Skip to content
GitLab
Commit 0b334636 authored by Amir MOHAMMADI's avatar Amir MOHAMMADI
Browse files

Merge branch 'dask-pipelines' into 'master' 

Remove deprecated code

See merge request !108
parents cdce425f 676fac8a
Pipeline #45486 failed with stages
in 5 minutes and 10 seconds
Expand all Hide whitespace changes
Inline Side-by-side
bob/pad/face/__init__.py
View file @ 0b334636
from . import extractor, preprocessor, database, test
from . import extractor, preprocessor, database
def get_config():
......
bob/pad/face/config/vanilla_pad/qm_svm.py
View file @ 0b334636
......@@ -8,7 +8,7 @@ from bob.pad.face.extractor import ImageQualityMeasure
from sklearn.svm import SVC
from sklearn.model_selection import GridSearchCV
from sklearn.pipeline import make_pipeline
from bob.pad.base.pipelines.vanilla_pad import FrameContainersToFrames
from bob.pad.face.transformer import VideoToFrames
import bob.pipelines as mario
database = globals().get("database")
......@@ -42,7 +42,7 @@ extractor = mario.wrap(
)
# new stuff #
frame_cont_to_array = FrameContainersToFrames()
frame_cont_to_array = VideoToFrames()
param_grid = [
{"C": [1, 10, 100, 1000], "kernel": ["linear"]},
......
bob/pad/face/extractor/LTSS.py deleted 100644 → 0
View file @ cdce425f
#!/usr/bin/env python
# encoding: utf-8
import numpy
from bob.bio.base.extractor import Extractor
from bob.core.log import setup
logger = setup("bob.pad.face")
from scipy.fftpack import rfft
class LTSS(Extractor, object):
"""Compute Long-term spectral statistics of a pulse signal.
The features are described in the following article:
H. Muckenhirn, P. Korshunov, M. Magimai-Doss, and S. Marcel
Long-Term Spectral Statistics for Voice Presentation Attack Detection,
IEEE/ACM Trans. Audio, Speech and Language Processing. vol 25, n. 11, 2017
Attributes
----------
window_size : :obj:`int`
The size of the window where FFT is computed
framerate : :obj:`int`
The sampling frequency of the signal (i.e the framerate ...)
nfft : :obj:`int`
Number of points to compute the FFT
debug : :obj:`bool`
Plot stuff
concat : :obj:`bool`
Flag if you would like to concatenate features from the 3 color channels
time : :obj:`int`
The length of the signal to consider (in seconds)
"""
def __init__(self, window_size=25, framerate=25, nfft=64, concat=False, debug=False, time=0, **kwargs):
"""Init function
Parameters
----------
window_size : :obj:`int`
The size of the window where FFT is computed
framerate : :obj:`int`
The sampling frequency of the signal (i.e the framerate ...)
nfft : :obj:`int`
Number of points to compute the FFT
debug : :obj:`bool`
Plot stuff
concat : :obj:`bool`
Flag if you would like to concatenate features from the 3 color channels
time : :obj:`int`
The length of the signal to consider (in seconds)
"""
super(LTSS, self).__init__(**kwargs)
self.framerate = framerate
# TODO: try to use window size as NFFT - Guillaume HEUSCH, 04-07-2018
self.nfft = nfft
self.debug = debug
self.window_size = window_size
self.concat = concat
self.time = time
def _get_ltss(self, signal):
"""Compute long term spectral statistics for a signal
Parameters
----------
signal : :py:class:`numpy.ndarray`
The signal
Returns
-------
:py:class:`numpy.ndarray`
The spectral statistics of the signal.
"""
window_stride = int(self.window_size / 2)
# log-magnitude of DFT coefficients
log_mags = []
# go through windows
for w in range(0, (signal.shape[0] - self.window_size), window_stride):
# n is even, as a consequence the fft is as follows [y(0), Re(y(1)), Im(y(1)), ..., Re(y(n/2))]
# i.e. each coefficient, except first and last, is represented by two numbers (real + imaginary)
fft = rfft(signal[w:w+self.window_size], n=self.nfft)
# the magnitude is the norm of the complex numbers, so its size is n/2 + 1
mags = numpy.zeros((int(self.nfft/2) + 1), dtype=numpy.float64)
# first coeff is real
if abs(fft[0]) < 1:
mags[0] = 1
else:
mags[0] = abs(fft[0])
# go through coeffs 2 to n/2
index = 1
for i in range(1, (fft.shape[0]-1), 2):
mags[index] = numpy.sqrt(fft[i]**2 + fft[i+1]**2)
if mags[index] < 1:
mags[index] = 1
index += 1
# last coeff is real too
if abs(fft[-1]) < 1:
mags[index] = 1
else:
mags[index] = abs(fft[-1])
log_mags.append(numpy.log(mags))
# build final feature
log_mags = numpy.array(log_mags)
mean = numpy.mean(log_mags, axis=0)
std = numpy.std(log_mags, axis=0)
ltss = numpy.concatenate([mean, std])
return ltss
def __call__(self, signal):
"""Computes the long-term spectral statistics for given pulse signals.
Parameters
----------
signal: numpy.ndarray
The signal
Returns
-------
feature: numpy.ndarray
the computed LTSS features
"""
# sanity check
if signal.ndim == 1:
if numpy.isnan(numpy.sum(signal)):
return
if signal.ndim == 2 and (signal.shape[1] == 3):
for i in range(signal.shape[1]):
if numpy.isnan(numpy.sum(signal[:, i])):
return
# truncate the signal according to time
if self.time > 0:
number_of_frames = self.time * self.framerate
# check that the truncated signal is not longer
# than the original one
if number_of_frames < signal.shape[0]:
if signal.ndim == 1:
signal = signal[:number_of_frames]
if signal.ndim == 2 and (signal.shape[1] == 3):
new_signal = numpy.zeros((number_of_frames, 3))
for i in range(signal.shape[1]):
new_signal[:, i] = signal[:number_of_frames, i]
signal = new_signal
else:
logger.warning("Sequence should be truncated to {}, but only contains {} => keeping original one".format(number_of_frames, signal.shape[0]))
# also, be sure that the window_size is not bigger that the signal
if self.window_size > int(signal.shape[0] / 2):
self.window_size = int(signal.shape[0] / 2)
logger.warning("Window size reduced to {}".format(self.window_size))
# we have a single pulse
if signal.ndim == 1:
feature = self._get_ltss(signal)
# pulse for the 3 color channels
if signal.ndim == 2 and (signal.shape[1] == 3):
if not self.concat:
feature = self._get_ltss(signal[:, 1])
else:
ltss = []
for i in range(signal.shape[1]):
ltss.append(self._get_ltss(signal[:, i]))
feature = numpy.concatenate([ltss[0], ltss[1], ltss[2]])
if numpy.isnan(numpy.sum(feature)):
logger.warn("Feature not extracted")
return
if numpy.sum(feature) == 0:
logger.warn("Feature not extracted")
return
return feature
bob/pad/face/extractor/LiSpectralFeatures.py deleted 100644 → 0
View file @ cdce425f
#!/usr/bin/env python
# encoding: utf-8
import numpy
from bob.bio.base.extractor import Extractor
from bob.core.log import setup
logger = setup("bob.pad.face")
class LiSpectralFeatures(Extractor, object):
"""Compute features from pulse signals in the three color channels.
The features are described in the following article:
X. Li, J. Komulainen, G. Zhao, P-C Yuen and M. Pietikainen,
Generalized Face Anti-spoofing by Detecting Pulse From Face Videos
Intl Conf. on Pattern Recognition (ICPR), 2016.
Attributes
----------
framerate : :obj:`int`
The sampling frequency of the signal (i.e the framerate ...)
nfft : :obj:`int`
Number of points to compute the FFT
debug : :obj:`bool`
Plot stuff
"""
def __init__(self, framerate=25, nfft=512, debug=False, **kwargs):
"""Init function
Parameters
----------
framerate : :obj:`int`
The sampling frequency of the signal (i.e the framerate ...)
nfft : :obj:`int`
Number of points to compute the FFT
debug : :obj:`bool`
Plot stuff
"""
super(LiSpectralFeatures, self).__init__()
self.framerate = framerate
self.nfft = nfft
self.debug = debug
def __call__(self, signal):
"""Compute the frequency features for the given signal.
Parameters
----------
signal : :py:class:`numpy.ndarray`
The signal
Returns
-------
:py:class:`numpy.ndarray`
the computed features
"""
# sanity check
assert signal.ndim == 2 and signal.shape[1] == 3, "You should provide 3 pulse signals"
for i in range(3):
if numpy.isnan(numpy.sum(signal[:, i])):
return
feature = numpy.zeros(6)
# when keypoints have not been detected, the pulse is zero everywhere
# hence, no psd and no features
zero_counter = 0
for i in range(3):
if numpy.sum(signal[:, i]) == 0:
zero_counter += 1
if zero_counter == 3:
logger.warn("Feature is all zeros")
return feature
# get the frequency spectrum
spectrum_dim = int((self.nfft / 2) + 1)
ffts = numpy.zeros((3, spectrum_dim))
f = numpy.fft.fftfreq(self.nfft) * self.framerate
f = abs(f[:spectrum_dim])
for i in range(3):
ffts[i] = abs(numpy.fft.rfft(signal[:, i], n=self.nfft))
# find the max of the frequency spectrum in the range of interest
first = numpy.where(f > 0.7)[0]
last = numpy.where(f < 4)[0]
first_index = first[0]
last_index = last[-1]
range_of_interest = range(first_index, last_index + 1, 1)
# build the feature vector
for i in range(3):
total_power = numpy.sum(ffts[i, range_of_interest])
max_power = numpy.max(ffts[i, range_of_interest])
feature[i] = max_power
if total_power == 0:
feature[i+3] = 0
else:
feature[i+3] = max_power / total_power
# plot stuff, if asked for
if self.debug:
from matplotlib import pyplot
for i in range(3):
max_idx = numpy.argmax(ffts[i, range_of_interest])
f_max = f[range_of_interest[max_idx]]
logger.debug("Inferred HR = {}".format(f_max*60))
pyplot.plot(f, ffts[i], 'k')
xmax, xmin, ymax, ymin = pyplot.axis()
pyplot.vlines(f[range_of_interest[max_idx]], ymin, ymax, color='red')
pyplot.vlines(f[first_index], ymin, ymax, color='green')
pyplot.vlines(f[last_index], ymin, ymax, color='green')
pyplot.show()
if numpy.isnan(numpy.sum(feature)):
logger.warn("Feature not extracted")
return
return feature
bob/pad/face/extractor/NormalizeLength.py deleted 100644 → 0
View file @ cdce425f
#!/usr/bin/env python
# encoding: utf-8
import numpy
from bob.bio.base.extractor import Extractor
import logging
logger = logging.getLogger("bob.pad.face")
class NormalizeLength(Extractor, object):
"""
Normalize the length of feature vectors, such that
they all have the same dimensions
**Parameters:**
length: int
The final length of the final feature vector
requires_training: boolean
This extractor actually may requires "training".
The goal here is to retrieve the length of the shortest sequence
debug: boolean
Plot stuff
"""
def __init__(self, length=-1, debug=False, requires_training=True, **kwargs):
super(NormalizeLength, self).__init__(requires_training=requires_training, **kwargs)
self.length = length
self.debug = debug
def __call__(self, signal):
"""
Normalize the length of the signal
**Parameters:**
signal: numpy.array
The signal
**Returns:**
signal: numpy.array
the signal with the provided length
"""
# we have a single pulse signal
if signal.ndim == 1:
signal = signal[:self.length]
# we have 3 pulse signal (Li's preprocessing)
# in this case, return the signal corresponding to the green channel
if signal.ndim == 2 and (signal.shape[1] == 3):
signal = signal[:self.length, 1]
if numpy.isnan(numpy.sum(signal)):
return
if signal.shape[0] < self.length:
logger.debug("signal shorter than training shape: {} vs {}".format(signal.shape[0], self.length))
import sys
sys.exit()
tmp = numpy.zeros((self.length), dtype=signal.dtype)
tmp[:, signal.shape[0]]
signal = tmp
if self.debug:
from matplotlib import pyplot
pyplot.plot(signal, 'k')
pyplot.title('Signal truncated')
pyplot.show()
return signal
def train(self, training_data, extractor_file):
"""
This function determines the shortest length across the training set.
It will be used to normalize the length of all the sequences.
**Parameters:**
training_data : [object] or [[object]]
A list of *preprocessed* data that can be used for training the extractor.
Data will be provided in a single list, if ``split_training_features_by_client = False`` was specified in the constructor,
otherwise the data will be split into lists, each of which contains the data of a single (training-)client.
extractor_file : str
The file to write.
This file should be readable with the :py:meth:`load` function.
"""
self.length = 100000
for i in range(len(training_data)):
if training_data[i].shape[0] < self.length:
self.length = training_data[i].shape[0]
logger.info("Signals will be truncated to {} dimensions".format(self.length))
bob/pad/face/extractor/OpticalFlow.py deleted 100644 → 0
View file @ cdce425f
from bob.bio.video import FrameContainer
from bob.io.base import HDF5File
from bob.ip.optflow.liu.cg import flow
from collections import Iterator
from functools import partial
import bob.pipelines as mario
import logging
logger = logging.getLogger(__name__)
def _check_frame(frame):
if frame.dtype == "uint8":
return frame.astype("float64") / 255.0
return frame.astype("float64")
class _Reader:
def __init__(self, i1, flow_method):
self.i1 = _check_frame(i1)
self.flow_method = flow_method
def __call__(self, i2):
i2 = _check_frame(i2)
flows = self.flow_method(self.i1, i2)[:2]
self.i1 = i2
return flows
class OpticalFlow(object):
"""An optical flow extractor
For more information see :any:`bob.ip.optflow.liu.cg.flow`.
Attributes
----------
alpha : float
Regularization weight
inner : int
The number of inner fixed point iterations
iterations : int
The number of conjugate-gradient (CG) iterations
min_width : int
Width of the coarsest level
outer : int
The number of outer fixed point iterations
ratio : float
Downsample ratio
"""
def __init__(
self,
alpha=0.02,
ratio=0.75,
min_width=30,
outer=20,
inner=1,
iterations=50,
**kwargs
):
super().__init__(**kwargs)
self.alpha = alpha
self.ratio = ratio
self.min_width = min_width
self.outer = outer
self.inner = inner
self.iterations = iterations
def __call__(self, video):
"""Computes optical flows on a video
Please note that the video should either be uint8 or float64 with values from 0
to 1.
Parameters
----------
video : numpy.ndarray
The video. Can be a FrameContainer, generator, bob.io.video.reader, or a
numpy array.
Returns
-------
numpy.ndarray
The flows calculated for each pixel. The output shape will be
[number_of_frames - 1, 2, height, width].
"""
if isinstance(video, FrameContainer):
video = video.as_array()
if not isinstance(video, Iterator):
video = iter(video)
i1 = next(video)
reader = _Reader(
i1,
partial(
flow,
alpha=self.alpha,
ratio=self.ratio,
min_width=self.min_width,
n_outer_fp_iterations=self.outer,
n_inner_fp_iterations=self.inner,
n_cg_iterations=self.iterations,
),
)
flows = mario.utils.vstack_features(reader, video)
shape = list(flows.shape)
shape[0] = 2
shape.insert(0, -1)
return flows.reshape(shape)
def write_feature(self, feature, feature_file):
if not isinstance(feature_file, HDF5File):
feature_file = HDF5File(feature_file, "w")
feature_file.set("uv", feature)
feature_file.set_attribute("method", "liu.cg", "uv")
feature_file.set_attribute("alpha", self.alpha, "uv")
feature_file.set_attribute("ratio", self.ratio, "uv")
feature_file.set_attribute("min_width", self.min_width, "uv")
feature_file.set_attribute("n_outer_fp_iterations", self.outer, "uv")
feature_file.set_attribute("n_inner_fp_iterations", self.inner, "uv")
feature_file.set_attribute("n_iterations", self.iterations, "uv")
def read_feature(self, feature_file):
if not isinstance(feature_file, HDF5File):
feature_file = HDF5File(feature_file, "r")
return feature_file["uv"]
bob/pad/face/extractor/PPGSecure.py deleted 100644 → 0
View file @ cdce425f
#!/usr/bin/env python
# encoding: utf-8
import numpy
from bob.bio.base.extractor import Extractor
from bob.core.log import setup
logger = setup("bob.pad.face")
class PPGSecure(Extractor, object):
"""Extract frequency spectra from pulse signals.
The feature are extracted according to what is described in
the following article:
E.M Nowara, A. Sabharwal and A. Veeraraghavan,
"PPGSecure: Biometric Presentation Attack Detection using Photoplethysmograms",
IEEE Intl Conf. on Automatic Face and Gesture Recognition, 2017.
Attributes
----------
framerate : :obj:`int`
The sampling frequency of the signal (i.e the framerate ...)
nfft : :obj:`int`
Number of points to compute the FFT
debug : :obj:`bool`
Plot stuff