diff --git a/bob/pad/face/config/lbp_svm.py b/bob/pad/face/config/lbp_svm.py
new file mode 100644
index 0000000000000000000000000000000000000000..1649e702ad52b70c49cd51a359833c3eb92aaad0
--- /dev/null
+++ b/bob/pad/face/config/lbp_svm.py
@@ -0,0 +1,81 @@
+import dask_ml.model_selection as dcv
+
+from sklearn.model_selection import StratifiedGroupKFold
+from sklearn.pipeline import Pipeline
+from sklearn.svm import SVC
+
+from bob.bio.face.annotator import MTCNN
+from bob.bio.face.preprocessor import INormLBP
+from bob.bio.face.utils import make_cropper, pad_default_cropping
+from bob.pad.face.transformer import VideoToFrames
+from bob.pad.face.transformer.histogram import SpatialHistogram
+from bob.pipelines.wrappers import SampleWrapper
+
+
+def _init_pipeline(database, crop_size=(112, 112), grid_size=(3, 3)):
+ # Face Crop
+ # --------------------------
+ annotator = MTCNN(thresholds=(0.1, 0.2, 0.2))
+ cropped_pos = pad_default_cropping(crop_size, database.annotation_type)
+ cropper = make_cropper(
+ cropped_image_size=crop_size,
+ cropped_positions=cropped_pos,
+ fixed_positions=database.fixed_positions,
+ color_channel="rgb",
+ annotator=annotator,
+ )
+ face_cropper = SampleWrapper(
+ cropper[0], transform_extra_arguments=cropper[1], delayed_output=False
+ )
+
+ # Extract LBP
+ # --------------------------
+ lbp_extractor = INormLBP(face_cropper=None, color_channel="gray")
+ lbp_extractor = SampleWrapper(lbp_extractor, delayed_output=False)
+
+ # Histogram
+ # --------------------------
+ histo = SpatialHistogram(grid_size=grid_size, nbins=256)
+ # histo = VideoWrapper(histo)
+ histo = SampleWrapper(histo, delayed_output=False)
+
+ # Classifier
+ # --------------------------
+ sk_classifier = SVC()
+ param_grid = [
+ {
+ "C": [2**p for p in range(-3, 14, 2)],
+ "gamma": [2**p for p in range(-15, 0, 2)],
+ "kernel": ["rbf"],
+ }
+ ]
+ cv = StratifiedGroupKFold(n_splits=3)
+ sk_classifier = dcv.GridSearchCV(
+ sk_classifier, param_grid=param_grid, cv=cv
+ )
+ fit_extra_arguments = [("y", "is_bonafide"), ("groups", "video_key")]
+ classifier = SampleWrapper(
+ sk_classifier,
+ delayed_output=False,
+ fit_extra_arguments=fit_extra_arguments,
+ )
+
+ # Full Pipeline
+ # --------------------------
+ return Pipeline(
+ [
+ ("video2frames", VideoToFrames()),
+ ("cropper", face_cropper),
+ ("lbp", lbp_extractor),
+ ("spatial_histogram", histo),
+ ("classifier", classifier),
+ ]
+ )
+
+
+# Get database information, needed for face cropper
+db = globals()["database"]
+if db is None:
+ raise ValueError("Missing database!")
+# Pipeline #
+pipeline = _init_pipeline(database=db)
diff --git a/bob/pad/face/config/svm_frames.py b/bob/pad/face/config/svm_frames.py
deleted file mode 100644
index ebe9f1063db1185d6ae236e3e32662b6b7446371..0000000000000000000000000000000000000000
--- a/bob/pad/face/config/svm_frames.py
+++ /dev/null
@@ -1,43 +0,0 @@
-from sklearn.model_selection import GridSearchCV
-from sklearn.pipeline import Pipeline
-from sklearn.svm import SVC
-
-import bob.pipelines as mario
-
-from bob.pad.face.transformer import VideoToFrames
-
-preprocessor = globals()["preprocessor"]
-extractor = globals()["extractor"]
-
-# Classifier #
-param_grid = [
- {
- "C": [2**P for P in range(-3, 14, 2)],
- "gamma": [2**P for P in range(-15, 0, 2)],
- "kernel": ["rbf"],
- },
-]
-
-
-# TODO: The grid search below does not take into account splitting frames of
-# each video into a separate group. You might have frames of the same video in
-# both groups of training and validation.
-
-# TODO: This gridsearch can also be part of dask graph using dask-ml and the
-# ``bob_fit_supports_dask_array`` tag from bob.pipelines.
-classifier = GridSearchCV(SVC(), param_grid=param_grid, cv=3)
-classifier = mario.wrap(
- ["sample"],
- classifier,
- fit_extra_arguments=[("y", "is_bonafide")],
-)
-
-# Pipeline #
-pipeline = Pipeline(
- [
- ("preprocessor", preprocessor),
- ("extractor", extractor),
- ("video_to_frames", VideoToFrames()),
- ("svm", classifier),
- ]
-)
diff --git a/bob/pad/face/database/database.py b/bob/pad/face/database/database.py
index d779906ff0432acd411e67085294c087cb2206c7..0c5d53d774801f8b7826c35cd16958a6503fb7c5 100644
--- a/bob/pad/face/database/database.py
+++ b/bob/pad/face/database/database.py
@@ -38,7 +38,7 @@ def delayed_video_load(
step_size=step_size,
transform=get_transform(sample),
)
- delayed_attributes = None
+ delayed_attributes = {"annotations": lambda: None}
if annotation_directory:
path = sample.filename
if not keep_extension_for_annotation:
diff --git a/bob/pad/face/test/test_transformers.py b/bob/pad/face/test/test_transformers.py
index 90d3209cf79f205a334f19b6bb21f93b4988492c..641c3deb5d5124e58a256becac303b115abbf3fc 100644
--- a/bob/pad/face/test/test_transformers.py
+++ b/bob/pad/face/test/test_transformers.py
@@ -14,6 +14,9 @@ def test_video_to_frames():
samples = [mario.Sample(v, key=i) for i, v in enumerate(video_container)]
frame_samples = VideoToFrames().transform(samples)
assert len(frame_samples) == 4
- assert all(s.key == 0 for s in frame_samples)
+ assert all(
+ s.key == "0_{}".format(k) for s, k in zip(frame_samples, [0, 1, 2, 4])
+ )
+ assert all(s.video_key == 0 for s in frame_samples)
assert [s.data for s in frame_samples] == [0, 1, 2, 3]
assert [s.frame_id for s in frame_samples] == [0, 1, 2, 4]
diff --git a/bob/pad/face/transformer/VideoToFrames.py b/bob/pad/face/transformer/VideoToFrames.py
index 8a1bdc5eb31131b9c70a25cc1e9ec865e8371ab9..8db36bdf66886928edcbbacdf955c990f12ec513 100644
--- a/bob/pad/face/transformer/VideoToFrames.py
+++ b/bob/pad/face/transformer/VideoToFrames.py
@@ -4,8 +4,7 @@ from functools import partial
from sklearn.base import BaseEstimator, TransformerMixin
-import bob.pipelines as mario
-
+from bob.pipelines.sample import DelayedSample, Sample
from bob.pipelines.wrappers import _frmt
logger = logging.getLogger(__name__)
@@ -18,33 +17,70 @@ def _get(sth):
class VideoToFrames(TransformerMixin, BaseEstimator):
"""Expands video samples to frame-based samples only when transform is called."""
+ def __init__(self, delayed_output=True):
+ self.delayed_output = delayed_output
+
def transform(self, video_samples):
logger.debug(f"{_frmt(self)}.transform")
- output = []
- for sample in video_samples:
- annotations = getattr(sample, "annotations", {}) or {}
-
+ outputs = []
+ for vid_sample in video_samples:
+ annotations = getattr(vid_sample, "annotations", None)
+ # Define groups with `sample.key`` since we need a unique ID for
+ # each video. The `groups` attribute is used to do cross-validation
+ if not hasattr(vid_sample, "key"):
+ raise ValueError(
+ "Video sample must have a unique `key` "
+ "attribute to be used with {}".format(
+ self.__class__.__name__
+ )
+ )
+ groups = vid_sample.key
# video is an instance of VideoAsArray or VideoLikeContainer
- video = sample.data
-
+ video = vid_sample.data
for frame, frame_id in zip(video, video.indices):
if frame is None:
continue
- # create a load method so that we can create DelayedSamples because
- # the input samples could be DelayedSamples with delayed attributes
- # as well and we don't want to load those delayed attributes.
- new_sample = mario.DelayedSample(
- partial(_get, frame),
- frame_id=frame_id,
- annotations=annotations.get(str(frame_id)),
- parent=sample,
- )
- output.append(new_sample)
-
- return output
+ # Do we have frame annotations?
+ frame_annotations = None
+ if annotations is not None:
+ # Global annotation are present -> query them
+ frame_annotations = annotations.get(str(frame_id))
+ # Update key, otherwise get the one from parent and each frames
+ # get the same one, breaking checkpoint mechanic for steps
+ # later down the pipelines
+ key = "{}_{}".format(vid_sample.key, frame_id)
+ if self.delayed_output:
+ # create a load method so that we can create DelayedSamples
+ # because the input samples could be DelayedSamples with
+ # delayed attributes as well and we don't want to load
+ # those delayed attributes.
+ sample = DelayedSample(
+ partial(_get, frame),
+ frame_id=frame_id,
+ video_key=groups,
+ parent=vid_sample,
+ # Override parent's attributes
+ annotations=frame_annotations,
+ key=key,
+ )
+ else:
+ sample = Sample(
+ frame,
+ frame_id=frame_id,
+ video_key=groups,
+ parent=vid_sample,
+ # Override parent's attributes
+ annotations=frame_annotations,
+ key=key,
+ )
+ outputs.append(sample)
+ return outputs
def fit(self, X, y=None):
return self
def _more_tags(self):
- return {"requires_fit": False, "bob_checkpoint_features": False}
+ return {
+ "requires_fit": False,
+ "bob_checkpoint_features": False,
+ }
diff --git a/bob/pad/face/transformer/histogram.py b/bob/pad/face/transformer/histogram.py
new file mode 100644
index 0000000000000000000000000000000000000000..a3bc4ff7d1724839d3d70900c935bcfed37ce48b
--- /dev/null
+++ b/bob/pad/face/transformer/histogram.py
@@ -0,0 +1,75 @@
+import numpy as np
+
+from sklearn.base import BaseEstimator, TransformerMixin
+from sklearn.utils import check_array
+
+
+def _get_cropping_size(image_size, patch_size):
+ # How many pixels missing to cover the whole image
+ r = image_size % patch_size
+ # Spit gap into two evenly
+ before = r // 2
+ after = image_size - (r - before)
+ return before, after
+
+
+def _extract_patches(image, patch_size):
+ # https://stackoverflow.com/a/16858283
+ h, w = image.shape
+ nrows, ncols = patch_size
+ if h % nrows != 0 or w % ncols != 0:
+ w_left, w_right = _get_cropping_size(w, ncols)
+ h_top, h_bottom = _get_cropping_size(h, nrows)
+ # Perform center crop
+ image = image[h_top:h_bottom, w_left:w_right]
+ return (
+ image.reshape(h // nrows, nrows, -1, ncols)
+ .swapaxes(1, 2)
+ .reshape(-1, nrows, ncols)
+ )
+
+
+class SpatialHistogram(TransformerMixin, BaseEstimator):
+ """
+ Split images into a grid of patches, compute histogram on each one of them
+ and concatenate them to obtain the final descriptor.
+ """
+
+ def __init__(self, grid_size=(4, 4), range=(0, 256), nbins=256):
+ """
+ Constructor
+ :param grid_size: Tuple `(grid_y, grid_x)` indicating the number of
+ patches to extract in each directions
+ :param range: Tuple `(h_min, h_max)` indicating the histogram range.
+ cf numpy.histogram
+ :param nbins: Number of bins in the histogram, cf numpy.histogram
+ """
+ self.grid_size = grid_size
+ self.range = range
+ self.nbins = nbins
+
+ def fit(self, X, y):
+ return self
+
+ def transform(self, X):
+ X = check_array(X, allow_nd=True) # X.shape == (N, H, W)
+ histo = []
+ for sample in X:
+ h = self._spatial_histogram(sample) # [grid_x * grid_y * nbins]
+ histo.append(h)
+ return np.asarray(histo)
+
+ def _spatial_histogram(self, image):
+ """Compute spatial histogram for a given images"""
+ patch_size = [s // g for s, g in zip(image.shape, self.grid_size)]
+ patches = _extract_patches(image=image, patch_size=patch_size)
+ hist = []
+ for patch in patches:
+ h, _ = np.histogram(
+ patch, bins=self.nbins, range=self.range, density=True
+ )
+ hist.append(h)
+ return np.asarray(hist).reshape(-1)
+
+ def _more_tags(self):
+ return {"stateless": True, "requires_fit": False}
diff --git a/doc/baselines.rst b/doc/baselines.rst
index b5a6b19ff6c24134f221656c6a40b4312db24bb5..578ff1bf15d5dbbdf18cda5e025dc81476388b12 100644
--- a/doc/baselines.rst
+++ b/doc/baselines.rst
@@ -83,14 +83,16 @@ Replay-Attack is given here :ref:`bob.pad.face.resources.databases.replay_attack
understand the settings in more detail you can check the corresponding
configuration file: ``bob/pad/face/config/replay_attack.py``.
-Deep-Pix-BiS Baseline
+Running Baseline
~~~~~~~~~~~~~~~~~~~~~
-(see :ref:`bob.pad.face.resources.deep_pix_bis_pad`)
+
+The baseline is run by invoking the command below
.. code-block:: sh
- $ bob pad run-pipeline -vvv replay-attack deep-pix-bis --output <OUTPUT> --dask-client <CLIENT>
+ $ bob pad run-pipeline -vvv replay-attack <BASELINE> --output <OUTPUT> --dask-client <CLIENT>
+where ``<BASELINE>`` can be any of the following: ``lbp-svm``, ``deep-pix-bis``.
This baseline reports scores per frame. To obtain scores per video, you can run::
$ bob pad finalize-scores -vvv <OUTPUT>/scores-{dev,eval}.csv
@@ -99,7 +101,38 @@ Finally, you can evaluate this baseline using::
$ bob pad metrics -vvv --eval <OUTPUT>/scores-{dev,eval}.csv
-which should give you::
+LBP-SVM Baseline
+~~~~~~~~~~~~~~~~~~~~~
+
+The metrics for this baseline should give you::
+
+ [Min. criterion: EER ] Threshold on Development set ``<OUTPUT>/scores-dev.csv`: -1.042440e+00
+ ============== ============== ==============
+ .. Development Evaluation
+ ============== ============== ==============
+ APCER (attack) 20.7% 19.5%
+ APCER_AP 20.7% 19.5%
+ BPCER 20.0% 13.8%
+ ACER 20.3% 16.6%
+ FTA 0.0% 0.0%
+ FPR 20.7% (62/300) 19.5% (78/400)
+ FNR 20.0% (12/60) 13.8% (11/80)
+ HTER 20.3% 16.6%
+ FAR 20.7% 19.5%
+ FRR 20.0% 13.8%
+ PRECISION 0.4 0.5
+ RECALL 0.8 0.9
+ F1_SCORE 0.6 0.6
+ AUC 0.9 0.9
+ AUC-LOG-SCALE 1.4 1.5
+ ============== ============== ==============
+
+
+Deep-Pix-BiS Baseline
+~~~~~~~~~~~~~~~~~~~~~
+(see :ref:`bob.pad.face.resources.deep_pix_bis_pad`)
+
+The metrics for this baseline should give you::
[Min. criterion: EER ] Threshold on Development set `<OUTPUT>/scores-dev.csv`: 1.919391e-01
============== ============== ===============
@@ -136,22 +169,57 @@ settings in more detail you can check the corresponding configuration file :
``bob/pad/face/config/replay_mobile.py``.
-Deep-Pix-BiS Baseline
+Running Baseline
~~~~~~~~~~~~~~~~~~~~~
+The baseline is run by invoking the command below
+
.. code-block:: sh
- $ bob pad run-pipeline -vv replay-mobile deep-pix-bis --output <OUTPUT> --dask-client <CLIENT>
+ $ bob pad run-pipeline -vvv replay-mobile <BASELINE> --output <OUTPUT> --dask-client <CLIENT>
+where ``<BASELINE>`` can be any of the following: ``lbp-svm``, ``deep-pix-bis``.
This baseline reports scores per frame. To obtain scores per video, you can run::
- $ bob pad finalize-scores -vv <OUTPUT>/scores-{dev,eval}.csv
+ $ bob pad finalize-scores -vvv <OUTPUT>/scores-{dev,eval}.csv
Finally, you can evaluate this baseline using::
- $ bob pad metrics -vv --eval <OUTPUT>/scores-{dev,eval}.csv
+ $ bob pad metrics -vvv --eval <OUTPUT>/scores-{dev,eval}.csv
-which should give you::
+
+LBP-SVM Baseline
+~~~~~~~~~~~~~~~~~~~~~
+
+The metrics for this baseline should give you::
+
+ [Min. criterion: EER ] Threshold on Development set `<OUTPUT>/scores-dev.csv`: -5.045229e-01
+ =================== ============== ==============
+ .. Development Evaluation
+ =================== ============== ==============
+ APCER (mattescreen) 7.0% 7.3%
+ APCER (print) 9.4% 3.1%
+ APCER_AP 9.4% 7.3%
+ BPCER 8.1% 0.9%
+ ACER 8.8% 4.1%
+ FTA 0.0% 0.0%
+ FPR 8.2% (21/256) 5.2% (10/192)
+ FNR 8.1% (13/160) 0.9% (1/110)
+ HTER 8.2% 3.1%
+ FAR 8.2% 5.2%
+ FRR 8.1% 0.9%
+ PRECISION 0.9 0.9
+ RECALL 0.9 1.0
+ F1_SCORE 0.9 1.0
+ AUC 1.0 1.0
+ AUC-LOG-SCALE 2.2 2.2
+ =================== ============== ==============
+
+
+Deep-Pix-BiS Baseline
+~~~~~~~~~~~~~~~~~~~~~
+
+The metrics for this baseline should give you::
[Min. criterion: EER ] Threshold on Development set `<OUTPUT>/scores-dev.csv`: 4.051177e-01
=================== ============== ==============
@@ -189,22 +257,59 @@ settings in more detail you can check the corresponding configuration file :
``bob/pad/face/config/oulu_npu.py``.
-Deep-Pix-BiS Baseline
+Running Baseline
~~~~~~~~~~~~~~~~~~~~~
+The baseline is run by invoking the command below
+
.. code-block:: sh
- $ bob pad run-pipeline -vv oulu-npu deep-pix-bis --output <OUTPUT> --dask-client <CLIENT>
+ $ bob pad run-pipeline -vvv oulu-npu <BASELINE> --output <OUTPUT> --dask-client <CLIENT>
+where ``<BASELINE>`` can be any of the following: ``lbp-svm``, ``deep-pix-bis``.
This baseline reports scores per frame. To obtain scores per video, you can run::
- $ bob pad finalize-scores -vv <OUTPUT>/scores-{dev,eval}.csv
+ $ bob pad finalize-scores -vvv <OUTPUT>/scores-{dev,eval}.csv
Finally, you can evaluate this baseline using::
- $ bob pad metrics -vv --eval <OUTPUT>/scores-{dev,eval}.csv
+ $ bob pad metrics -vvv --eval <OUTPUT>/scores-{dev,eval}.csv
+
+
+LBP-SVM Baseline
+~~~~~~~~~~~~~~~~~~~~~
+
+The metrics for this baseline should give you::
+
+ [Min. criterion: EER ] Threshold on Development set `<OUTPUT>/scores-dev.csv`: 6.161214e-02
+ ====================== ============= ============
+ .. Development Evaluation
+ ====================== ============= ============
+ APCER (print/1) 13.3% 10.0%
+ APCER (print/2) 11.1% 6.7%
+ APCER (video_replay/1) 4.4% 5.8%
+ APCER (video_replay/2) 8.3% 11.7%
+ APCER_AP 13.3% 11.7%
+ BPCER 9.4% 57.5%
+ ACER 11.4% 34.6%
+ FTA 0.0% 0.0%
+ FPR 9.3% (67/720) 8.5% (41/480)
+ FNR 9.4% (17/180) 57.5% (69/120)
+ HTER 9.4% 33.0%
+ FAR 9.3% 8.5%
+ FRR 9.4% 57.5%
+ PRECISION 0.7 0.6
+ RECALL 0.9 0.4
+ F1_SCORE 0.8 0.5
+ AUC 1.0 0.8
+ AUC-LOG-SCALE 2.1 1.1
+ ====================== ============= ============
+
-which should give you::
+Deep-Pix-BiS Baseline
+~~~~~~~~~~~~~~~~~~~~~
+
+The metrics for this baseline should give you::
[Min. criterion: EER ] Threshold on Development set `<OUTPUT>/scores-dev.csv`: 4.326179e-01
====================== ============= ============
@@ -244,22 +349,58 @@ settings in more detail you can check the corresponding configuration file :
``bob/pad/face/config/swan.py``.
-Deep-Pix-BiS Baseline
+Running Baseline
~~~~~~~~~~~~~~~~~~~~~
+The baseline is run by invoking the command below
+
.. code-block:: sh
- $ bob pad run-pipeline -vv swan deep-pix-bis --output <OUTPUT> --dask-client <CLIENT>
+ $ bob pad run-pipeline -vvv swan <BASELINE> --output <OUTPUT> --dask-client <CLIENT>
+where ``<BASELINE>`` can be any of the following: ``lbp-svm``, ``deep-pix-bis``.
This baseline reports scores per frame. To obtain scores per video, you can run::
- $ bob pad finalize-scores -vv <OUTPUT>/scores-{dev,eval}.csv
+ $ bob pad finalize-scores -vvv <OUTPUT>/scores-{dev,eval}.csv
Finally, you can evaluate this baseline using::
- $ bob pad metrics -vv --eval <OUTPUT>/scores-{dev,eval}.csv
+ $ bob pad metrics -vvv --eval <OUTPUT>/scores-{dev,eval}.csv
+
+
+LBP-SVM Baseline
+~~~~~~~~~~~~~~~~~~~~~
+
+The metrics for this baseline should give you::
+
+ [Min. criterion: EER ] Threshold on Development set `<OUTPUT>/scores-dev.csv`: 9.408851e-02
+ ====================== ============= ============
+ .. Development Evaluation
+ ====================== ============= ============
+ APCER (PA.F.1) 6.7% 11.1%
+ APCER (PA.F.5) 0.8% 0.8%
+ APCER (PA.F.6) 11.2% 10.8%
+ APCER_AP 11.2% 11.1%
+ BPCER 6.0% 25.2%
+ ACER 8.6% 18.2%
+ FTA 0.0% 0.0%
+ FPR 6.0% (30/502) 6.0% (45/749)
+ FNR 6.0% (18/300) 25.2% (568/2250)
+ HTER 6.0% 15.6%
+ FAR 6.0% 6.0%
+ FRR 6.0% 25.2%
+ PRECISION 0.9 1.0
+ RECALL 0.9 0.7
+ F1_SCORE 0.9 0.8
+ AUC 1.0 1.0
+ AUC-LOG-SCALE 2.4 2.0
+ ====================== ============= ============
+
+
+Deep-Pix-BiS Baseline
+~~~~~~~~~~~~~~~~~~~~~
-which should give you::
+The metrics for this baseline should give you::
[Min. criterion: EER ] Threshold on Development set `<OUTPUT>/scores-dev.csv`: 4.867174e-01
============== ============== ================
diff --git a/doc/other_pad_algorithms.rst b/doc/other_pad_algorithms.rst
index 9feb8a3b1e1fd0911a3105e451e534ef90458617..1429d88d953a715d5266809e02efea4cee820d54 100644
--- a/doc/other_pad_algorithms.rst
+++ b/doc/other_pad_algorithms.rst
@@ -43,10 +43,11 @@ Usually, it is a good idea to have at least verbose level 2 (i.e., calling
line option. To run experiments in parallel on the local machine, add the
``--dask-client local-parallel`` option.
- See :any:`this <bob.bio.base.pipeline_simple_advanced_features>` for more for more
+ See :any:`this <bob.bio.base.pipeline_simple_advanced_features>` for more
details on dask configurations.
+
Database setups and face PAD algorithms are encoded using
``bob.bio.base.configuration-files``, all stored inside the package structure,
in the directory ``bob/pad/face/config``. Documentation for each resource
diff --git a/setup.py b/setup.py
index 8a9cc734f7f4c434bb287d46a20f54addf8cafdb..16986887d71e5363f5e75324e282b06bae028f8b 100644
--- a/setup.py
+++ b/setup.py
@@ -79,7 +79,7 @@ setup(
"replay-mobile = bob.pad.face.config.replay_mobile",
"swan = bob.pad.face.config.swan",
# pipelines
- "svm-frames = bob.pad.face.config.svm_frames",
+ "lbp-svm = bob.pad.face.config.lbp_svm",
"deep-pix-bis = bob.pad.face.config.deep_pix_bis",
],
# registered ``bob pad ...`` commands