def _create_sample_sets(raw_data, offset, references=None):
if references is None:
return [
SampleSet(
[
Sample(
s,
reference_id=str(i + offset),
key=str(uuid.uuid4()),
)
],
key=str(i + offset),
reference_id=str(i + offset),
subject_id=str(i + offset),
) for i, s in enumerate(raw_data)
]
else:
return [
SampleSet(
[
Sample(
s,
reference_id=str(i + offset),
key=str(uuid.uuid4()),
)
],
key=str(i + offset),
reference_id=str(i + offset),
subject_id=str(i + offset),
references=references,
) for i, s in enumerate(raw_data)
]
def test_extractor_fittable():
with tempfile.TemporaryDirectory() as dir_name:
extractor_file = os.path.join(dir_name, "Extractor.hdf5")
extractor = FakeExtractorFittable()
extractor_transformer = ExtractorTransformer(extractor,
model_path=extractor_file)
# Testing sample
sample_transformer = SampleWrapper(extractor_transformer)
# Fitting
training_data = np.arange(4).reshape(2, 2)
training_samples = [Sample(training_data, key="1")]
sample_transformer = sample_transformer.fit(training_samples)
test_data = [np.zeros((2, 2)), np.ones((2, 2))]
oracle = [np.zeros((2, 2)), np.ones((2, 2)) @ training_data]
test_sample = [Sample(d, key=str(i)) for i, d in enumerate(test_data)]
transformed_sample = sample_transformer.transform(test_sample)
assert assert_sample(transformed_sample, oracle)
# Testing checkpoint
checkpointing_transformer = CheckpointWrapper(
sample_transformer,
features_dir=dir_name,
load_func=extractor.read_feature,
save_func=extractor.write_feature,
)
transformed_sample = checkpointing_transformer.transform(test_sample)
assert assert_sample(transformed_sample, oracle)
assert assert_checkpoints(transformed_sample, dir_name)
def test_sampleset_collection():
n_samples = 10
X = np.ones(shape=(n_samples, 2), dtype=int)
sampleset = SampleSet(
[Sample(data, key=str(i)) for i, data in enumerate(X)], key="1")
assert len(sampleset) == n_samples
# Testing insert
sample = Sample(X, key=100)
sampleset.insert(1, sample)
assert len(sampleset) == n_samples + 1
# Testing delete
del sampleset[0]
assert len(sampleset) == n_samples
# Testing set
sampleset[0] = copy.deepcopy(sample)
# Testing iterator
for i in sampleset:
assert isinstance(i, Sample)
def _load(path):
return pickle.loads(open(path, "rb").read())
# Testing delayed sampleset
with tempfile.TemporaryDirectory() as dir_name:
samples = [Sample(data, key=str(i)) for i, data in enumerate(X)]
filename = os.path.join(dir_name, "samples.pkl")
with open(filename, "wb") as f:
f.write(pickle.dumps(samples))
sampleset = DelayedSampleSet(functools.partial(_load, filename), key=1)
assert len(sampleset) == n_samples
assert sampleset.samples == samples
# Testing delayed sampleset cached
with tempfile.TemporaryDirectory() as dir_name:
samples = [Sample(data, key=str(i)) for i, data in enumerate(X)]
filename = os.path.join(dir_name, "samples.pkl")
with open(filename, "wb") as f:
f.write(pickle.dumps(samples))
sampleset = DelayedSampleSetCached(functools.partial(_load, filename),
key=1)
assert len(sampleset) == n_samples
assert sampleset.samples == samples
def fit(self, t_scores, y=None):
# TODO: THIS IS SUPER INNEFICIENT, BUT
# IT'S THE MOST READABLE SOLUTION
# Stacking scores by biometric reference
self.t_stats = dict()
for sset in t_scores:
self.t_stats[sset.reference_id] = Sample([s.data for s in sset],
parent=sset)
# Now computing the statistics in place
for key in self.t_stats:
data = self.t_stats[key].data
# Selecting the top scores
if self.top_norm:
# Sorting in ascending order
data = -np.sort(-data)
proportion = int(
np.floor(len(data) * self.top_norm_score_fraction))
data = data[0:proportion]
self.t_stats[key].mu = np.mean(self.t_stats[key].data)
self.t_stats[key].std = np.std(self.t_stats[key].data)
# self._z_stats[key].std = legacy_std(
# self._z_stats[key].mu, self._z_stats[key].data
# )
self.t_stats[key].data = []
return self
def test_mod_4hz():
"""Loading and running the mod-4hz annotator."""
# Test setup and config
annotator = bob.bio.base.load_resource("mod-4hz", "annotator")
assert isinstance(annotator, bob.bio.spear.annotator.Mod_4Hz)
# Read input
rate, wav = _wav()
# Test the VAD annotator
annotator = bob.bio.spear.annotator.Mod_4Hz()
_compare(
annotator.transform_one(wav, sample_rate=rate),
pkg_resources.resource_filename(
"bob.bio.spear.test", "data/vad_mod_4hz.hdf5"
),
)
# Test the processing of Sample objects and tags of annotator transformer
wrapped_annotator = wrap(["sample"], annotator)
samples = [Sample(data=wav, rate=rate)]
# Attribute `rate` should be passed as `sample_rate` argument of transform (tags)
result = wrapped_annotator.transform(samples)
# Annotations should be in attribute `annotations` of result samples (tags)
_compare(
result[0].annotations,
pkg_resources.resource_filename(
"bob.bio.spear.test", "data/vad_mod_4hz.hdf5"
),
)
def test_preprocessor():
preprocessor = FakePreprocesor()
preprocessor_transformer = PreprocessorTransformer(preprocessor)
# Testing sample
transform_extra_arguments = [("annotations", "annotations")]
sample_transformer = SampleWrapper(preprocessor_transformer,
transform_extra_arguments)
data = np.zeros((2, 2))
oracle = [np.ones((2, 2))]
annotations = 1
sample = [Sample(data, key="1", annotations=annotations)]
transformed_sample = sample_transformer.transform(sample)
assert assert_sample(transformed_sample, oracle)
# Testing checkpoint
with tempfile.TemporaryDirectory() as dir_name:
checkpointing_transformer = CheckpointWrapper(
sample_transformer,
features_dir=dir_name,
load_func=preprocessor.read_data,
save_func=preprocessor.write_data,
)
transformed_sample = checkpointing_transformer.transform(sample)
assert assert_sample(transformed_sample, oracle)
assert assert_checkpoints(transformed_sample, dir_name)
def test_extractor():
extractor = FakeExtractor()
extractor_transformer = ExtractorTransformer(extractor)
# Testing sample
sample_transformer = SampleWrapper(extractor_transformer)
data = np.zeros((2, 2))
oracle = [np.zeros((1, 4))]
sample = [Sample(data, key="1")]
transformed_sample = sample_transformer.transform(sample)
assert assert_sample(transformed_sample, oracle)
# Testing checkpoint
with tempfile.TemporaryDirectory() as dir_name:
checkpointing_transformer = CheckpointWrapper(
sample_transformer,
features_dir=dir_name,
load_func=extractor.read_feature,
save_func=extractor.write_feature,
)
transformed_sample = checkpointing_transformer.transform(sample)
assert assert_sample(transformed_sample, oracle)
assert assert_checkpoints(transformed_sample, dir_name)
def test_sample_hdf5():
n_samples = 10
X = np.ones(shape=(n_samples, 2), dtype=int)
samples = [
Sample(data, key=str(i), subject="Subject") for i, data in enumerate(X)
]
with tempfile.TemporaryDirectory() as dir_name:
# Single sample
filename = os.path.join(dir_name, "sample.hdf5")
with h5py.File(filename, "w", driver="core") as hdf5:
sample_to_hdf5(samples[0], hdf5)
with h5py.File(filename, "r") as hdf5:
sample = hdf5_to_sample(hdf5)
assert sample == samples[0]
# List of samples
filename = os.path.join(dir_name, "samples.hdf5")
with h5py.File(filename, "w", driver="core") as hdf5:
sample_to_hdf5(samples, hdf5)
with h5py.File(filename, "r") as hdf5:
samples_deserialized = hdf5_to_sample(hdf5)
compare = [a == b for a, b in zip(samples_deserialized, samples)]
assert np.sum(compare) == 10
def _transform_samples(X, stats):
scores = []
for no_normed_score in X:
score = (no_normed_score.data - stats.mu) / stats.std
t_score = Sample(score, parent=no_normed_score)
scores.append(t_score)
return scores
def get_fake_samples_for_training():
data = np.random.rand(10, 3, 400, 400)
annotations = {"reye": (131, 176), "leye": (222, 170)}
return [
Sample(x, key=str(i), reference_id=str(i), annotations=annotations)
for i, x in enumerate(data)
]
def _transform_samples(X):
scores = []
for no_normed_score in X:
score = (no_normed_score.data -
self.z_stats[no_normed_score.reference_id].mu
) / self.z_stats[no_normed_score.reference_id].std
z_score = Sample(score, parent=no_normed_score)
scores.append(z_score)
return scores
def test_resample():
"""Resample using the transformer."""
audio_path = resource_filename("bob.bio.spear.test", "data/sample.wav")
audio_n_samples = 77760
audio_sample_rate = 16000
sample = Sample(data=audio_path, channel=None, rate=audio_sample_rate)
pipeline = make_pipeline(
PathToAudio(), wrap(["sample"], Resample(audio_sample_rate // 2)))
results = pipeline.transform([sample])[0]
assert results.data.shape == (audio_n_samples // 2, ), results.data.shape
def _create_random_2dsamples(self, n_samples, offset, dim):
return [
Sample(
np.random.rand(dim, dim),
key=str(uuid.uuid4()),
annotations=1,
reference_id=str(i),
subject_id=str(i),
)
for i in range(offset, offset + n_samples)
]
def create_templates_from_samplesets(self, list_of_samplesets, enroll):
"""Creates enroll or probe templates from multiple SampleSets.
Parameters
----------
list_of_samplesets : list
A list (length N) of SampleSets.
enroll : bool
If True, the SampleSets are for enrollment. If False, the SampleSets
are for probe.
Returns
-------
templates : list
A list of Samples which has the same length as ``list_of_samplesets``.
Each Sample contains a template.
"""
logger.debug(
f"{_frmt(self)}.create_templates_from_samplesets(... enroll={enroll})"
)
# create templates from .data attribute of samples inside sample_sets
list_of_feature_sets = []
for sampleset in list_of_samplesets:
data = [s.data for s in sampleset.samples]
valid_data = [d for d in data if d is not None]
if len(data) != len(valid_data):
logger.warning(
f"Removed {len(data)-len(valid_data)} invalid enrollment samples."
)
if not valid_data and enroll:
# we do not support failure to enroll cases currently
raise NotImplementedError(
f"None of the enrollment samples were valid for {sampleset}."
)
list_of_feature_sets.append(valid_data)
templates = self.create_templates(list_of_feature_sets, enroll)
expected_size = len(list_of_samplesets)
assert len(templates) == expected_size, (
"The number of (%s) templates (%d) created by the algorithm does not match "
"the number of sample sets (%d)" % (
"enroll" if enroll else "probe",
len(templates),
expected_size,
))
# return a list of Samples (one per template)
templates = [
Sample(t, parent=sampleset)
for t, sampleset in zip(templates, list_of_samplesets)
]
return templates
def test_path_to_audio():
"""Tries to load the audio data from a file."""
audio_path = resource_filename("bob.bio.spear.test", "data/sample.wav")
audio_n_samples = 77760
audio_sample_rate = 16000
sample = Sample(data=audio_path)
transformer = PathToAudio()
results = transformer.transform([sample])[0]
assert results.rate == audio_sample_rate, results.rate
assert isinstance(results.data, np.ndarray)
assert results.data.shape == (audio_n_samples, ), results.data.shape
assert results.data.dtype == np.float32, results.data.dtype
# Force a different sample rate
sample = Sample(data=audio_path)
transformer = PathToAudio(forced_sr=audio_sample_rate // 2)
results = transformer.transform([sample])[0]
assert results.rate == audio_sample_rate // 2, results.rate
assert isinstance(results.data, np.ndarray)
assert results.data.shape == (audio_n_samples // 2, ), results.data.shape
def test_delayed_samples():
def load_data():
return 0
def load_annot():
return "annotation"
def load_annot_variant():
return "annotation_variant"
delayed_attr_read = False
def load_check():
nonlocal delayed_attr_read
delayed_attr_read = True
return "delayed_attr_data"
delayed_sample = DelayedSample(load_data,
delayed_attributes=dict(annot=load_annot))
assert delayed_sample.data == 0, delayed_sample.data
assert delayed_sample.annot == "annotation", delayed_sample.annot
child_sample = Sample(1, parent=delayed_sample)
assert child_sample.data == 1, child_sample.data
assert child_sample.annot == "annotation", child_sample.annot
assert child_sample.__dict__ == {
"data": 1,
"annot": "annotation",
}, child_sample.__dict__
# Overwriting and adding delayed_attributes to the child
new_delayed_attr = {
"annot": load_annot_variant, # Override parent's annot
"new_annot": load_annot, # Add the new_annot attribute
"read_check": load_check,
}
child_sample = DelayedSample(load_data,
parent=delayed_sample,
delayed_attributes=new_delayed_attr)
assert child_sample.data == 0, child_sample.data
assert child_sample.annot == "annotation_variant", child_sample.annot
assert child_sample.new_annot == "annotation", child_sample.new_annot
assert not delayed_attr_read, "delayed attribute has been read early"
assert child_sample.read_check == "delayed_attr_data", child_sample.read_check
assert delayed_attr_read, "delayed attribute should have been read by now"
delayed_sample.annot = "changed"
assert delayed_sample.annot == "changed", delayed_sample.annot
def generate_samples(n_subjects,
n_samples_per_subject,
shape=(2, 2),
annotations=1):
"""
Simple sample generator that generates a certain number of samples per
subject, whose data is np.zeros + subject index
"""
samples = []
for i in range(n_subjects):
data = np.zeros(shape) + i
for j in range(n_samples_per_subject):
samples += [
Sample(
data,
subject=str(i),
key=str(i * n_subjects + j),
annotations=annotations,
)
]
return samples
def _delayed_samples_to_samples(delayed_samples):
return [Sample(sample.data, parent=sample) for sample in delayed_samples]
def score_sample_templates(self, probe_samples, enroll_samples,
score_all_vs_all):
"""Computes the similarity score between all probe and enroll templates.
Parameters
----------
probe_samples : list
A list (length N) of Samples containing probe templates.
enroll_samples : list
A list (length M) of Samples containing enroll templates.
score_all_vs_all : bool
If True, the similarity scores between all probe and enroll templates
are computed. If False, the similarity scores between the probes and
their associated enroll templates are computed.
Returns
-------
score_samplesets : list
A list of N SampleSets each containing a list of M score Samples if score_all_vs_all
is True. Otherwise, a list of N SampleSets each containing a list of <=M score Samples
depending on the database.
"""
logger.debug(
f"{_frmt(self)}.score_sample_templates(... score_all_vs_all={score_all_vs_all})"
)
# Returns a list of SampleSets where a Sampleset for each probe
# SampleSet where each Sample inside the SampleSets contains the score
# for one enroll SampleSet
score_samplesets = []
if score_all_vs_all:
probe_data = [s.data for s in probe_samples]
valid_probe_indices = [
i for i, d in enumerate(probe_data) if _data_valid(d)
]
valid_probe_data = [probe_data[i] for i in valid_probe_indices]
scores = self.compare(SampleBatch(enroll_samples),
valid_probe_data)
scores = np.asarray(scores, dtype=float)
if len(valid_probe_indices) != len(probe_data):
# inject None scores for invalid probe samples
scores: list = scores.T.tolist()
for i in range(len(probe_data)):
if i not in valid_probe_indices:
scores.insert(i, [None] * len(enroll_samples))
# transpose back to original shape
scores = np.array(scores, dtype=float).T
expected_shape = (len(enroll_samples), len(probe_samples))
assert scores.shape == expected_shape, (
"The shape of the similarity scores (%s) does not match the expected shape (%s)"
% (scores.shape, expected_shape))
for j, probe in enumerate(probe_samples):
samples = []
for i, enroll in enumerate(enroll_samples):
samples.append(Sample(scores[i, j], parent=enroll))
score_samplesets.append(SampleSet(samples, parent=probe))
else:
for probe in probe_samples:
references = [str(ref) for ref in probe.references]
# get the indices of references for enroll samplesets
indices = [
i for i, enroll in enumerate(enroll_samples)
if str(enroll.reference_id) in references
]
if not indices:
raise ValueError(
f"No enroll sampleset found for probe {probe} and its required references {references}. "
"Did you mean to set score_all_vs_all=True?")
if not _data_valid(probe.data):
scores = [[None]] * len(indices)
else:
scores = self.compare(
SampleBatch([enroll_samples[i] for i in indices]),
SampleBatch([probe]),
)
scores = np.asarray(scores, dtype=float)
expected_shape = (len(indices), 1)
assert scores.shape == expected_shape, (
"The shape of the similarity scores (%s) does not match the expected shape (%s)"
% (scores.shape, expected_shape))
samples = []
for i, j in enumerate(indices):
samples.append(
Sample(scores[i, 0], parent=enroll_samples[j]))
score_samplesets.append(SampleSet(samples, parent=probe))
return score_samplesets