1] + 1, tf.cast(tf.cast(h, tf.float32) * (1 - max_perc_to_trim), tf.int32))

    x_frames = tf.cast(tf.reshape(tf.where(tf.reduce_any(has_content, axis=0)), [-1]), tf.int32)
    nvbars = tf.shape(x_frames)[0]
    x_frames = tf.cond(nvbars > 0, lambda: x_frames, lambda: tf.expand_dims(tf.cast(w // 2, tf.int32), axis=0))

    x1 = tf.minimum(x_frames[0], tf.cast(tf.cast(w, tf.float32) * max_perc_to_trim, tf.int32))
    x2 = tf.maximum(x_frames[-1] + 1, tf.cast(tf.cast(w, tf.float32) * (1 - max_perc_to_trim), tf.int32))

    frames = frames[:, y1:y2, x1:x2]
    return frames

def convert_video_dtype(video,dtype):
    """
    Converts tensor to dtype and scales the values. 
    Video equivalent of tf.convert_image_dtype: https://www.tensorflow.org/api_docs/python/tf/image/convert_image_dtype
    """
    return tf.map_fn(
        fn=functools.partial(
            tf.image.convert_image_dtype,
            dtype=dtype),
        elems=video,
        fn_output_signature=dtype)


def stateless_shuffle(x: tf.Tensor, seed):
  if hasattr(tf.random.experimental, 'stateless_shuffle'):
    return tf.random.experimental.stateless_shuffle(x, seed=seed)
  else:
    vals = tf.random.stateless_uniform(tf.shape(x)[:1], seed)
    ixs = tf.argsort(vals)
    return tf.gather(x, ixs)


def stateless_permutation(n: int, seed):
    if hasattr(tf.random.experimental, 'stateless_shuffle'):
        ix = tf.range(0, n, dtype=tf.int32)
        return tf.random.experimental.stateless_shuffle(ix, seed=seed)
    else:
        vals = tf.random.stateless_uniform(n, seed)
        return tf.argsort(vals)


@seqio.map_over_dataset
def _strip_metadata(example):
    return pop_metadata(example)[0]


def sample_patches(mask, n_patches, stateless=False, seeds=None):
  input_sample_valid = tf.boolean_mask(tf.range(tf.shape(mask)[0]), mask)
  input_sample_masked = tf.boolean_mask(tf.range(tf.shape(mask)[0]), mask == 0)
  if stateless:
    encoder_pos_ids = tf.concat([
      stateless_shuffle(input_sample_valid, seeds[0]),
      stateless_shuffle(input_sample_masked, seeds[1])], axis=0)[:n_patches]
  else:
    encoder_pos_ids = tf.concat([
      tf.random.shuffle(input_sample_valid),
      tf.random.shuffle(input_sample_masked)], axis=0)[:n_patches]
  encoder_pos_ids = tf.reshape(encoder_pos_ids, (n_patches,))
  encoder_pos_ids = tf.cast(encoder_pos_ids, tf.int32)
  return encoder_pos_ids


@gin.configurable()
def normalize_image(image,
                    offset=(0.48145466, 0.4578275, 0.40821073),
                    scale=(0.26862954, 0.26130258, 0.27577711)):
  """Normalizes the image to zero mean and unit variance."""
  offset = tf.constant(offset)
  offset = tf.expand_dims(offset, axis=0)
  offset = tf.expand_dims(offset, axis=0)
  image -= tf.cast(offset, image.dtype)

  scale = tf.constant(scale)
  scale = tf.expand_dims(scale, axis=0)
  scale = tf.expand_dims(scale, axis=0)
  image /= tf.cast(scale, image.dtype)
  return image


def unnormalize_image(image,
                    offset=(0.48145466, 0.4578275, 0.40821073),
                    scale=(0.26862954, 0.26130258, 0.27577711)):
  """Normalizes the image to zero mean and unit variance."""
  scale = tf.cast(tf.expand_dims(tf.expand_dims(tf.constant(scale), axis=0), axis=0), image.dtype)
  image *= scale

  offset = tf.cast(tf.expand_dims(tf.expand_dims(tf.constant(offset), axis=0), axis=0), image.dtype)
  image += offset
  return image


def flatten_parts(ds: tf.data.Dataset, parts: List[str], add_index=False, dataset_size=None) -> tf.data.Dataset:
    def _flatten(ex):
        flat_key = {k: ex[k] for k in parts}
        if add_index:
            flat_key['index'] = tf.range(len(ex[parts[0]]))

        flat_ds = tf.data.Dataset.from_tensor_slices(flat_key)

        def _merge(_flat_ex):
            for k, v in ex.items():
                if k not in parts:
                    _flat_ex[k] = v
            return _flat_ex
        return flat_ds.map(_merge)

    ds = ds.flat_map(_flatten)
    if dataset_size is not None:
        ds = tf.data.experimental.assert_cardinality(dataset_size)(ds)
    return ds