Merge branch 'testfix'

README.rst

@@ -9,7 +9,8 @@ Presentation Attack Detection in Voice Biometrics
 
 This package is an extension to the ``bob.pad.base`` package, which provides the basic presentation attack detection (PAD) framework.
 
-The ``bob.pad.voice`` contains additional functionality to run PAD experiments using speech databases. 
+The ``bob.pad.voice`` contains additional functionality to run PAD experiments using speech databases.
+Wrappers for speech databases are also included in this package.
 
 
 Installation

bob/pad/voice/algorithm/histdistance_algorithm.py

@@ -69,15 +69,15 @@ class HistDistanceAlgorithm(Algorithm):
         # print ("Min attack ", numpy.min(attack_features))
         # print ("Max attack ", numpy.max(attack_features))
 
-        from antispoofing.utils.ml import norm
+        from bob.pad.voice.utils import extraction
 
         mean = None
         std = None
         # normalize features column-wise
         if self.normalize_features:
-            mean, std = norm.calc_mean_std(real_features, attack_features, nonStdZero=True)
-            real_features = norm.zeromean_unitvar_norm(real_features, mean, std)
-            attack_features = norm.zeromean_unitvar_norm(attack_features, mean, std)
+            mean, std = extraction.calc_mean_std(real_features, attack_features, nonStdZero=True)
+            real_features = extraction.zeromean_unitvar_norm(real_features, mean, std)
+            attack_features = extraction.zeromean_unitvar_norm(attack_features, mean, std)
 
         # compute average histogram for each type of features
         self.real_mean = numpy.mean(real_features, axis=0)

bob/pad/voice/algorithm/logregr_algorithm.py

@@ -74,15 +74,15 @@ class LogRegrAlgorithm(Algorithm):
         # save the trained model to file for future use
         hdf5file = bob.io.base.HDF5File(projector_file, "w")
 
-        from antispoofing.utils.ml import norm
+        from bob.pad.voice.utils import extraction
 
         mean = None
         std = None
         # reduce the feature space using PCA
         if self.use_PCA_training or self.normalize_features:
-            mean, std = norm.calc_mean_std(real_features, attack_features, nonStdZero=True)
-            real_features = norm.zeromean_unitvar_norm(real_features, mean, std)
-            attack_features = norm.zeromean_unitvar_norm(attack_features, mean, std)
+            mean, std = extraction.calc_mean_std(real_features, attack_features, nonStdZero=True)
+            real_features = extraction.zeromean_unitvar_norm(real_features, mean, std)
+            attack_features = extraction.zeromean_unitvar_norm(attack_features, mean, std)
 
         if self.use_PCA_training:
             pca_trainer = bob.learn.linear.PCATrainer()

bob/pad/voice/test/data/test_energylbp/scores-dev

-001 001 genuine_laptop2_sentence01 19.762860966727
-001 attack attack_phone_sentence01 -11.477293810506
+001 001 genuine_laptop2_sentence01 9.668502723115
+001 attack attack_phone_sentence01 -27.475504915505

bob/pad/voice/test/data/test_energylbp/scores-dev-attack

-001 attack attack_phone_sentence01 -11.477293810506
+001 attack attack_phone_sentence01 -27.475504915505

bob/pad/voice/test/data/test_energylbp/scores-dev-real

-001 001 genuine_laptop2_sentence01 19.762860966727
+001 001 genuine_laptop2_sentence01 9.668502723115

bob/pad/voice/test/data/test_energylbp/scores-eval

-001 001 genuine_phone_sentence01 27.051880353577
-001 attack attack_ss_sentence01 14.836701843382
+001 001 genuine_phone_sentence01 16.425012903955
+001 attack attack_ss_sentence01 8.957517280861

bob/pad/voice/test/data/test_energylbp/scores-eval-attack

-001 attack attack_ss_sentence01 14.836701843382
+001 attack attack_ss_sentence01 8.957517280861

bob/pad/voice/test/data/test_energylbp/scores-eval-real

-001 001 genuine_phone_sentence01 27.051880353577
+001 001 genuine_phone_sentence01 16.425012903955

bob/pad/voice/test/test_spoofspeech.py

@@ -68,7 +68,8 @@ def _spoof(parameters, cur_test_dir, sub_dir, score_types=('dev-real',), scores_
             # assert that the data order is still correct
             assert (data2check[0][:, 0:3] == data2check[1][:, 0:3]).all()
             # assert that the values are OK
-            assert numpy.allclose(data2check[0][:, 3].astype(float), data2check[1][:, 3].astype(float), 1e-5)
+            print (data2check)
+            assert numpy.allclose(data2check[0][:, 3].astype(float), data2check[1][:, 3].astype(float), 1e-3)
 
     finally:
         shutil.rmtree(cur_test_dir)

bob/pad/voice/utils/extraction.py

@@ -24,6 +24,48 @@ import logging
 logger = logging.getLogger("bob.pad.voice")
 
 
+def zeromean_unitvar_norm(data, mean, std):
+    """ Normalized the data with zero mean and unit variance. Mean and variance are in numpy.ndarray format"""
+    return numpy.divide(data - mean, std)
+
+
+def calc_mean(c0, c1=[]):
+    """ Calculates the mean of the data."""
+    if c1 != []:
+        return (numpy.mean(c0, 0) + numpy.mean(c1, 0)) / 2.
+    else:
+        return numpy.mean(c0, 0)
+
+
+def calc_std(c0, c1=[]):
+    """ Calculates the variance of the data."""
+    if c1 == []:
+        return numpy.std(c0, 0)
+    prop = float(len(c0)) / float(len(c1))
+    if prop < 1:
+        p0 = int(math.ceil(1 / prop))
+        p1 = 1
+    else:
+        p0 = 1
+        p1 = int(math.ceil(prop))
+    return numpy.std(numpy.vstack(p0 * [c0] + p1 * [c1]), 0)
+
+
+"""
+@param c0
+@param c1
+@param nonStdZero if the std was zero, convert to one. This will avoid a zero division
+"""
+def calc_mean_std(c0, c1=[], nonStdZero=False):
+    """ Calculates both the mean of the data. """
+    mi = calc_mean(c0, c1)
+    std = calc_std(c0, c1)
+    if (nonStdZero):
+        std[std == 0] = 1
+
+    return mi, std
+
+
 def vad_filter_features(vad_labels, features, filter_frames="trim_silence"):
     """ Trim the spectrogram to remove silent head/tails from the speech sample.
     Keep all remaining frames or either speech or non-speech only
@@ -69,7 +111,7 @@ def vad_filter_features(vad_labels, features, filter_frames="trim_silence"):
                 elif filter_frames == "speech_only":
                     filtered_features = features[speech, :]
                 else:  # when we take all
-                    filtered_features = features[nzstart:nzend+1, :]  # numpy slicing is a non-closed interval [)
+                    filtered_features = features[nzstart:nzend + 1, :]  # numpy slicing is a non-closed interval [)
         else:
             logger.error("vad_filter_features(): VAD labels should be the same length as energy bands")
 

buildout.cfg

@@ -4,13 +4,22 @@
 
 [buildout]
 parts = scripts
-eggs =  gridtk
+eggs =  bob.pad.voice
+        bob.bio.spear
+        bob.pad.base
+        bob.bio.base
+        bob.db.base
+        gridtk
 
 extensions = bob.buildout
              mr.developer
 
 auto-checkout = *
-develop = .
+develop = src/bob.bio.spear
+         src/bob.pad.base
+         src/bob.bio.base
+         src/bob.db.base
+          .
          
 ; options for bob.buildout
 debug = true
@@ -18,6 +27,10 @@ verbose = true
 newest = false
 
 [sources]
+bob.bio.spear = git git@gitlab.idiap.ch:bob/bob.bio.spear.git
+bob.db.base = git git@gitlab.idiap.ch:bob/bob.db.base.git
+bob.bio.base = git git@gitlab.idiap.ch:bob/bob.bio.base.git
+bob.pad.base = git git@gitlab.idiap.ch:bob/bob.pad.base.git
 
 [scripts]
 recipe = bob.buildout:scripts

develop.cfg

@@ -44,7 +44,7 @@ bob.db.avspoof = git git@gitlab.idiap.ch:bob/bob.db.avspoof.git
 bob.db.asvspoof = git git@gitlab.idiap.ch:bob/bob.db.asvspoof.git
 bob.db.base = git git@gitlab.idiap.ch:bob/bob.db.base.git
 bob.bio.base = git git@gitlab.idiap.ch:bob/bob.bio.base.git
-bob.pad.base = git branch=move-pad-db-definition-here git@gitlab.idiap.ch:bob/bob.pad.base.git
+bob.pad.base = git git@gitlab.idiap.ch:bob/bob.pad.base.git
 bob.ap = git git@github.com:pkorshunov/bob.ap.git
 
 [scripts]