Skip to content

ocl.transforms

Module with data pipe transforms.

Transforms are callables which transform a input torchdata datapipe into a new datapipe. For further information see ocl.transforms.Transform.

Transform

Bases: ABC

Abstract Base Class representing a transformation of the input data pipe.

A transform is a callable which when called with a torchdata.datapipes.iter.IterDataPipe applies a transformation and returns a new torchdata.datapipes.iter.IterDataPipe.

Attributes:

Name Type Description
is_batch_transform bool

True if the transform should be applied to a batch of examples instead of individual examples. False otherwise.
fields Tuple[str]

Tuple of strings, that indicate which elements of the input are needed for this transform to be applied. This allows to avoid decoding parts of the input which are not needed for training/evaluating a particular model.
Source code in ocl/transforms.py 
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
class Transform(ABC):
    """Abstract Base Class representing a transformation of the input data pipe.

    A transform is a callable which when called with a [torchdata.datapipes.iter.IterDataPipe][]
    applies a transformation and returns a new [torchdata.datapipes.iter.IterDataPipe][].


    Attributes:
        is_batch_transform: True if the transform should be applied to a batch of
            examples instead of individual examples. False otherwise.
        fields: Tuple of strings, that indicate which elements of the input are needed
            for this transform to be applied.  This allows to avoid decoding parts of the
            input which are not needed for training/evaluating a particular model.

    """

    is_batch_transform: bool

    @property
    @abstractmethod
    def fields(self) -> Tuple[str]:
        """Fields that will be transformed with this transform."""

    @abstractmethod
    def __call__(self, input_pipe: IterDataPipe) -> IterDataPipe:
        """Application of transform to input pipe.

        Args:
            input_pipe: Input data pipe

        Returns:
            Transformed data pipe.
        """

fields: Tuple[str] property abstractmethod

Fields that will be transformed with this transform.

__call__ abstractmethod

Application of transform to input pipe.

Parameters:

Name Type Description Default
input_pipe IterDataPipe

Input data pipe

 required

Returns:

Type Description
IterDataPipe

Transformed data pipe.
Source code in ocl/transforms.py 
47
48
49
50
51
52
53
54
55
56
@abstractmethod
def __call__(self, input_pipe: IterDataPipe) -> IterDataPipe:
    """Application of transform to input pipe.

    Args:
        input_pipe: Input data pipe

    Returns:
        Transformed data pipe.
    """

SimpleTransform

Bases: Transform

Transform of individual key in input dict using different callables.

Example 
from torchdata.datapipes.iter import IterableWrapper
from ocl.transforms import SimpleTransform

input_dicts = [{"object_a": 1, "object_b": 2}]
transform = SimpleTransform(
    transforms={
        "object_a": lambda a: a*2,
        "object_b": lambda b: b*3
    }
)

input_pipe = IterableWrapper(input_dicts)
transformed_pipe = transform(input_pipe)

for transformed_dict in transformed_pipe:
    assert transformed_dict["object_a"] == 1 * 2
    assert transformed_dict["object_b"] == 2 * 3
Source code in ocl/transforms.py 
 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
class SimpleTransform(Transform):
    """Transform of individual key in input dict using different callables.

    Example:
        ```python
        from torchdata.datapipes.iter import IterableWrapper
        from ocl.transforms import SimpleTransform

        input_dicts = [{"object_a": 1, "object_b": 2}]
        transform = SimpleTransform(
            transforms={
                "object_a": lambda a: a*2,
                "object_b": lambda b: b*3
            }
        )

        input_pipe = IterableWrapper(input_dicts)
        transformed_pipe = transform(input_pipe)

        for transformed_dict in transformed_pipe:
            assert transformed_dict["object_a"] == 1 * 2
            assert transformed_dict["object_b"] == 2 * 3
        ```
    """

    def __init__(self, transforms: Dict[str, Callable], batch_transform: bool):
        """Initialize SimpleTransform.

        Args:
            transforms: Mapping of dict keys to callables that should be used to transform them.
            batch_transform: Set to true if you want your transform to be applied after the
                data has been batched.
        """
        self.transforms = transforms
        self.is_batch_transform = batch_transform

    @property
    def fields(self) -> Tuple[str]:
        return tuple(self.transforms.keys())

    def __call__(self, input_pipe: IterDataPipe) -> IterDataPipe:
        """Transform input data pipe using transforms.

        Args:
            input_pipe: Input data pipe

        Returns:
            Transformed data pipe.
        """
        return input_pipe.map_dict(self.transforms)

__init__

Initialize SimpleTransform.

Parameters:

Name Type Description Default
transforms Dict[str, Callable]

Mapping of dict keys to callables that should be used to transform them.

 required
batch_transform bool

Set to true if you want your transform to be applied after the data has been batched.

 required
Source code in ocl/transforms.py 
84
85
86
87
88
89
90
91
92
93
def __init__(self, transforms: Dict[str, Callable], batch_transform: bool):
    """Initialize SimpleTransform.

    Args:
        transforms: Mapping of dict keys to callables that should be used to transform them.
        batch_transform: Set to true if you want your transform to be applied after the
            data has been batched.
    """
    self.transforms = transforms
    self.is_batch_transform = batch_transform

__call__

Transform input data pipe using transforms.

Parameters:

Name Type Description Default
input_pipe IterDataPipe

Input data pipe

 required

Returns:

Type Description
IterDataPipe

Transformed data pipe.
Source code in ocl/transforms.py 
 99
100
101
102
103
104
105
106
107
108
def __call__(self, input_pipe: IterDataPipe) -> IterDataPipe:
    """Transform input data pipe using transforms.

    Args:
        input_pipe: Input data pipe

    Returns:
        Transformed data pipe.
    """
    return input_pipe.map_dict(self.transforms)

DuplicateFields

Bases: Transform

Transform to duplicate a key of a dictionary.

This is useful if your pipeline requires the same input to be transformed in different ways.

Example 
from torchdata.datapipes.iter import IterableWrapper
from ocl.transforms import DuplicateFields

input_dicts = [{"object_a": 1, "object_b": 2}]
transform = DuplicateFields(
    mapping={
        "object_a": "copy_of_object_a",
        "object_b": "copy_of_object_b"
    }
)

input_pipe = IterableWrapper(input_dicts)
transformed_pipe = transform(input_pipe)

for transformed_dict in transformed_pipe:
    assert transformed_dict["object_a"] == 1
    assert transformed_dict["copy_of_object_a"] == 1
    assert transformed_dict["object_b"] == 2
    assert transformed_dict["copy_of_object_b"] == 2
Source code in ocl/transforms.py 
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
class DuplicateFields(Transform):
    """Transform to duplicate a key of a dictionary.

    This is useful if your pipeline requires the same input to be transformed in different ways.

    Example:
        ```python
        from torchdata.datapipes.iter import IterableWrapper
        from ocl.transforms import DuplicateFields

        input_dicts = [{"object_a": 1, "object_b": 2}]
        transform = DuplicateFields(
            mapping={
                "object_a": "copy_of_object_a",
                "object_b": "copy_of_object_b"
            }
        )

        input_pipe = IterableWrapper(input_dicts)
        transformed_pipe = transform(input_pipe)

        for transformed_dict in transformed_pipe:
            assert transformed_dict["object_a"] == 1
            assert transformed_dict["copy_of_object_a"] == 1
            assert transformed_dict["object_b"] == 2
            assert transformed_dict["copy_of_object_b"] == 2
        ```
    """

    def __init__(self, mapping: Dict[str, str], batch_transform: bool):
        """Initialize DuplicateFields.

        Args:
            mapping: Source to target mapping for dupplicated fields. Keys are sources,
                values are the key for duplicated field.
            batch_transform: Apply to batched input.
        """
        self.mapping = mapping
        self.is_batch_transform = batch_transform

    def duplicate_keys(self, input_dict: Dict[str, Any]):
        for key, value in self.mapping.items():
            input_dict[value] = input_dict[key]
        return input_dict

    @property
    def fields(self) -> Tuple[str]:
        return tuple(self.mapping.keys())

    def __call__(self, input_pipe: IterDataPipe) -> IterDataPipe:
        return input_pipe.map(self.duplicate_keys)

__init__

Initialize DuplicateFields.

Parameters:

Name Type Description Default
mapping Dict[str, str]

Source to target mapping for dupplicated fields. Keys are sources, values are the key for duplicated field.

 required
batch_transform bool

Apply to batched input.

 required
Source code in ocl/transforms.py 
140
141
142
143
144
145
146
147
148
149
def __init__(self, mapping: Dict[str, str], batch_transform: bool):
    """Initialize DuplicateFields.

    Args:
        mapping: Source to target mapping for dupplicated fields. Keys are sources,
            values are the key for duplicated field.
        batch_transform: Apply to batched input.
    """
    self.mapping = mapping
    self.is_batch_transform = batch_transform

Map

Bases: Transform

Apply a function to the whole input dict to create a new output dict.

This transform requires explicitly defining the input fields as this cannot be determined from the provided callable alone.

Example 
from torchdata.datapipes.iter import IterableWrapper
from ocl.transforms import Map

input_dicts = [{"object_a": 1, "object_b": 2}]

def combine_a_and_b(input_dict):
    output_dict = input_dict.copy()
    output_dict["combined"] = input_dict["object_a"] + input_dict["object_b"]
    return output_dict

transform = Map(
    transform=combine_a_and_b,
    fields=("object_a", "object_b")
)

input_pipe = IterableWrapper(input_dicts)
transformed_pipe = transform(input_pipe)

for transformed_dict in transformed_pipe:
    a = transformed_dict["object_a"]
    b = transformed_dict["object_b"]
    assert transformed_dict["combined"] == a + b
Source code in ocl/transforms.py 
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
class Map(Transform):
    """Apply a function to the whole input dict to create a new output dict.

    This transform requires explicitly defining the input fields as this
    cannot be determined from the provided callable alone.

    Example:
        ```python
        from torchdata.datapipes.iter import IterableWrapper
        from ocl.transforms import Map

        input_dicts = [{"object_a": 1, "object_b": 2}]

        def combine_a_and_b(input_dict):
            output_dict = input_dict.copy()
            output_dict["combined"] = input_dict["object_a"] + input_dict["object_b"]
            return output_dict

        transform = Map(
            transform=combine_a_and_b,
            fields=("object_a", "object_b")
        )

        input_pipe = IterableWrapper(input_dicts)
        transformed_pipe = transform(input_pipe)

        for transformed_dict in transformed_pipe:
            a = transformed_dict["object_a"]
            b = transformed_dict["object_b"]
            assert transformed_dict["combined"] == a + b
        ```
    """

    def __init__(
        self,
        transform: Callable[[Dict[str, Any]], Dict[str, Any]],
        fields: Tuple[str],
        batch_transform: bool,
    ):
        """Initialize Map transform.

        Args:
            transform: Callable which is applied to the individual input dictionaries.
            fields: The fields the transform requires to operate.
            batch_transform: Apply to batched input.
        """
        self.transfrom = transform
        self._fields = fields
        self.is_batch_transform = batch_transform

    @property
    def fields(self) -> Tuple[str]:
        return self._fields

    def __call__(self, input_pipe: IterDataPipe) -> IterDataPipe:
        return input_pipe.map(self.transfrom)

__init__

Initialize Map transform.

Parameters:

Name Type Description Default
transform Callable[[Dict[str, Any]], Dict[str, Any]]

Callable which is applied to the individual input dictionaries.

 required
fields Tuple[str]

The fields the transform requires to operate.

 required
batch_transform bool

Apply to batched input.

 required
Source code in ocl/transforms.py 
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
def __init__(
    self,
    transform: Callable[[Dict[str, Any]], Dict[str, Any]],
    fields: Tuple[str],
    batch_transform: bool,
):
    """Initialize Map transform.

    Args:
        transform: Callable which is applied to the individual input dictionaries.
        fields: The fields the transform requires to operate.
        batch_transform: Apply to batched input.
    """
    self.transfrom = transform
    self._fields = fields
    self.is_batch_transform = batch_transform

Filter

Bases: Transform

Filter samples according to predicate.

Remove samples from input data pipe by evaluating a predicate.

Example 
from torchdata.datapipes.iter import IterableWrapper
from ocl.transforms import Filter

input_dicts = [{"myvalue": 5}, {"myvalue": 10}]

transform = Filter(
    predicate=lambda a: a > 5,
    fields=("myvalue",)
)

input_pipe = IterableWrapper(input_dicts)
transformed_pipe = transform(input_pipe)

for transformed_dict in transformed_pipe:
    assert transformed_dict["myvalue"] > 5
Source code in ocl/transforms.py 
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
class Filter(Transform):
    """Filter samples according to predicate.

    Remove samples from input data pipe by evaluating a predicate.

    Example:
        ```python
        from torchdata.datapipes.iter import IterableWrapper
        from ocl.transforms import Filter

        input_dicts = [{"myvalue": 5}, {"myvalue": 10}]

        transform = Filter(
            predicate=lambda a: a > 5,
            fields=("myvalue",)
        )

        input_pipe = IterableWrapper(input_dicts)
        transformed_pipe = transform(input_pipe)

        for transformed_dict in transformed_pipe:
            assert transformed_dict["myvalue"] > 5
        ```


    """

    def __init__(self, predicate: Callable[..., bool], fields: Sequence[str]):
        """Transform to create a subset of a dataset by discarding samples.

        Args:
            predicate: Function which determines if elements should be kept (return value is True)
                or discarded (return value is False). The function is only provided with the fields
                specified in the `fields` parameter.
            fields (Sequence[str]): The fields from the input which should be passed on to the
                predicate for evaluation.
        """
        self.predicate = predicate
        self._fields = tuple(fields)
        self.is_batch_transform = False

    @property
    def fields(self):
        return self._fields

    def _filter_using_predicate(self, d: Dict[str, Any]):
        return self.predicate(*(d[field] for field in self._fields))

    def __call__(self, input_pipe: IterDataPipe):
        return input_pipe.filter(self._filter_using_predicate)

__init__

Transform to create a subset of a dataset by discarding samples.

Parameters:

Name Type Description Default
predicate Callable[..., bool]

Function which determines if elements should be kept (return value is True) or discarded (return value is False). The function is only provided with the fields specified in the fields parameter.

 required
fields Sequence[str]

The fields from the input which should be passed on to the predicate for evaluation.

 required
Source code in ocl/transforms.py 
249
250
251
252
253
254
255
256
257
258
259
260
261
def __init__(self, predicate: Callable[..., bool], fields: Sequence[str]):
    """Transform to create a subset of a dataset by discarding samples.

    Args:
        predicate: Function which determines if elements should be kept (return value is True)
            or discarded (return value is False). The function is only provided with the fields
            specified in the `fields` parameter.
        fields (Sequence[str]): The fields from the input which should be passed on to the
            predicate for evaluation.
    """
    self.predicate = predicate
    self._fields = tuple(fields)
    self.is_batch_transform = False

SampleSlices

Bases: Transform

Transform to sample slices from input tensors / numpy arrays.

If multiple fields are provided the input tensors are assumed to be of same length along slicing axes and the same slices will be returned.

Example 
import numpy as np
from torchdata.datapipes.iter import IterableWrapper
from ocl.transforms import SampleSlices

my_array = np.random.randn(100, 10)

input_dicts = [{"array1": my_array, "array2": my_array.copy()}]

transform = SampleSlices(
    n_slices_per_input=5,
    fields=("array1", "array2"),
    dim=0,
    shuffle_buffer_size=1
)

input_pipe = IterableWrapper(input_dicts)
transformed_pipe = transform(input_pipe)

for transformed_dict in transformed_pipe:
    assert transformed_dict["array1"].shape == (5, 10)
    assert transformed_dict["array2"].shape == (5, 10)
    assert np.allclose(transformed_dict["array1"], transformed_dict["array2"])
Source code in ocl/transforms.py 
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
class SampleSlices(Transform):
    """Transform to sample slices from input tensors / numpy arrays.

    If multiple fields are provided the input tensors are assumed to be of
    same length along slicing axes and the same slices will be returned.

    Example:
        ```python
        import numpy as np
        from torchdata.datapipes.iter import IterableWrapper
        from ocl.transforms import SampleSlices

        my_array = np.random.randn(100, 10)

        input_dicts = [{"array1": my_array, "array2": my_array.copy()}]

        transform = SampleSlices(
            n_slices_per_input=5,
            fields=("array1", "array2"),
            dim=0,
            shuffle_buffer_size=1
        )

        input_pipe = IterableWrapper(input_dicts)
        transformed_pipe = transform(input_pipe)

        for transformed_dict in transformed_pipe:
            assert transformed_dict["array1"].shape == (5, 10)
            assert transformed_dict["array2"].shape == (5, 10)
            assert np.allclose(transformed_dict["array1"], transformed_dict["array2"])
        ```
    """

    def __init__(
        self,
        n_slices_per_input: int,
        fields: Sequence[str],
        dim: int = 0,
        seed: int = 39480234,
        per_epoch: bool = False,
        shuffle_buffer_size: int = 1000,
    ):
        """Initialize SampleSlices.

        Args:
            n_slices_per_input: Number of slices per input to sample. -1 indicates that all possible
                slices should be sampled.
            fields: The fields that should be considered video data and thus sliced
                according to the frame sampling during training.
            dim: The dimension along which to slice the tensors.
            seed: Random number generator seed to deterministic sampling during evaluation.
            per_epoch: Sampling of frames over epochs, this ensures that after
                n_frames / n_frames_per_video epochs all frames have been seen at least once.
                In the case of uneven division, some frames will be seen more than once.
            shuffle_buffer_size: Size of shuffle buffer used during training. An additional
                shuffling step ensures each batch contains a diverse set of images and not only
                images from the same video.
        """
        self.n_slices_per_input = n_slices_per_input
        self._fields = tuple(fields)
        self.dim = dim
        self.seed = seed
        self.per_epoch = per_epoch
        self.shuffle_buffer_size = shuffle_buffer_size
        self.is_batch_transform = False

    def slice_data(self, data, index: int):
        """Small utility method to slice a numpy array along a specified axis."""
        n_dims_before = self.dim
        n_dims_after = data.ndim - 1 - self.dim
        slices = (slice(None),) * n_dims_before + (index,) + (slice(None),) * n_dims_after
        return data[slices]

    def sample_frames_using_key(self, data):
        """Sample frames deterministically from generator of videos using the __key__ field."""
        for sample in data:
            # Initialize random number generator dependent on instance key. This should make the
            # sampling process deterministic, which is useful when sampling frames for the
            # validation/test data.
            key = sample["__key__"]
            # TODO (hornmax): We assume all fields to have the same size. I do not want to check
            # this here as it seems a bit verbose.
            n_frames = sample[self._fields[0]].shape[self.dim]
            frames_per_video = self.n_slices_per_input if self.n_slices_per_input != -1 else n_frames

            if self.per_epoch and self.n_slices_per_input != -1:
                n_different_epochs_per_seed = int(math.ceil(n_frames / frames_per_video))
                try:
                    epoch = int(os.environ["EPOCH"])
                except KeyError:
                    raise RuntimeError(
                        "Using SampleSlices with per_epoch=True "
                        "requires environment variable `EPOCH` to be set. "
                        "You might need to add the SetEpochEnvironmentVariable callback."
                    )
                # Only update the seed after n_frames / n_frames_per_video epochs.
                # This ensures that we get the same random order of frames until
                # we have sampled all of them.
                rand = np.random.RandomState(
                    int(key) + self.seed + (epoch // n_different_epochs_per_seed)
                )
                indices = rand.permutation(n_frames)
                selected_frames = indices[
                    epoch * self.n_slices_per_input : (epoch + 1) * self.n_slices_per_input
                ].tolist()
                if len(selected_frames) < self.n_slices_per_input:
                    # Input cannot be evenly split, take some frames from the first batch of frames.
                    n_missing = self.n_slices_per_input - len(selected_frames)
                    selected_frames.extend(indices[0:n_missing].tolist())
            else:
                rand = random.Random(int(key) + self.seed)
                selected_frames = rand.sample(range(n_frames), k=frames_per_video)

            for frame in selected_frames:
                # Slice the fields according to the frame.
                sliced_fields = {
                    field: self.slice_data(sample[field], frame) for field in self._fields
                }
                # Leave all fields besides the sliced ones as before, augment the __key__ field to
                # include the frame number.
                sliced_fields["__key__"] = f"{key}_{frame}"
                yield {**sample, **sliced_fields}

            # Delete fields to be sure we remove all references.
            for field in self.fields:
                del sample[field]

    @property
    def fields(self):
        return self._fields

    def __call__(self, input_pipe: IterDataPipe) -> IterDataPipe:
        output_pipe = input_pipe.then(self.sample_frames_using_key)
        if self.shuffle_buffer_size > 1:
            output_pipe = output_pipe.shuffle(buffer_size=self.shuffle_buffer_size)
        return output_pipe

__init__

Initialize SampleSlices.

Parameters:

Name Type Description Default
n_slices_per_input int

Number of slices per input to sample. -1 indicates that all possible slices should be sampled.

 required
fields Sequence[str]

The fields that should be considered video data and thus sliced according to the frame sampling during training.

 required
dim int

The dimension along which to slice the tensors.

 0
seed int

Random number generator seed to deterministic sampling during evaluation.

 39480234
per_epoch bool

Sampling of frames over epochs, this ensures that after n_frames / n_frames_per_video epochs all frames have been seen at least once. In the case of uneven division, some frames will be seen more than once.

 False
shuffle_buffer_size int

Size of shuffle buffer used during training. An additional shuffling step ensures each batch contains a diverse set of images and not only images from the same video.

 1000
Source code in ocl/transforms.py 
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
def __init__(
    self,
    n_slices_per_input: int,
    fields: Sequence[str],
    dim: int = 0,
    seed: int = 39480234,
    per_epoch: bool = False,
    shuffle_buffer_size: int = 1000,
):
    """Initialize SampleSlices.

    Args:
        n_slices_per_input: Number of slices per input to sample. -1 indicates that all possible
            slices should be sampled.
        fields: The fields that should be considered video data and thus sliced
            according to the frame sampling during training.
        dim: The dimension along which to slice the tensors.
        seed: Random number generator seed to deterministic sampling during evaluation.
        per_epoch: Sampling of frames over epochs, this ensures that after
            n_frames / n_frames_per_video epochs all frames have been seen at least once.
            In the case of uneven division, some frames will be seen more than once.
        shuffle_buffer_size: Size of shuffle buffer used during training. An additional
            shuffling step ensures each batch contains a diverse set of images and not only
            images from the same video.
    """
    self.n_slices_per_input = n_slices_per_input
    self._fields = tuple(fields)
    self.dim = dim
    self.seed = seed
    self.per_epoch = per_epoch
    self.shuffle_buffer_size = shuffle_buffer_size
    self.is_batch_transform = False

slice_data

Small utility method to slice a numpy array along a specified axis.

Source code in ocl/transforms.py 
340
341
342
343
344
345
def slice_data(self, data, index: int):
    """Small utility method to slice a numpy array along a specified axis."""
    n_dims_before = self.dim
    n_dims_after = data.ndim - 1 - self.dim
    slices = (slice(None),) * n_dims_before + (index,) + (slice(None),) * n_dims_after
    return data[slices]

sample_frames_using_key

Sample frames deterministically from generator of videos using the key field.

Source code in ocl/transforms.py 
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
def sample_frames_using_key(self, data):
    """Sample frames deterministically from generator of videos using the __key__ field."""
    for sample in data:
        # Initialize random number generator dependent on instance key. This should make the
        # sampling process deterministic, which is useful when sampling frames for the
        # validation/test data.
        key = sample["__key__"]
        # TODO (hornmax): We assume all fields to have the same size. I do not want to check
        # this here as it seems a bit verbose.
        n_frames = sample[self._fields[0]].shape[self.dim]
        frames_per_video = self.n_slices_per_input if self.n_slices_per_input != -1 else n_frames

        if self.per_epoch and self.n_slices_per_input != -1:
            n_different_epochs_per_seed = int(math.ceil(n_frames / frames_per_video))
            try:
                epoch = int(os.environ["EPOCH"])
            except KeyError:
                raise RuntimeError(
                    "Using SampleSlices with per_epoch=True "
                    "requires environment variable `EPOCH` to be set. "
                    "You might need to add the SetEpochEnvironmentVariable callback."
                )
            # Only update the seed after n_frames / n_frames_per_video epochs.
            # This ensures that we get the same random order of frames until
            # we have sampled all of them.
            rand = np.random.RandomState(
                int(key) + self.seed + (epoch // n_different_epochs_per_seed)
            )
            indices = rand.permutation(n_frames)
            selected_frames = indices[
                epoch * self.n_slices_per_input : (epoch + 1) * self.n_slices_per_input
            ].tolist()
            if len(selected_frames) < self.n_slices_per_input:
                # Input cannot be evenly split, take some frames from the first batch of frames.
                n_missing = self.n_slices_per_input - len(selected_frames)
                selected_frames.extend(indices[0:n_missing].tolist())
        else:
            rand = random.Random(int(key) + self.seed)
            selected_frames = rand.sample(range(n_frames), k=frames_per_video)

        for frame in selected_frames:
            # Slice the fields according to the frame.
            sliced_fields = {
                field: self.slice_data(sample[field], frame) for field in self._fields
            }
            # Leave all fields besides the sliced ones as before, augment the __key__ field to
            # include the frame number.
            sliced_fields["__key__"] = f"{key}_{frame}"
            yield {**sample, **sliced_fields}

        # Delete fields to be sure we remove all references.
        for field in self.fields:
            del sample[field]

SplitConsecutive

Bases: Transform

Source code in ocl/transforms.py 
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
class SplitConsecutive(Transform):
    def __init__(
        self,
        n_consecutive: int,
        fields: Sequence[str],
        dim: int = 0,
        shuffle_buffer_size: int = 1000,
        drop_last: bool = True,
    ):
        self.n_consecutive = n_consecutive
        self._fields = tuple(fields)
        self.dim = dim
        self.shuffle_buffer_size = shuffle_buffer_size
        self.drop_last = drop_last
        self.is_batch_transform = False

    @property
    def fields(self):
        return self._fields

    def split_to_consecutive_frames(self, data):
        """Sample frames deterministically from generator of videos using the __key__ field."""
        for sample in data:
            key = sample["__key__"]
            n_frames = sample[self._fields[0]].shape[self.dim]

            splitted_fields = [
                np.array_split(
                    sample[field],
                    range(self.n_consecutive, n_frames, self.n_consecutive),
                    axis=self.dim,
                )
                for field in self._fields
            ]

            for i, slices in enumerate(zip(*splitted_fields)):
                if self.drop_last and slices[0].shape[self.dim] < self.n_consecutive:
                    # Last slice of not equally divisible input, discard.
                    continue

                sliced_fields = dict(zip(self._fields, slices))
                sliced_fields["__key__"] = f"{key}_{i}"
                yield {**sample, **sliced_fields}

    def __call__(self, input_pipe: IterDataPipe) -> IterDataPipe:
        output_pipe = input_pipe.then(self.split_to_consecutive_frames)
        if self.shuffle_buffer_size > 1:
            output_pipe = output_pipe.shuffle(buffer_size=self.shuffle_buffer_size)
        return output_pipe

split_to_consecutive_frames

Sample frames deterministically from generator of videos using the key field.

Source code in ocl/transforms.py 
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
def split_to_consecutive_frames(self, data):
    """Sample frames deterministically from generator of videos using the __key__ field."""
    for sample in data:
        key = sample["__key__"]
        n_frames = sample[self._fields[0]].shape[self.dim]

        splitted_fields = [
            np.array_split(
                sample[field],
                range(self.n_consecutive, n_frames, self.n_consecutive),
                axis=self.dim,
            )
            for field in self._fields
        ]

        for i, slices in enumerate(zip(*splitted_fields)):
            if self.drop_last and slices[0].shape[self.dim] < self.n_consecutive:
                # Last slice of not equally divisible input, discard.
                continue

            sliced_fields = dict(zip(self._fields, slices))
            sliced_fields["__key__"] = f"{key}_{i}"
            yield {**sample, **sliced_fields}

SampleConsecutive

Bases: Transform

Select a random consecutive subsequence of frames in a strided manner.

Given a sequence of [1, 2, 3, 4, 5, 6, 7, 8, 9] this will return one of [1, 2, 3] [4, 5, 6] [7, 8, 9].

Source code in ocl/transforms.py 
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
class SampleConsecutive(Transform):
    """Select a random consecutive subsequence of frames in a strided manner.

    Given a sequence of [1, 2, 3, 4, 5, 6, 7, 8, 9] this will return one of
    [1, 2, 3] [4, 5, 6] [7, 8, 9].
    """

    def __init__(
        self,
        n_consecutive: int,
        fields: Sequence[str],
        dim: int = 0,
    ):
        self.n_consecutive = n_consecutive
        self._fields = tuple(fields)
        self.dim = dim
        self._random = None
        self.is_batch_transform = False

    @property
    def fields(self):
        return self._fields

    @property
    def random(self):
        if not self._random:
            worker_info = torch.utils.data.get_worker_info()
            if worker_info:
                self._random = random.Random(worker_info.seed)
            else:
                self._random = random.Random(torch.initial_seed())

        return self._random

    def split_to_consecutive_frames(self, sample):
        """Sample frames deterministically from generator of videos using the __key__ field."""
        key = sample["__key__"]
        n_frames = sample[self._fields[0]].shape[self.dim]

        splitted_fields = [
            np.array_split(
                sample[field],
                range(self.n_consecutive, n_frames, self.n_consecutive),
                axis=self.dim,
            )
            for field in self._fields
        ]

        n_fragments = len(splitted_fields[0])

        if len(splitted_fields[0][-1] < self.n_consecutive):
            # Discard last fragment if too short.
            n_fragments -= 1

        fragment_id = self.random.randint(0, n_fragments - 1)
        sliced_fields: Dict[str, Any] = {
            field_name: splitted_field[fragment_id]
            for field_name, splitted_field in zip(self._fields, splitted_fields)
        }
        sliced_fields["__key__"] = f"{key}_{fragment_id}"
        return {**sample, **sliced_fields}

    def __call__(self, input_pipe: IterDataPipe) -> IterDataPipe:
        return input_pipe.map(self.split_to_consecutive_frames)

split_to_consecutive_frames

Sample frames deterministically from generator of videos using the key field.

Source code in ocl/transforms.py 
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
def split_to_consecutive_frames(self, sample):
    """Sample frames deterministically from generator of videos using the __key__ field."""
    key = sample["__key__"]
    n_frames = sample[self._fields[0]].shape[self.dim]

    splitted_fields = [
        np.array_split(
            sample[field],
            range(self.n_consecutive, n_frames, self.n_consecutive),
            axis=self.dim,
        )
        for field in self._fields
    ]

    n_fragments = len(splitted_fields[0])

    if len(splitted_fields[0][-1] < self.n_consecutive):
        # Discard last fragment if too short.
        n_fragments -= 1

    fragment_id = self.random.randint(0, n_fragments - 1)
    sliced_fields: Dict[str, Any] = {
        field_name: splitted_field[fragment_id]
        for field_name, splitted_field in zip(self._fields, splitted_fields)
    }
    sliced_fields["__key__"] = f"{key}_{fragment_id}"
    return {**sample, **sliced_fields}

VideoDecoder

Bases: Transform

Video decoder based on Decord.

Video decoding is implemented as a preprocessing transform instead of a part of the decoding mechanics as this allows sparse decoding if we only require parts of the input video.

Source code in ocl/transforms.py 
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
class VideoDecoder(Transform):
    """Video decoder based on Decord.

    Video decoding is implemented as a preprocessing transform instead of a
    part of the decoding mechanics as this allows sparse decoding if we
    only require parts of the input video.
    """

    def __init__(
        self,
        fields: Sequence[str],
        stride: int = 1,
        split_extension: bool = True,
        video_reader_kwargs: Optional[Dict[str, Any]] = None,
    ):
        """Video decoder based on decord.

        It will decode the whole video into a single tensor and can be used with other downstream
        processing plugins.

        Args:
            fields (Sequence[str]): The field of the input dictionary containing the video bytes.
            stride (int): Downsample frames by using striding. Default: 1
            split_extension (bool): Split the extension off the field name.
            video_reader_kwargs (Dict[str, Any]): Arguments to decord.VideoReader.
        """
        self._fields = tuple(fields)
        self.stride = stride
        self.split_extension = split_extension
        self.video_reader_kwargs = video_reader_kwargs if video_reader_kwargs else {}
        self.is_batch_transform = False

    @property
    def fields(self):
        return self._fields

    def _chunk_iterator(
        self, vrs: Dict[str, decord.VideoReader], key: str, inputs: Dict[str, Any]
    ) -> Iterable[Tuple[str, Dict]]:
        """Iterate over chunks of the video.

        For the video decoder we simply return a single chunk containing the whole video, subclasses
        might override this method though.

        Returns:
            str: Derived key which combines chunk and video key.
            torch.Tensor: Chunk of video data.
            Dict: Additional information, for example which frames where selected.  This might be of
                relevance when different modalities need to be sliced in a similar fashion as the
                video input.
        """
        # Get whole video.
        indices = list(range(0, len(next(iter(vrs.values()))), self.stride))
        videos = {output_name: vr.get_batch(indices) for output_name, vr in vrs.items()}
        yield (key, {**videos, "decoded_indices": indices})

    def video_decoding(self, input_generator):
        for input_data in input_generator:
            key = input_data["__key__"]
            vrs = {}
            for field in self.fields:
                video_file: StreamWrapper = input_data[field]
                # Remove the input field
                del input_data[field]
                if self.split_extension:
                    output_field, _ = os.path.splitext(field)
                else:
                    output_field = field

                vr = decord.VideoReader(video_file.file_obj, **self.video_reader_kwargs)
                video_file.autoclose()
                vrs[output_field] = vr

            for derived_key, videos_and_additional_info in self._chunk_iterator(
                vrs, key, input_data
            ):
                yield {
                    **input_data,
                    "__key__": derived_key,
                    **videos_and_additional_info,
                }

    def __call__(self, input_pipe: IterDataPipe) -> IterDataPipe:
        return input_pipe.then(self.video_decoding)

__init__

Video decoder based on decord.

It will decode the whole video into a single tensor and can be used with other downstream processing plugins.

Parameters:

Name Type Description Default
fields Sequence[str]

The field of the input dictionary containing the video bytes.

 required
stride int

Downsample frames by using striding. Default: 1

 1
split_extension bool

Split the extension off the field name.

 True
video_reader_kwargs Dict[str, Any]

Arguments to decord.VideoReader.

 None
Source code in ocl/transforms.py 
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
def __init__(
    self,
    fields: Sequence[str],
    stride: int = 1,
    split_extension: bool = True,
    video_reader_kwargs: Optional[Dict[str, Any]] = None,
):
    """Video decoder based on decord.

    It will decode the whole video into a single tensor and can be used with other downstream
    processing plugins.

    Args:
        fields (Sequence[str]): The field of the input dictionary containing the video bytes.
        stride (int): Downsample frames by using striding. Default: 1
        split_extension (bool): Split the extension off the field name.
        video_reader_kwargs (Dict[str, Any]): Arguments to decord.VideoReader.
    """
    self._fields = tuple(fields)
    self.stride = stride
    self.split_extension = split_extension
    self.video_reader_kwargs = video_reader_kwargs if video_reader_kwargs else {}
    self.is_batch_transform = False

DecodeRandomWindow

Bases: VideoDecoder

Decode a random window of the video.

Source code in ocl/transforms.py 
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
class DecodeRandomWindow(VideoDecoder):
    """Decode a random window of the video."""

    def __init__(self, n_consecutive_frames: int, **video_decoder_args):
        self.n_consecutive_frames = n_consecutive_frames
        self._random = None
        super().__init__(**video_decoder_args)

    @property
    def random(self):
        if not self._random:
            worker_info = torch.utils.data.get_worker_info()
            if worker_info:
                self._random = random.Random(worker_info.seed)
            else:
                self._random = random.Random(torch.initial_seed())

        return self._random

    def _chunk_iterator(
        self, vrs: Mapping[str, decord.VideoReader], key: str, inputs: Dict[str, Any]
    ) -> Iterable[Tuple[str, Dict]]:
        """Iterate over chunks of the video.

        Returns:
            str: Derived key which combines chunk and video key.
            torch.Tensor: Chunk of video data.
            Dict: Additional information, for example which frames where selected.  This might be of
                relevance when different modalities need to be sliced in a similar fashion as the
                video input.
        """
        n_frames = len(next(iter(vrs.values())))
        assert self.n_consecutive_frames * self.stride < n_frames
        starting_index = self.random.randint(0, n_frames - self.n_consecutive_frames * self.stride)
        indices = list(
            range(
                starting_index, starting_index + self.n_consecutive_frames * self.stride, self.stride
            )
        )
        videos = {output_field: vr.get_batch(indices) for output_field, vr in vrs.items()}
        yield f"{key}_{starting_index}", {**videos, "decoded_indices": indices}

DecodeRandomStridedWindow

Bases: DecodeRandomWindow

Decode random strided segment of input video.

Source code in ocl/transforms.py 
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
class DecodeRandomStridedWindow(DecodeRandomWindow):
    """Decode random strided segment of input video."""

    def _chunk_iterator(
        self, vrs: Mapping[str, decord.VideoReader], key: str, inputs: Dict[str, Any]
    ) -> Iterable[Tuple[str, Dict]]:
        """Iterate over chunks of the video.

        For the video decoder we simply return a single chunk containing the whole video, subclasses
        might override this method though.

        Returns:
            str: Derived key which combines chunk and video key.
            torch.Tensor: Chunk of video data.
            Dict: Additional information, for example which frames where selected.  This might be of
                relevance when different modalities need to be sliced in a similar fashion as the
                video input.
        """
        n_frames = len(next(iter(vrs.values())))
        segment_indices = list(range(0, n_frames + 1, self.n_consecutive_frames * self.stride))
        segment_index = self.random.randint(0, len(segment_indices) - 2)
        indices = list(
            range(segment_indices[segment_index], segment_indices[segment_index + 1], self.stride)
        )
        videos = {output_field: vr.get_batch(indices) for output_field, vr in vrs.items()}
        yield f"{key}_{segment_index}", {**videos, "decoded_indices": indices}

SpatialSlidingWindow

Bases: Transform

Split image data spatially by sliding a window across.

Source code in ocl/transforms.py 
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
class SpatialSlidingWindow(Transform):
    """Split image data spatially by sliding a window across."""

    def __init__(
        self,
        window_size: Tuple[int, int],
        stride: Tuple[int, int],
        padding: Tuple[int, int, int, int],
        fields: Sequence[str],
        expected_n_windows: Optional[int] = None,
    ):
        self.window_size = window_size
        self.stride = stride
        self.padding = padding
        self.expected_n_windows = expected_n_windows
        self._fields = tuple(fields)
        self.is_batch_transform = False

    @property
    def fields(self):
        return self._fields

    @staticmethod
    def pad(elem, padding):
        if elem.shape[-1] != 1 and elem.shape[-1] != 3:
            elem = elem[..., None]
        orig_height = elem.shape[-3]
        orig_width = elem.shape[-2]

        p_left, p_top, p_right, p_bottom = padding
        height = orig_height + p_top + p_bottom
        width = orig_width + p_left + p_right

        padded_shape = list(elem.shape[:-3]) + [height, width, elem.shape[-1]]
        elem_padded = np.zeros_like(elem, shape=padded_shape)
        elem_padded[..., p_top : p_top + orig_height, p_left : p_left + orig_width, :] = elem

        return elem_padded

    def sliding_window(self, data):
        for sample in data:
            key = sample["__key__"]

            window_x, window_y = self.window_size
            stride_x, stride_y = self.stride
            padded_elems = {key: self.pad(sample[key], self.padding) for key in self._fields}

            n_windows = 0
            x = 0
            y = 0
            while True:
                shape = None
                windowed_fields = {}
                for key in self._fields:
                    elem_padded = padded_elems[key]
                    if shape is None:
                        shape = elem_padded.shape
                    else:
                        if shape[-3:-1] != elem_padded.shape[-3:-1]:
                            raise ValueError("Element height, width after padding do not match")
                    windowed_fields[key] = elem_padded[..., y : y + window_y, x : x + window_x, :]

                    window_height, window_width = windowed_fields[key].shape[-3:-1]
                    assert (
                        window_y == window_height and window_x == window_width
                    ), f"Expected {window_y}, {window_x}, received {window_height}, {window_width}"

                windowed_fields["__key__"] = f"{key}_{x - self.padding[0]}_{y - self.padding[1]}"
                yield {**sample, **windowed_fields}
                n_windows += 1

                x += stride_x
                if x >= shape[-2]:
                    y += stride_y
                    x = 0
                if y >= shape[-3]:
                    break

            if self.expected_n_windows is not None and self.expected_n_windows != n_windows:
                raise ValueError(f"Expected {self.expected_n_windows} windows, but got {n_windows}")

    def __call__(self, input_pipe: IterDataPipe):
        return input_pipe.then(self.sliding_window)