yt8m_input.py 19.2 KB
Newer Older
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
1
# Copyright 2022 The TensorFlow Authors. All Rights Reserved.
Hye Yoon's avatar
Hye Yoon committed
2
3
4
5
6
7
8
9
10
11
12
13
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
14
15
16

"""class YT8MFrameFeatureReader(BaseReader).

Hye Yoon's avatar
Hye Yoon committed
17
18
19
20
21
22
23
  Reads TFRecords of SequenceExamples.

  The TFRecords must contain SequenceExamples with the sparse in64 'labels'
  context feature and a fixed length byte-quantized feature vector, obtained
  from the features in 'feature_names'. The quantized features will be mapped
  back into a range between min_quantized_value and max_quantized_value.
  link for details: https://research.google.com/youtube8m/download.html
24
"""
Chaochao Yan's avatar
Chaochao Yan committed
25
from typing import Any, Dict
Hye Yoon's avatar
Hye Yoon committed
26
27

import tensorflow as tf
Yeqing Li's avatar
Yeqing Li committed
28
from official.projects.yt8m.dataloaders import utils
Yeqing Li's avatar
Yeqing Li committed
29
30
31
from official.vision.configs import video_classification as exp_cfg
from official.vision.dataloaders import decoder
from official.vision.dataloaders import parser
Hye Yoon's avatar
Hye Yoon committed
32
33
34


def resize_axis(tensor, axis, new_size, fill_value=0):
Yulv-git's avatar
Yulv-git committed
35
  """Truncates or pads a tensor to new_size on a given axis.
Hye Yoon's avatar
Hye Yoon committed
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

  Truncate or extend tensor such that tensor.shape[axis] == new_size. If the
  size increases, the padding will be performed at the end, using fill_value.

  Args:
    tensor: The tensor to be resized.
    axis: An integer representing the dimension to be sliced.
    new_size: An integer or 0d tensor representing the new value for
      tensor.shape[axis].
    fill_value: Value to use to fill any new entries in the tensor. Will be cast
      to the type of tensor.

  Returns:
    The resized tensor.
  """
  tensor = tf.convert_to_tensor(tensor)
  shape = tf.unstack(tf.shape(tensor))

  pad_shape = shape[:]
  pad_shape[axis] = tf.maximum(0, new_size - shape[axis])

  shape[axis] = tf.minimum(shape[axis], new_size)
  shape = tf.stack(shape)

  resized = tf.concat([
61
62
      tf.slice(tensor, tf.zeros_like(shape), shape),
      tf.fill(tf.stack(pad_shape), tf.cast(fill_value, tensor.dtype))
Hye Yoon's avatar
Hye Yoon committed
63
64
65
66
67
68
69
70
71
  ], axis)

  # Update shape.
  new_shape = tensor.shape.as_list()  # A copy is being made.
  new_shape[axis] = new_size
  resized = tf.ensure_shape(resized, new_shape)
  return resized


72
73
def _process_segment_and_label(video_matrix, num_frames, contexts,
                               segment_labels, segment_size,
Hye Yoon's avatar
Hye Yoon committed
74
75
                               num_classes) -> Dict[str, tf.Tensor]:
  """Processes a batched Tensor of frames.
76

Hye Yoon's avatar
Hye Yoon committed
77
78
79
80
81
82
  The same parameters used in process should be used here.
  Args:
    video_matrix: different features concatenated into one matrix
    num_frames: Number of frames per subclip.
    contexts: context information extracted from decoder
    segment_labels: if we read segment labels instead.
83
    segment_size: the segment_size used for reading segments. Segment length.
Hye Yoon's avatar
Hye Yoon committed
84
85
86
87
88
89
    num_classes: a positive integer for the number of classes.

  Returns:
    output: dictionary containing batch information
  """
  # Partition frame-level feature matrix to segment-level feature matrix.
90
  batch_video_ids = None
Hye Yoon's avatar
Hye Yoon committed
91
92
93
94
  if segment_labels:
    start_times = contexts["segment_start_times"].values
    # Here we assume all the segments that started at the same start time has
    # the same segment_size.
95
    uniq_start_times, seg_idxs = tf.unique(start_times, out_idx=tf.dtypes.int64)
Hye Yoon's avatar
Hye Yoon committed
96
    # Range gather matrix, e.g., [[0,1,2],[1,2,3]] for segment_size == 3.
97
98
99
    range_mtx = tf.expand_dims(
        uniq_start_times, axis=-1) + tf.expand_dims(
            tf.range(0, segment_size, dtype=tf.int64), axis=0)
Hye Yoon's avatar
Hye Yoon committed
100
101
102
103
    # Shape: [num_segment, segment_size, feature_dim].
    batch_video_matrix = tf.gather_nd(video_matrix,
                                      tf.expand_dims(range_mtx, axis=-1))
    num_segment = tf.shape(batch_video_matrix)[0]
104
105
106
    if "id" in contexts:
      batch_video_ids = tf.reshape(
          tf.tile([contexts["id"]], [num_segment]), (num_segment,))
107
108
    batch_frames = tf.reshape(
        tf.tile([segment_size], [num_segment]), (num_segment,))
Hye Yoon's avatar
Hye Yoon committed
109
110
111
112
113
114
115
116
117
118
119
120
121
122
    batch_frames = tf.cast(tf.expand_dims(batch_frames, 1), tf.float32)

    # For segment labels, all labels are not exhaustively rated. So we only
    # evaluate the rated labels.

    # Label indices for each segment, shape: [num_segment, 2].
    label_indices = tf.stack([seg_idxs, contexts["segment_labels"].values],
                             axis=-1)
    label_values = contexts["segment_scores"].values
    sparse_labels = tf.sparse.SparseTensor(label_indices, label_values,
                                           (num_segment, num_classes))
    batch_labels = tf.sparse.to_dense(sparse_labels, validate_indices=False)

    sparse_label_weights = tf.sparse.SparseTensor(
123
124
125
126
        label_indices, tf.ones_like(label_values, dtype=tf.float32),
        (num_segment, num_classes))
    batch_label_weights = tf.sparse.to_dense(
        sparse_label_weights, validate_indices=False)
Hye Yoon's avatar
Hye Yoon committed
127
128
129
130
131
    # output_dict = utils.get_segments(batch_video_matrix, batch_frames, 5)
  else:
    # Process video-level labels.
    label_indices = contexts["labels"].values
    sparse_labels = tf.sparse.SparseTensor(
132
133
134
135
        tf.expand_dims(label_indices, axis=-1),
        tf.ones_like(contexts["labels"].values, dtype=tf.bool), (num_classes,))
    labels = tf.sparse.to_dense(
        sparse_labels, default_value=False, validate_indices=False)
Hye Yoon's avatar
Hye Yoon committed
136
137

    # convert to batch format.
138
139
    if "id" in contexts:
      batch_video_ids = tf.expand_dims(contexts["id"], 0)
Hye Yoon's avatar
Hye Yoon committed
140
141
142
143
144
145
    batch_video_matrix = tf.expand_dims(video_matrix, 0)
    batch_labels = tf.expand_dims(labels, 0)
    batch_frames = tf.expand_dims(num_frames, 0)
    batch_label_weights = None

  output_dict = {
146
147
148
      "video_matrix": batch_video_matrix,
      "labels": batch_labels,
      "num_frames": batch_frames,
Hye Yoon's avatar
Hye Yoon committed
149
  }
150
151
  if batch_video_ids is not None:
    output_dict["video_ids"] = batch_video_ids
Hye Yoon's avatar
Hye Yoon committed
152
153
154
155
156
157
  if batch_label_weights is not None:
    output_dict["label_weights"] = batch_label_weights

  return output_dict


158
159
def _get_video_matrix(features, feature_size, dtype, max_frames,
                      max_quantized_value, min_quantized_value):
Hye Yoon's avatar
Hye Yoon committed
160
161
  """Decodes features from an input string and quantizes it.

162
  Args:
163
164
165
166
    features: raw feature values.
    feature_size: length of each frame feature vector.
    dtype: raw type of the feature.
    max_frames: number of frames (rows) in the output feature_matrix.
167
168
    max_quantized_value: the maximum of the quantized value.
    min_quantized_value: the minimum of the quantized value.
Hye Yoon's avatar
Hye Yoon committed
169

170
171
172
173
  Returns:
    feature_matrix: matrix of all frame-features
    num_frames: number of frames in the sequence
  """
174
  decoded_features = tf.reshape(features, [-1, feature_size])
Hye Yoon's avatar
Hye Yoon committed
175
176

  num_frames = tf.math.minimum(tf.shape(decoded_features)[0], max_frames)
177
  if dtype.is_integer:
178
    feature_matrix = utils.dequantize(decoded_features, max_quantized_value,
179
180
181
                                      min_quantized_value)
  else:
    feature_matrix = decoded_features
Hye Yoon's avatar
Hye Yoon committed
182
183
184
185
  feature_matrix = resize_axis(feature_matrix, 0, max_frames)
  return feature_matrix, num_frames


186
187
def _concat_features(features, feature_names, feature_sizes, feature_dtypes,
                     max_frames, max_quantized_value, min_quantized_value):
188
  """Loads (potentially) different types of features and concatenates them.
Hye Yoon's avatar
Hye Yoon committed
189

190
191
192
193
  Args:
      features: raw feature values
      feature_names: list of feature names
      feature_sizes: list of features sizes
194
      feature_dtypes: dtype of the feature.
195
196
197
      max_frames: number of frames in the sequence
      max_quantized_value: the maximum of the quantized value.
      min_quantized_value: the minimum of the quantized value.
Hye Yoon's avatar
Hye Yoon committed
198

199
200
201
202
  Returns:
      video_matrix: different features concatenated into one matrix
      num_frames: the number of frames in the video
  """
Hye Yoon's avatar
Hye Yoon committed
203
204
205
206
207

  num_features = len(feature_names)
  assert num_features > 0, "No feature selected: feature_names is empty!"

  assert len(feature_names) == len(feature_sizes), (
208
209
      "length of feature_names (={}) != length of feature_sizes (={})".format(
          len(feature_names), len(feature_sizes)))
210
211
212
  assert len(feature_names) == len(feature_dtypes), (
      "length of feature_names (={}) != length of feature_sizes (={})".format(
          len(feature_names), len(feature_dtypes)))
Hye Yoon's avatar
Hye Yoon committed
213
214
215

  num_frames = -1  # the number of frames in the video
  feature_matrices = [None] * num_features  # an array of different features
216
  for i in range(num_features):
Hye Yoon's avatar
Hye Yoon committed
217
    feature_matrix, num_frames_in_this_feature = _get_video_matrix(
Chaochao Yan's avatar
Chaochao Yan committed
218
219
        features[feature_names[i]], feature_sizes[i],
        tf.dtypes.as_dtype(feature_dtypes[i]), max_frames, max_quantized_value,
220
        min_quantized_value)
Hye Yoon's avatar
Hye Yoon committed
221
222
    if num_frames == -1:
      num_frames = num_frames_in_this_feature
223
    feature_matrices[i] = feature_matrix
Hye Yoon's avatar
Hye Yoon committed
224
225
226
227
228
229
230
231
232
233
234

  # cap the number of frames at self.max_frames
  num_frames = tf.minimum(num_frames, max_frames)

  # concatenate different features
  video_matrix = tf.concat(feature_matrices, 1)

  return video_matrix, num_frames


class Decoder(decoder.Decoder):
Chaochao Yan's avatar
Chaochao Yan committed
235
  """A tf.train.SequeneExample decoder for classification task."""
Hye Yoon's avatar
Hye Yoon committed
236

237
238
239
240
  def __init__(
      self,
      input_params: exp_cfg.DataConfig,
  ):
Hye Yoon's avatar
Hye Yoon committed
241
242
243

    self._segment_labels = input_params.segment_labels
    self._feature_names = input_params.feature_names
244
245
246
247
248
    self._feature_sources = input_params.feature_sources
    self._feature_sizes = input_params.feature_sizes
    self._feature_dtypes = input_params.feature_dtypes
    self._feature_from_bytes = input_params.feature_from_bytes
    self._include_video_id = input_params.include_video_id
249
    self._label_field = input_params.label_field
250
251
252
253
254
255
256
257
258
259

    assert len(self._feature_names) == len(self._feature_sources), (
        "length of feature_names (={}) != length of feature_sizes (={})".format(
            len(self._feature_names), len(self._feature_sources)))

    self._context_features = {}
    self._sequence_features = {}
    if self._include_video_id:
      self._context_features["id"] = tf.io.FixedLenFeature([], tf.string)

Hye Yoon's avatar
Hye Yoon committed
260
261
    if self._segment_labels:
      self._context_features.update({
262
263
264
265
266
          # There is no need to read end-time given we always assume the segment
          # has the same size.
          "segment_labels": tf.io.VarLenFeature(tf.int64),
          "segment_start_times": tf.io.VarLenFeature(tf.int64),
          "segment_scores": tf.io.VarLenFeature(tf.float32)
Hye Yoon's avatar
Hye Yoon committed
267
268
      })
    else:
Chaochao Yan's avatar
Chaochao Yan committed
269
      self._add_labels_specification()
Hye Yoon's avatar
Hye Yoon committed
270

271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
    for i, name in enumerate(self._feature_names):
      if self._feature_from_bytes[i]:
        feature_type = tf.io.FixedLenSequenceFeature([], dtype=tf.string)
      else:
        dtype = tf.dtypes.as_dtype(self._feature_dtypes[i])
        feature_shape = [self._feature_sizes[i]]
        if self._feature_sources[i] == "feature":
          feature_type = tf.io.FixedLenSequenceFeature(feature_shape, dtype)
        else:
          feature_type = tf.io.FixedLenFeature(feature_shape, dtype)
      if self._feature_sources[i] == "feature":
        self._sequence_features[name] = feature_type
      elif self._feature_sources[i] == "context":
        self._context_features[name] = feature_type
      else:
        raise ValueError(
            f"Unknow feature source {self._feature_sources[i]} for {name}")
Hye Yoon's avatar
Hye Yoon committed
288

Chaochao Yan's avatar
Chaochao Yan committed
289
290
291
292
293
294
295
296
297
  def _add_labels_specification(self):
    if not self._label_field:
      raise ValueError(f"Invalid label field: {self._label_field}!")
    self._context_features.update(
        {self._label_field: tf.io.VarLenFeature(tf.int64)})

  def decode(self,
             serialized_example: tf.train.SequenceExample) -> Dict[str, Any]:
    """Parses a single tf.train.SequenceExample into video and label tensors."""
Hye Yoon's avatar
Hye Yoon committed
298
    contexts, features = tf.io.parse_single_sequence_example(
299
300
301
        serialized_example,
        context_features=self._context_features,
        sequence_features=self._sequence_features)
302
303
304
305
306
307
308
309
310
311
312
    decoded_tensor = {**contexts, **features}
    for i, name in enumerate(self._feature_names):
      # Convert the VarLen feature to dense tensor.
      if self._feature_from_bytes[i]:
        dtype = tf.dtypes.as_dtype(self._feature_dtypes[i])
        decoded_tensor[name] = tf.cast(
            tf.io.decode_raw(decoded_tensor[name], dtype), tf.float32),
      else:
        if isinstance(decoded_tensor[name], tf.SparseTensor):
          decoded_tensor[name] = tf.sparse.to_dense(decoded_tensor[name])
    return decoded_tensor
Hye Yoon's avatar
Hye Yoon committed
313
314
315
316


class Parser(parser.Parser):
  """Parses a video and label dataset.
317

Hye Yoon's avatar
Hye Yoon committed
318
319
320
321
322
323
    takes the decoded raw tensors dict
    and parse them into a dictionary of tensors
    that can be consumed by the model.
    It will be executed after decoder.
  """

324
325
326
327
328
329
  def __init__(
      self,
      input_params: exp_cfg.DataConfig,
      max_quantized_value=2,
      min_quantized_value=-2,
  ):
Hye Yoon's avatar
Hye Yoon committed
330
    self._num_classes = input_params.num_classes
Chaochao Yan's avatar
Chaochao Yan committed
331
    self._label_field = input_params.label_field
Hye Yoon's avatar
Hye Yoon committed
332
333
    self._segment_size = input_params.segment_size
    self._segment_labels = input_params.segment_labels
334
    self._include_video_id = input_params.include_video_id
Hye Yoon's avatar
Hye Yoon committed
335
336
    self._feature_names = input_params.feature_names
    self._feature_sizes = input_params.feature_sizes
337
    self._feature_dtypes = input_params.feature_dtypes
Hye Yoon's avatar
Hye Yoon committed
338
339
340
341
342
343
344
345
    self._max_frames = input_params.max_frames
    self._max_quantized_value = max_quantized_value
    self._min_quantized_value = min_quantized_value

  def _parse_train_data(self, decoded_tensors):
    """Parses data for training."""
    # loads (potentially) different types of features and concatenates them
    self.video_matrix, self.num_frames = _concat_features(
346
347
348
349
350
        decoded_tensors, self._feature_names, self._feature_sizes,
        self._feature_dtypes, self._max_frames, self._max_quantized_value,
        self._min_quantized_value)
    if not self._include_video_id and "id" in decoded_tensors:
      del decoded_tensors["id"]
Chaochao Yan's avatar
Chaochao Yan committed
351
352
353

    return self._process_label(self.video_matrix, self.num_frames,
                               decoded_tensors)
Hye Yoon's avatar
Hye Yoon committed
354
355
356
357
358

  def _parse_eval_data(self, decoded_tensors):
    """Parses data for evaluation."""
    # loads (potentially) different types of features and concatenates them
    self.video_matrix, self.num_frames = _concat_features(
359
360
361
362
363
        decoded_tensors, self._feature_names, self._feature_sizes,
        self._feature_dtypes, self._max_frames, self._max_quantized_value,
        self._min_quantized_value)
    if not self._include_video_id and "id" in decoded_tensors:
      del decoded_tensors["id"]
Chaochao Yan's avatar
Chaochao Yan committed
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378

    return self._process_label(self.video_matrix, self.num_frames,
                               decoded_tensors)

  def _process_label(self, video_matrix, num_frames, contexts):
    """Processes a batched Tensor of frames.

    Args:
      video_matrix: video feature matric.
      num_frames: number of frames in this video.
      contexts: context information extracted from decoder.

    Returns:
      output: dictionary containing batch information
    """
Chaochao Yan's avatar
Chaochao Yan committed
379
380
    if self._label_field and not self._segment_labels:
      contexts["labels"] = contexts[self._label_field]
Chaochao Yan's avatar
Chaochao Yan committed
381
    output_dict = _process_segment_and_label(video_matrix, num_frames, contexts,
382
383
384
                                             self._segment_labels,
                                             self._segment_size,
                                             self._num_classes)
Chaochao Yan's avatar
Chaochao Yan committed
385
    return output_dict
Hye Yoon's avatar
Hye Yoon committed
386
387
388
389
390
391
392
393
394
395
396
397

  def parse_fn(self, is_training):
    """Returns a parse fn that reads and parses raw tensors from the decoder.

    Args:
      is_training: a `bool` to indicate whether it is in training mode.

    Returns:
      parse: a `callable` that takes the serialized example and generate the
        images, labels tuple where labels is a dict of Tensors that contains
        labels.
    """
398

Hye Yoon's avatar
Hye Yoon committed
399
400
401
402
403
404
405
406
407
    def parse(decoded_tensors):
      """Parses the serialized example data."""
      if is_training:
        return self._parse_train_data(decoded_tensors)
      else:
        return self._parse_eval_data(decoded_tensors)

    return parse

408

Hye Yoon's avatar
Hye Yoon committed
409
class TransformBatcher():
410
411
412
  """Performs manual batching on input dataset."""

  def __init__(self, input_params: exp_cfg.DataConfig):
Hye Yoon's avatar
Hye Yoon committed
413
414
415
    self._segment_labels = input_params.segment_labels
    self._global_batch_size = input_params.global_batch_size
    self._is_training = input_params.is_training
416
    self._include_video_id = input_params.include_video_id
Yeqing Li's avatar
Yeqing Li committed
417
    self._drop_remainder = input_params.drop_remainder
Hye Yoon's avatar
Hye Yoon committed
418
419

  def batch_fn(self, dataset, input_context):
420
    """Add padding when segment_labels is true."""
Hye Yoon's avatar
Hye Yoon committed
421
422
423
    per_replica_batch_size = input_context.get_per_replica_batch_size(
        self._global_batch_size) if input_context else self._global_batch_size
    if not self._segment_labels:
Yeqing Li's avatar
Yeqing Li committed
424
425
      dataset = dataset.batch(
          per_replica_batch_size, drop_remainder=self._drop_remainder)
Hye Yoon's avatar
Hye Yoon committed
426
427
    else:
      # add padding
428
429
430
431
432
433
434
435
436
437
438
439
      pad_shapes = {
          "video_matrix": [None, None, None],
          "labels": [None, None],
          "num_frames": [None, None],
          "label_weights": [None, None]
      }
      pad_values = {
          "video_matrix": 0.0,
          "labels": -1.0,
          "num_frames": 0.0,
          "label_weights": 0.0
      }
440
441
442
      if self._include_video_id:
        pad_shapes["video_ids"] = [None]
        pad_values["video_ids"] = None
Hye Yoon's avatar
Hye Yoon committed
443
      dataset = dataset.padded_batch(
444
445
          per_replica_batch_size,
          padded_shapes=pad_shapes,
Yeqing Li's avatar
Yeqing Li committed
446
          drop_remainder=self._drop_remainder,
447
          padding_values=pad_values)
Hye Yoon's avatar
Hye Yoon committed
448
    return dataset
Chaochao Yan's avatar
Chaochao Yan committed
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


class PostBatchProcessor():
  """Processes a video and label dataset which is batched."""

  def __init__(self, input_params: exp_cfg.DataConfig):
    self.segment_labels = input_params.segment_labels
    self.num_classes = input_params.num_classes
    self.segment_size = input_params.segment_size
    self.num_features = sum(input_params.feature_sizes)

  def post_fn(self, batched_tensors: Dict[str,
                                          tf.Tensor]) -> Dict[str, tf.Tensor]:
    """Processes batched Tensors."""
    video_ids = batched_tensors.get("video_ids", None)
    video_matrix = batched_tensors["video_matrix"]
    labels = batched_tensors["labels"]
    num_frames = batched_tensors["num_frames"]

    if self.segment_labels:
      # [batch x num_segment x segment_size x num_features]
      # -> [batch * num_segment x segment_size x num_features]
      if video_ids is not None:
        video_ids = tf.reshape(video_ids, [-1])
      video_matrix = tf.reshape(video_matrix,
                                [-1, self.segment_size, self.num_features])
      labels = tf.reshape(labels, [-1, self.num_classes])
      num_frames = tf.reshape(num_frames, [-1, 1])
      batched_tensors["label_weights"] = tf.reshape(
          batched_tensors["label_weights"], [-1, self.num_classes])
    else:
      # NOTE(b/237445211): Must provide axis argument to tf.squeeze.
      video_matrix = tf.squeeze(video_matrix, axis=1)
      labels = tf.squeeze(labels, axis=1)
      num_frames = tf.reshape(num_frames, [-1, 1])
      if "label_weights" in batched_tensors:
        batched_tensors["label_weights"] = tf.squeeze(
            batched_tensors["label_weights"], axis=1)

    batched_tensors.update({
        "video_matrix": video_matrix,
        "labels": labels,
        "num_frames": num_frames,
    })
    if video_ids is not None:
      batched_tensors["video_ids"] = video_ids

    return batched_tensors