csv_dataset.py 16.9 KB
Newer Older
1
2
3
4
5
6
#!/usr/bin/env python
# vim: set fileencoding=utf-8 :


import os
from bob.pipelines import Sample, DelayedSample, SampleSet
7
from bob.db.base.utils import check_parameters_for_validity
8
9
10
11
import csv
import bob.io.base
import functools
from abc import ABCMeta, abstractmethod
12
13
import numpy as np
import itertools
14
import logging
15

16
logger = logging.getLogger(__name__)
17

Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
18
class CSVBaseSampleLoader(metaclass=ABCMeta):
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
    """
    Convert CSV files in the format below to either a list of
    :any:`bob.pipelines.DelayedSample` or :any:`bob.pipelines.SampleSet`

    .. code-block:: text

       PATH,SUBJECT
       path_1,subject_1
       path_2,subject_2
       path_i,subject_j
       ...

    .. note::
       This class should be extended

    Parameters
    ----------

        data_loader:
            A python function that can be called parameterlessly, to load the
            sample in question from whatever medium

        extension:
            The file extension

    """

Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
46
    def __init__(self, data_loader, dataset_original_directory="", extension=""):
47
48
        self.data_loader = data_loader
        self.extension = extension
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
49
        self.dataset_original_directory = dataset_original_directory
50
51
52
53
54
55
56
57
58
59

    @abstractmethod
    def __call__(self, filename):
        pass

    @abstractmethod
    def convert_row_to_sample(self, row, header):
        pass

    @abstractmethod
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
60
61
62
    def convert_samples_to_samplesets(
        self, samples, group_by_subject=True, references=None
    ):
63
64
65
        pass


Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
66
class CSVToSampleLoader(CSVBaseSampleLoader):
67
68
69
70
71
    """
    Simple mechanism to convert CSV files in the format below to either a list of
    :any:`bob.pipelines.DelayedSample` or :any:`bob.pipelines.SampleSet`
    """

72
73
74
75
76
77
78
    def check_header(self, header):
        """
        A header should have at least "SUBJECT" AND "PATH"
        """
        header = [h.lower() for h in header]
        if not "subject" in header:
            raise ValueError("The field `subject` is not available in your dataset.")
79

80
81
82
83
        if not "path" in header:
            raise ValueError("The field `path` is not available in your dataset.")

    def __call__(self, filename):
84
85
86
87
88

        with open(filename) as cf:
            reader = csv.reader(cf)
            header = next(reader)

89
            self.check_header(header)
90
91
92
93
94
            return [self.convert_row_to_sample(row, header) for row in reader]

    def convert_row_to_sample(self, row, header):
        path = row[0]
        subject = row[1]
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
95
        kwargs = dict([[h, r] for h, r in zip(header[2:], row[2:])])
96
        return DelayedSample(
97
98
99
100
            functools.partial(
                self.data_loader,
                os.path.join(self.dataset_original_directory, path + self.extension),
            ),
101
102
103
104
105
            key=path,
            subject=subject,
            **kwargs,
        )

Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
106
107
108
    def convert_samples_to_samplesets(
        self, samples, group_by_subject=True, references=None
    ):
109
110
111
112
113
114
115
        if group_by_subject:

            # Grouping sample sets
            sample_sets = dict()
            for s in samples:
                if s.subject not in sample_sets:
                    sample_sets[s.subject] = SampleSet(
116
                        [s], parent=s, references=references
117
                    )
118
119
                else:
                    sample_sets[s.subject].append(s)
120
121
122
            return list(sample_sets.values())

        else:
123
            return [
124
                SampleSet([s], parent=s, references=references)
125
126
                for s in samples
            ]
127
128
129
130


class CSVDatasetDevEval:
    """
131
132
    Generic filelist dataset for :any:` bob.bio.base.pipelines.vanilla_biometrics.VanillaBiometricsPipeline` pipeline.
    Check :any:`vanilla_biometrics_features` for more details about the Vanilla Biometrics Dataset
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
    interface.

    To create a new dataset, you need to provide a directory structure similar to the one below:

    .. code-block:: text

       my_dataset/
       my_dataset/my_protocol/
       my_dataset/my_protocol/train.csv
       my_dataset/my_protocol/train.csv/dev_enroll.csv
       my_dataset/my_protocol/train.csv/dev_probe.csv
       my_dataset/my_protocol/train.csv/eval_enroll.csv
       my_dataset/my_protocol/train.csv/eval_probe.csv
       ...


    In the above directory structure, inside of `my_dataset` should contain the directories with all
    evaluation protocols this dataset might have.
    Inside of the `my_protocol` directory should contain at least two csv files:

     - dev_enroll.csv
     - dev_probe.csv


    Those csv files should contain in each row i-) the path to raw data and ii-) the subject label
158
159
    for enrollment (:any:`bob.bio.base.pipelines.vanilla_biometrics.Database.references`) and
    probing (:any:`bob.bio.base.pipelines.vanilla_biometrics.Database.probes`).
160
161
162
163
164
165
166
167
168
169
    The structure of each CSV file should be as below:

    .. code-block:: text

       PATH,SUBJECT
       path_1,subject_1
       path_2,subject_2
       path_i,subject_j
       ...

170

171
172
173
174
175
176
177
178
179
180
181
182
183
184
    You might want to ship metadata within your Samples (e.g gender, age, annotation, ...)
    To do so is simple, just do as below:

    .. code-block:: text

       PATH,SUBJECT,METADATA_1,METADATA_2,METADATA_k
       path_1,subject_1,A,B,C
       path_2,subject_2,A,B,1
       path_i,subject_j,2,3,4
       ...


    The files `my_dataset/my_protocol/train.csv/eval_enroll.csv` and `my_dataset/my_protocol/train.csv/eval_probe.csv`
    are optional and it is used in case a protocol contains data for evaluation.
185

186
    Finally, the content of the file `my_dataset/my_protocol/train.csv` is used in the case a protocol
187
    contains data for training (`bob.bio.base.pipelines.vanilla_biometrics.Database.background_model_samples`)
188
189
190
191
192
193
194

    Parameters
    ----------

        dataset_path: str
          Absolute path of the dataset protocol description

Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
195
        protocol_na,e: str
196
197
          The name of the protocol

198
199
        csv_to_sample_loader: :any:`bob.bio.base.database.CSVBaseSampleLoader`
            Base class that whose objective is to generate :any:`bob.pipelines.Sample`
200
            and/or :any:`bob.pipelines.SampleSet` from csv rows
201
202
203
204
205

    """

    def __init__(
        self,
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
206
207
        dataset_protocol_path,
        protocol_name,
208
        csv_to_sample_loader=CSVToSampleLoader(
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
209
            data_loader=bob.io.base.load, dataset_original_directory="", extension=""
210
211
212
213
        ),
    ):
        def get_paths():

Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
214
215
            if not os.path.exists(dataset_protocol_path):
                raise ValueError(f"The path `{dataset_protocol_path}` was not found")
216
217

            # TODO: Unzip file if dataset path is a zip
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
218
            protocol_path = os.path.join(dataset_protocol_path, protocol_name)
219
            if not os.path.exists(protocol_path):
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
220
                raise ValueError(f"The protocol `{protocol_name}` was not found")
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

            train_csv = os.path.join(protocol_path, "train.csv")
            dev_enroll_csv = os.path.join(protocol_path, "dev_enroll.csv")
            dev_probe_csv = os.path.join(protocol_path, "dev_probe.csv")
            eval_enroll_csv = os.path.join(protocol_path, "eval_enroll.csv")
            eval_probe_csv = os.path.join(protocol_path, "eval_probe.csv")

            # The minimum required is to have `dev_enroll_csv` and `dev_probe_csv`
            train_csv = train_csv if os.path.exists(train_csv) else None

            # Eval
            eval_enroll_csv = (
                eval_enroll_csv if os.path.exists(eval_enroll_csv) else None
            )
            eval_probe_csv = eval_probe_csv if os.path.exists(eval_probe_csv) else None

            # Dev
            if not os.path.exists(dev_enroll_csv):
                raise ValueError(
                    f"The file `{dev_enroll_csv}` is required and it was not found"
                )

            if not os.path.exists(dev_probe_csv):
                raise ValueError(
                    f"The file `{dev_probe_csv}` is required and it was not found"
                )

            return (
                train_csv,
                dev_enroll_csv,
                dev_probe_csv,
                eval_enroll_csv,
                eval_probe_csv,
            )

        (
            self.train_csv,
            self.dev_enroll_csv,
            self.dev_probe_csv,
            self.eval_enroll_csv,
            self.eval_probe_csv,
        ) = get_paths()

        def get_dict_cache():
            cache = dict()
            cache["train"] = None
            cache["dev_enroll_csv"] = None
            cache["dev_probe_csv"] = None
            cache["eval_enroll_csv"] = None
            cache["eval_probe_csv"] = None
            return cache

        self.cache = get_dict_cache()
        self.csv_to_sample_loader = csv_to_sample_loader

    def background_model_samples(self):

        self.cache["train"] = (
            self.csv_to_sample_loader(self.train_csv)
            if self.cache["train"] is None
            else self.cache["train"]
        )

        return self.cache["train"]

    def _get_samplesets(self, group="dev", purpose="enroll", group_by_subject=False):

Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
288
        fetching_probes = False
289
290
291
        if purpose == "enroll":
            cache_label = "dev_enroll_csv" if group == "dev" else "eval_enroll_csv"
        else:
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
292
            fetching_probes = True
293
294
295
296
297
            cache_label = "dev_probe_csv" if group == "dev" else "eval_probe_csv"

        if self.cache[cache_label] is not None:
            return self.cache[cache_label]

Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
298
299
        references = None
        if fetching_probes:
300
            references = list(set([s.subject for s in self.references(group=group)]))
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
301
302

        samples = self.csv_to_sample_loader(self.__dict__[cache_label])
303
304

        sample_sets = self.csv_to_sample_loader.convert_samples_to_samplesets(
Tiago de Freitas Pereira's avatar
Tiago de Freitas Pereira committed
305
            samples, group_by_subject=group_by_subject, references=references
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
        )

        self.cache[cache_label] = sample_sets

        return self.cache[cache_label]

    def references(self, group="dev"):
        return self._get_samplesets(
            group=group, purpose="enroll", group_by_subject=True
        )

    def probes(self, group="dev"):
        return self._get_samplesets(
            group=group, purpose="probe", group_by_subject=False
        )
321

322
323
324
325
326
327
328
    def all_samples(self, groups=None):
        """
        Reads and returns all the samples in `groups`.

        Parameters
        ----------
        groups: list or None
329
330
331
332
333
334
335
            Groups to consider ('train', 'dev', and/or 'eval'). If `None` is
            given, returns the samples from all groups.

        Returns
        -------
        samples: list
            List of :class:`bob.pipelines.Sample` objects.
336
        """
337
338
339
340
341
342
343
344
345
346
347
348
349
350
        valid_groups = ["train"]
        if self.dev_enroll_csv and self.dev_probe_csv:
            valid_groups.append("dev")
        if self.eval_enroll_csv and self.eval_probe_csv:
            valid_groups.append("eval")
        groups = check_parameters_for_validity(
            parameters=groups,
            parameter_description="groups",
            valid_parameters=valid_groups,
            default_parameters=valid_groups,
        )

        samples = []

351
        # Get train samples (background_model_samples returns a list of samples)
352
353
354
        if "train" in groups:
            samples = samples + self.background_model_samples()
            groups.remove("train")
355
356
357
358
359

        # Get enroll and probe samples
        for group in groups:
            for purpose in ("enroll", "probe"):
                label = f"{group}_{purpose}_csv"
360
                samples = samples + self.csv_to_sample_loader(self.__dict__[label])
361
362
        return samples

363
364
365

class CSVDatasetCrossValidation:
    """
366
    Generic filelist dataset for :any:`bob.bio.base.pipelines.vanilla_biometrics.VanillaBiometricsPipeline` pipeline that
367
368
    handles **CROSS VALIDATION**.

369
    Check :any:`vanilla_biometrics_features` for more details about the Vanilla Biometrics Dataset
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
    interface.


    This interface will take one `csv_file` as input and split into i-) data for training and
    ii-) data for testing.
    The data for testing will be further split in data for enrollment and data for probing.
    The input CSV file should be casted in the following format:

    .. code-block:: text

       PATH,SUBJECT
       path_1,subject_1
       path_2,subject_2
       path_i,subject_j
       ...

    Parameters
    ----------

    csv_file_name: str
      CSV file containing all the samples from your database

    random_state: int
      Pseudo-random number generator seed

    test_size: float
      Percentage of the subjects used for testing

    samples_for_enrollment: float
      Number of samples used for enrollment

401
402
    csv_to_sample_loader: :any:`bob.bio.base.database.CSVBaseSampleLoader`
        Base class that whose objective is to generate :any:`bob.pipelines.Sample`
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
        and/or :any:`bob.pipelines.SampleSet` from csv rows

    """

    def __init__(
        self,
        csv_file_name="metadata.csv",
        random_state=0,
        test_size=0.8,
        samples_for_enrollment=1,
        csv_to_sample_loader=CSVToSampleLoader(
            data_loader=bob.io.base.load, dataset_original_directory="", extension=""
        ),
    ):
        def get_dict_cache():
            cache = dict()
            cache["train"] = None
            cache["dev_enroll_csv"] = None
            cache["dev_probe_csv"] = None
            return cache

        self.random_state = random_state
        self.cache = get_dict_cache()
        self.csv_to_sample_loader = csv_to_sample_loader
        self.csv_file_name = csv_file_name
        self.samples_for_enrollment = samples_for_enrollment
        self.test_size = test_size

        if self.test_size < 0 and self.test_size > 1:
            raise ValueError(
                f"`test_size` should be between 0 and 1. {test_size} is provided"
            )

    def _do_cross_validation(self):

        # Shuffling samples by subject
        samples_by_subject = group_samples_by_subject(
            self.csv_to_sample_loader(self.csv_file_name)
        )
        subjects = list(samples_by_subject.keys())
        np.random.seed(self.random_state)
        np.random.shuffle(subjects)

        # Getting the training data
        n_samples_for_training = len(subjects) - int(self.test_size * len(subjects))
        self.cache["train"] = list(
            itertools.chain(
                *[samples_by_subject[s] for s in subjects[0:n_samples_for_training]]
            )
        )

        # Splitting enroll and probe
        self.cache["dev_enroll_csv"] = []
        self.cache["dev_probe_csv"] = []
        for s in subjects[n_samples_for_training:]:
            samples = samples_by_subject[s]
            if len(samples) < self.samples_for_enrollment:
                raise ValueError(
                    f"Not enough samples ({len(samples)}) for enrollment for the subject {s}"
                )

            # Enrollment samples
            self.cache["dev_enroll_csv"].append(
                self.csv_to_sample_loader.convert_samples_to_samplesets(
                    samples[0 : self.samples_for_enrollment]
                )[0]
            )

            self.cache[
                "dev_probe_csv"
            ] += self.csv_to_sample_loader.convert_samples_to_samplesets(
                samples[self.samples_for_enrollment :],
                group_by_subject=False,
                references=subjects[n_samples_for_training:],
            )

    def _load_from_cache(self, cache_key):
        if self.cache[cache_key] is None:
            self._do_cross_validation()
        return self.cache[cache_key]

    def background_model_samples(self):
        return self._load_from_cache("train")

    def references(self, group="dev"):
        return self._load_from_cache("dev_enroll_csv")

    def probes(self, group="dev"):
        return self._load_from_cache("dev_probe_csv")

493
494
495
496
497
498
499
    def all_samples(self, groups=None):
        """
        Reads and returns all the samples in `groups`.

        Parameters
        ----------
        groups: list or None
500
501
502
503
504
505
506
            Groups to consider ('train' and/or 'dev'). If `None` is given,
            returns the samples from all groups.

        Returns
        -------
        samples: list
            List of :class:`bob.pipelines.Sample` objects.
507
        """
508
509
510
511
512
513
514
515
516
517
        valid_groups = ["train", "dev"]
        groups = check_parameters_for_validity(
            parameters=groups,
            parameter_description="groups",
            valid_parameters=valid_groups,
            default_parameters=valid_groups,
        )

        samples = []

518
        # Get train samples (background_model_samples returns a list of samples)
519
520
521
        if "train" in groups:
            samples = samples + self.background_model_samples()
            groups.remove("train")
522
523
524

        # Get enroll and probe samples
        for group in groups:
525
526
            samples = samples + [s for s_set in self.references(group) for s in s_set]
            samples = samples + [s for s_set in self.probes(group) for s in s_set]
527
528
        return samples

529
530
531
532
533
534
535
536
537
538

def group_samples_by_subject(samples):

    # Grouping sample sets
    samples_by_subject = dict()
    for s in samples:
        if s.subject not in samples_by_subject:
            samples_by_subject[s.subject] = []
        samples_by_subject[s.subject].append(s)
    return samples_by_subject