diff --git a/bob/bio/base/pipelines/vanilla_biometrics/wrappers.py b/bob/bio/base/pipelines/vanilla_biometrics/wrappers.py
index 157c0def30ece6b451796f480306906f821526ad..fed15da90ea07dfff467b338e193da49378eeec5 100644
--- a/bob/bio/base/pipelines/vanilla_biometrics/wrappers.py
+++ b/bob/bio/base/pipelines/vanilla_biometrics/wrappers.py
@@ -38,16 +38,24 @@ class BioAlgorithmCheckpointWrapper(BioAlgorithm):
"""
- def __init__(self, biometric_algorithm, base_dir, force=False, **kwargs):
+ def __init__(self, biometric_algorithm, base_dir, group=None, force=False, **kwargs):
super().__init__(**kwargs)
- self.biometric_reference_dir = os.path.join(base_dir, "biometric_references")
- self.score_dir = os.path.join(base_dir, "scores")
+ self.base_dir = base_dir
+ self.set_score_references_path(group)
+
self.biometric_algorithm = biometric_algorithm
self.force = force
self._biometric_reference_extension = ".hdf5"
- self._score_extension = ".pkl"
- self.base_dir = base_dir
+ self._score_extension = ".pkl"
+
+ def set_score_references_path(self, group):
+ if group is None:
+ self.biometric_reference_dir = os.path.join(self.base_dir, "biometric_references")
+ self.score_dir = os.path.join(self.base_dir, "scores")
+ else:
+ self.biometric_reference_dir = os.path.join(self.base_dir, group, "biometric_references")
+ self.score_dir = os.path.join(self.base_dir, group, "scores")
def enroll(self, enroll_features):
return self.biometric_algorithm.enroll(enroll_features)
@@ -190,6 +198,9 @@ class BioAlgorithmDaskWrapper(BioAlgorithm):
biometric_references, data
)
+ def set_score_references_path(self, group):
+ self.biometric_algorithm.set_score_references_path(group)
+
def dask_vanilla_biometrics(pipeline, npartitions=None, partition_size=None):
"""
diff --git a/bob/bio/base/pipelines/vanilla_biometrics/zt_norm.py b/bob/bio/base/pipelines/vanilla_biometrics/zt_norm.py
index 0b58fb88e0a93115f5cb5aa95c5b289409fdfa79..38b66e73b65f2758479d3a1be05c693ecf4a3760 100644
--- a/bob/bio/base/pipelines/vanilla_biometrics/zt_norm.py
+++ b/bob/bio/base/pipelines/vanilla_biometrics/zt_norm.py
@@ -47,6 +47,17 @@ class ZTNormPipeline(object):
If True, applies TScore normalization on top of raw scores.
If both, z_norm and t_norm are true, it applies score normalization
+ score_writer:
+
+ adaptive_score_fraction: float
+ Set the proportion of the impostor scores used to compute :math:`\mu` and :math:`\std` for the T normalization
+ This is also called as adaptative T-Norm (https://ieeexplore.ieee.org/document/1415220) or
+ Top-Norm (https://ieeexplore.ieee.org/document/4013533)
+
+ adaptive_score_descending_sort bool
+ It true, during the Top-norm statistics computations, sort the scores in descending order
+
+
"""
def __init__(
@@ -55,9 +66,13 @@ class ZTNormPipeline(object):
z_norm=True,
t_norm=True,
score_writer=FourColumnsScoreWriter("./scores.txt"),
+ adaptive_score_fraction=1.0,
+ adaptive_score_descending_sort=True,
):
self.vanilla_biometrics_pipeline = vanilla_biometrics_pipeline
- self.ztnorm_solver = ZTNorm()
+ self.ztnorm_solver = ZTNorm(
+ adaptive_score_fraction, adaptive_score_descending_sort
+ )
self.z_norm = z_norm
self.t_norm = t_norm
@@ -78,7 +93,6 @@ class ZTNormPipeline(object):
allow_scoring_with_all_biometric_references=False,
):
-
self.transformer = self.train_background_model(background_model_samples)
# Create biometric samples
@@ -110,6 +124,7 @@ class ZTNormPipeline(object):
return z_normed_scores
# T NORM
+
t_normed_scores, t_scores, t_biometric_references = self.compute_tnorm_scores(
t_biometric_reference_samples,
probe_features,
@@ -128,8 +143,17 @@ class ZTNormPipeline(object):
t_scores,
allow_scoring_with_all_biometric_references,
)
-
- return raw_scores, z_normed_scores, t_normed_scores, zt_normed_scores
+
+ # S-norm
+ s_normed_scores = self.compute_snorm_scores(z_normed_scores, t_normed_scores)
+
+ return (
+ raw_scores,
+ z_normed_scores,
+ t_normed_scores,
+ zt_normed_scores,
+ s_normed_scores,
+ )
def train_background_model(self, background_model_samples):
return self.vanilla_biometrics_pipeline.train_background_model(
@@ -194,7 +218,7 @@ class ZTNormPipeline(object):
)
t_normed_scores = self.ztnorm_solver.compute_tnorm_scores(
- probe_scores, t_scores, t_biometric_references
+ probe_scores, t_scores, t_biometric_references,
)
return t_normed_scores, t_scores, t_biometric_references
@@ -217,22 +241,31 @@ class ZTNormPipeline(object):
# Z Normalizing the T-normed scores
z_normed_t_normed = self.ztnorm_solver.compute_znorm_scores(
- t_scores, zt_scores, t_biometric_references
+ t_scores, zt_scores, t_biometric_references,
)
# (Z Normalizing the T-normed scores) the Z normed scores
zt_normed_scores = self.ztnorm_solver.compute_tnorm_scores(
- z_normed_scores, z_normed_t_normed, t_biometric_references
+ z_normed_scores, z_normed_t_normed, t_biometric_references,
)
return zt_normed_scores
+ def compute_snorm_scores(self, znormed_scores, tnormed_scores):
+
+ s_normed_scores = self.ztnorm_solver.compute_snorm_scores(
+ znormed_scores, tnormed_scores
+ )
+
+ return s_normed_scores
+
def write_scores(self, scores):
return self.vanilla_biometrics_pipeline.write_scores(scores)
def post_process(self, score_paths, filename):
return self.vanilla_biometrics_pipeline.post_process(score_paths, filename)
+
class ZTNorm(object):
"""
Computes Z, T and ZT Score Normalization of a :any:`BioAlgorithm`
@@ -240,8 +273,24 @@ class ZTNorm(object):
Reference bibliography from: A Generative Model for Score Normalization in Speaker Recognition
https://arxiv.org/pdf/1709.09868.pdf
+
+ Parameters
+ ----------
+
+ adaptive_score_fraction: float
+ Set the proportion of the impostor scores used to compute :math:`\mu` and :math:`\std` for the T normalization
+ This is also called as adaptative T-Norm (https://ieeexplore.ieee.org/document/1415220) or
+ Top-Norm (https://ieeexplore.ieee.org/document/4013533)
+
+ adaptive_score_descending_sort bool
+ It true, during the Top-norm statistics computations, sort the scores in descending order
+
"""
+ def __init__(self, adaptive_score_fraction, adaptive_score_descending_sort):
+ self.adaptive_score_fraction = adaptive_score_fraction
+ self.adaptive_score_descending_sort = adaptive_score_descending_sort
+
def _norm(self, score, mu, std):
# Reference: https://gitlab.idiap.ch/bob/bob.learn.em/-/blob/master/bob/learn/em/test/test_ztnorm.py
# Axis 0=ZNORM
@@ -294,9 +343,7 @@ class ZTNorm(object):
axis=1 computes CORRECTLY the statistics for TNorm
"""
# Dumping all scores
- score_floats = np.array(
- [s.data for sset in sampleset_for_norm for s in sset]
- )
+ score_floats = np.array([s.data for sset in sampleset_for_norm for s in sset])
# Reshaping in PROBE vs BIOMETRIC_REFERENCES
n_probes = len(sampleset_for_norm)
@@ -304,8 +351,25 @@ class ZTNorm(object):
score_floats = score_floats.reshape((n_probes, n_references))
# AXIS ON THE MODELS
- big_mu = np.mean(score_floats, axis=axis)
- big_std = self._compute_std(big_mu, score_floats, axis=axis)
+
+ proportion = int(
+ np.floor(score_floats.shape[axis] * self.adaptive_score_fraction)
+ )
+
+
+ sorted_scores = (
+ -np.sort(-score_floats, axis=axis)
+ if self.adaptive_score_descending_sort
+ else np.sort(score_floats, axis=axis)
+ )
+
+ if axis == 0:
+ top_scores = sorted_scores[0:proportion, :]
+ else:
+ top_scores = sorted_scores[:, 0:proportion]
+
+ big_mu = np.mean(top_scores, axis=axis)
+ big_std = self._compute_std(big_mu, top_scores, axis=axis)
# Creating statistics structure with subject id as the key
stats = {}
@@ -390,10 +454,33 @@ class ZTNorm(object):
Base T-normalization function
"""
- stats = self._compute_stats(sampleset_for_tnorm, t_biometric_references, axis=1)
+ stats = self._compute_stats(
+ sampleset_for_tnorm, t_biometric_references, axis=1,
+ )
return self._tnorm_samplesets(probe_scores, stats)
+ def _snorm(self, z_score, t_score):
+ return 0.5 * (z_score + t_score)
+
+ def _snorm_samplesets(self, znormed_scores, tnormed_scores):
+
+ s_normed_samplesets = []
+ for z, t in zip(znormed_scores, tnormed_scores):
+ s_normed_scores = SampleSet([], parent=z)
+ for b_z, b_t in zip(z, t):
+ score = self._snorm(b_z.data, b_t.data)
+
+ new_sample = Sample(score, parent=b_z)
+ s_normed_scores.samples.append(new_sample)
+ s_normed_samplesets.append(s_normed_scores)
+
+ return s_normed_samplesets
+
+ def compute_snorm_scores(self, znormed_scores, tnormed_scores):
+
+ return self._snorm_samplesets(znormed_scores, tnormed_scores)
+
class ZTNormDaskWrapper(object):
"""
@@ -435,6 +522,11 @@ class ZTNormDaskWrapper(object):
return probe_scores.map_partitions(self.ztnorm._tnorm_samplesets, stats)
+ def compute_snorm_scores(self, znormed_scores, tnormed_scores):
+ return znormed_scores.map_partitions(
+ self.ztnorm._snorm_samplesets, tnormed_scores
+ )
+
class ZTNormCheckpointWrapper(object):
"""
@@ -458,7 +550,6 @@ class ZTNormCheckpointWrapper(object):
self.force = force
self.base_dir = base_dir
-
def _write_scores(self, samples, path):
os.makedirs(os.path.dirname(path), exist_ok=True)
open(path, "wb").write(cloudpickle.dumps(samples))
@@ -488,6 +579,28 @@ class ZTNormCheckpointWrapper(object):
return z_normed_score
+ def _apply_tnorm(self, probe_score, stats):
+
+ path = os.path.join(self.tnorm_score_path, str(probe_score.key) + ".pkl")
+
+ if self.force or not os.path.exists(path):
+ t_normed_score = self.ztnorm._apply_tnorm(probe_score, path)
+
+ self.write_scores(t_normed_score.samples)
+
+ t_normed_score = SampleSet(
+ [
+ DelayedSample(
+ functools.partial(self._load, path), parent=probe_score
+ )
+ ],
+ parent=probe_score,
+ )
+ else:
+ t_normed_score = SampleSet(self._load(path), parent=probe_score)
+
+ return t_normed_score
+
def compute_znorm_scores(
self, probe_scores, sampleset_for_znorm, biometric_references
):
@@ -499,11 +612,13 @@ class ZTNormCheckpointWrapper(object):
def compute_tnorm_scores(
self, probe_scores, sampleset_for_tnorm, t_biometric_references
):
-
return self.ztnorm.compute_tnorm_scores(
probe_scores, sampleset_for_tnorm, t_biometric_references
)
+ def compute_snorm_scores(self, znormed_scores, tnormed_scores):
+ return self.ztnorm.compute_snorm_scores(znormed_scores, tnormed_scores)
+
def _compute_stats(self, sampleset_for_norm, biometric_references, axis=0):
return self.ztnorm._compute_stats(
sampleset_for_norm, biometric_references, axis=axis
@@ -514,3 +629,6 @@ class ZTNormCheckpointWrapper(object):
def _tnorm_samplesets(self, probe_scores, stats):
return self.ztnorm._tnorm_samplesets(probe_scores, stats)
+
+ def _snorm_samplesets(self, probe_scores, stats):
+ return self.ztnorm._snorm_samplesets(probe_scores, stats)
diff --git a/bob/bio/base/test/test_vanilla_biometrics_score_norm.py b/bob/bio/base/test/test_vanilla_biometrics_score_norm.py
index eb9f1593ce7eb3e01f4d6ec35ed2dee5bcb77ae7..3871f641cd3a72ec6f007c83fee8722576e880d4 100644
--- a/bob/bio/base/test/test_vanilla_biometrics_score_norm.py
+++ b/bob/bio/base/test/test_vanilla_biometrics_score_norm.py
@@ -127,7 +127,11 @@ def zt_norm_stubs(references, probes, t_references, z_probes):
zt_normed_scores = _norm(z_normed_scores, z_t_scores, axis=0)
assert zt_normed_scores.shape == (n_reference, n_probes)
- return raw_scores, z_normed_scores, t_normed_scores, zt_normed_scores
+ s_normed_scores = (z_normed_scores+t_normed_scores)*0.5
+ assert s_normed_scores.shape == (n_reference, n_probes)
+
+
+ return raw_scores, z_normed_scores, t_normed_scores, zt_normed_scores, s_normed_scores
def test_norm_mechanics():
@@ -191,6 +195,7 @@ def test_norm_mechanics():
z_normed_scores_ref,
t_normed_scores_ref,
zt_normed_scores_ref,
+ s_normed_scores_ref,
) = zt_norm_stubs(references, probes, t_references, z_probes)
############
@@ -249,7 +254,7 @@ def test_norm_mechanics():
#############
z_vanilla_pipeline = ZTNormPipeline(
- vanilla_pipeline, z_norm=True, t_norm=False,
+ vanilla_pipeline, z_norm=True, t_norm=False
)
if with_checkpoint:
@@ -280,6 +285,7 @@ def test_norm_mechanics():
)
assert np.allclose(z_normed_scores, z_normed_scores_ref)
+
############
# TESTING T-NORM
#############
@@ -313,7 +319,7 @@ def test_norm_mechanics():
t_normed_scores = _dump_scores_from_samples(
t_normed_score_samples, shape=(n_probes, n_references)
- )
+ )
assert np.allclose(t_normed_scores, t_normed_scores_ref)
############
@@ -338,6 +344,7 @@ def test_norm_mechanics():
z_normed_score_samples,
t_normed_score_samples,
zt_normed_score_samples,
+ s_normed_score_samples,
) = zt_vanilla_pipeline(
[],
biometric_reference_sample_sets,
@@ -360,6 +367,11 @@ def test_norm_mechanics():
scheduler="single-threaded"
)
+ s_normed_score_samples = s_normed_score_samples.compute(
+ scheduler="single-threaded"
+ )
+
+
raw_scores = _dump_scores_from_samples(
raw_score_samples, shape=(n_probes, n_references)
)
@@ -380,6 +392,14 @@ def test_norm_mechanics():
)
assert np.allclose(zt_normed_scores, zt_normed_scores_ref)
+ s_normed_scores = _dump_scores_from_samples(
+ s_normed_score_samples, shape=(n_probes, n_references)
+ )
+ assert np.allclose(s_normed_scores, s_normed_scores_ref)
+
+
+
+
# No dask
run(False) # On memory
@@ -418,7 +438,7 @@ def test_znorm_on_memory():
vanilla_biometrics_pipeline, npartitions=2
)
- raw_scores, z_scores, t_scores, zt_scores = vanilla_biometrics_pipeline(
+ raw_scores, z_scores, t_scores, zt_scores, s_scores = vanilla_biometrics_pipeline(
database.background_model_samples(),
database.references(),
database.probes(),
@@ -448,15 +468,22 @@ def test_znorm_on_memory():
t_scores = _concatenate(
vanilla_biometrics_pipeline, t_scores, "scores-dev_tscores"
)
+
zt_scores = _concatenate(
vanilla_biometrics_pipeline, zt_scores, "scores-dev_ztscores"
)
+ s_scores = _concatenate(
+ vanilla_biometrics_pipeline, s_scores, "scores-dev_sscores"
+ )
+
if with_dask:
raw_scores = raw_scores.compute(scheduler="single-threaded")
z_scores = z_scores.compute(scheduler="single-threaded")
t_scores = t_scores.compute(scheduler="single-threaded")
zt_scores = zt_scores.compute(scheduler="single-threaded")
+ s_scores = s_scores.compute(scheduler="single-threaded")
+
if isinstance(score_writer, CSVScoreWriter):
n_lines = 51 if with_dask else 101
@@ -465,12 +492,14 @@ def test_znorm_on_memory():
assert len(open(z_scores[0], "r").readlines()) == n_lines
assert len(open(t_scores[0], "r").readlines()) == n_lines
assert len(open(zt_scores[0], "r").readlines()) == n_lines
+ assert len(open(s_scores[0], "r").readlines()) == n_lines
else:
assert len(raw_scores) == 10
assert len(z_scores) == 10
assert len(t_scores) == 10
assert len(zt_scores) == 10
+ assert len(s_scores) == 10
run_pipeline(False)
run_pipeline(False) # Testing checkpoint
diff --git a/bob/bio/base/transformers/preprocessor.py b/bob/bio/base/transformers/preprocessor.py
index 1b1e15c2a2c6d4552e791a70b642db6d0653faeb..f91c563cfd8c060531d814e6a40cf13a56cfb7d0 100644
--- a/bob/bio/base/transformers/preprocessor.py
+++ b/bob/bio/base/transformers/preprocessor.py
@@ -37,3 +37,6 @@ class PreprocessorTransformer(TransformerMixin, BaseEstimator):
def _more_tags(self):
return {"stateless": True, "requires_fit": False}
+
+ def fit(self, X, y=None):
+ return self