class DSPFWA(DeepForCls):
def __init__(self):
# Set up env
pwd = os.path.dirname(__file__)
root_dir = pwd + '/../'
sys.path.append(root_dir + '/externals/')
from DSP_FWA import dsp_fwa_utils
self.pointer = dsp_fwa_utils
self.net = self.pointer.init_model()
self.facelib = Flib()
self.facelib.set_face_detector()
self.facelib.set_landmarks_predictor(68)
def preproc(self, im):
return im
def crop_face(self, im):
loc, point = self.facelib.get_face_loc_landmarks(im)[0]
return self.pointer.crop(im, point), loc
def get_softlabel(self, im):
conf = self.pointer.predict(self.net, im) # fake conf
return conf
def run(self, im):
rois, _ = self.crop_face(im)
conf = self.get_softlabel(rois)
return conf
def __init__(self, mode='meso4'):
"""
mode = meso4, mesoinception4
"""
super(MesoNet, self).__init__()
# Set up env
pwd = os.path.dirname(__file__)
root_dir = pwd + '/../'
sys.path.append(root_dir + '/externals/')
from MesoNet.classifiers import Meso4, MesoInception4
import MesoNet.utils as pointer
self.pointer = pointer
if mode == 'meso4':
self.net = Meso4()
model_path = root_dir + '/externals/MesoNet/weights/Meso4_DF'
elif mode == 'mesoinception4':
self.net = MesoInception4()
model_path = root_dir + '/externals/MesoNet/weights/MesoInception_DF'
else:
raise ValueError('name should be meso4 or mesoinception4')
self.net.load(model_path)
self.facelib = Flib()
self.facelib.set_face_detector()
self.facelib.set_landmarks_predictor(68)
class XceptionNet(DeepForCls):
def __init__(self, mode='c23'):
"""
mode: c23, c40, raw
"""
super(XceptionNet, self).__init__()
# Set up env
pwd = os.path.dirname(__file__)
root_dir = pwd + '/../'
sys.path.append(root_dir + '/externals/')
from xceptionnet.classification import xception_main as df_xception
self.pointer = df_xception
self.model = self.pointer.init_model(mode)
self.facelib = Flib()
self.facelib.set_face_detector()
def preproc(self, im):
# im is face area
im = self.pointer.preprocess_image(im)
return im
def crop_face(self, im):
face_detector = self.facelib._face_detector
# Image size
height, width = im.shape[:2]
gray = cv2.cvtColor(im, cv2.COLOR_BGR2GRAY)
faces = face_detector(gray, 1)
if len(faces):
# For now only take biggest face
face = faces[0]
x, y, size = self.pointer.get_boundingbox(face, width, height)
cropped_face = im[y:y + size, x:x + size]
return cropped_face, [x, y, x + size, y + size]
else:
return None, []
def get_softlabel(self, im):
# Model prediction
output = self.pointer.predict(im, self.model)
conf = output.detach().cpu().numpy()[0][1]
return conf
def get_hardlabel(self, im):
conf = self.get_softlabel(im)
label = np.argmax(conf)
return label
def run(self, im):
cropped_face, _ = self.crop_face(im)
if cropped_face is not None:
preproced_face = self.preproc(cropped_face)
conf = self.get_softlabel(preproced_face)
return conf
else:
return 0.5
def __init__(self):
# Set up env
pwd = os.path.dirname(__file__)
root_dir = pwd + '/../'
sys.path.append(root_dir + '/externals/')
from DSP_FWA import dsp_fwa_utils
self.pointer = dsp_fwa_utils
self.net = self.pointer.init_model()
self.facelib = Flib()
self.facelib.set_face_detector()
self.facelib.set_landmarks_predictor(68)
def __init__(self):
super(CapsuleNet, self).__init__()
# Set up env
pwd = os.path.dirname(__file__)
root_dir = pwd + '/../'
sys.path.append(root_dir + '/externals/')
from Capsule import Capsule_main as df_Capsule
self.pointer = df_Capsule
self.vgg_ext, self.capnet = self.pointer.init_model()
self.facelib = Flib()
self.facelib.set_face_detector()
def __init__(self):
# Set up env
pwd = os.path.dirname(__file__)
root_dir = pwd + '/../'
sys.path.append(root_dir + '/externals/FWA')
import fwa_utils
self.pointer = fwa_utils
self.solver, self.cfg = self.pointer.init_model()
self.facelib = Flib()
self.facelib.set_face_detector()
self.facelib.set_landmarks_predictor(68)
def __init__(self):
super(ClassNSeg, self).__init__()
# Set up env
pwd = os.path.dirname(__file__)
root_dir = pwd + '/../'
sys.path.append(root_dir + '/externals/')
from ClassNSeg import ClassNSeg_main as df_classnseg
self.pointer = df_classnseg
self.model = self.pointer.init_model()
self.facelib = Flib()
self.facelib.set_face_detector()
def __init__(self):
super(VA, self).__init__()
# Set up env
pwd = os.path.dirname(__file__)
root_dir = pwd + '/../'
sys.path.append(root_dir + '/externals/')
from VA import VA_main as df_VA
self.pointer = df_VA
self.model = self.pointer.init_model('deepfake')
self.facelib = Flib()
self.facelib.set_face_detector()
self.facelib.set_landmarks_predictor(68)
class CapsuleNet(DeepForCls):
def __init__(self):
super(CapsuleNet, self).__init__()
# Set up env
pwd = os.path.dirname(__file__)
root_dir = pwd + '/../'
sys.path.append(root_dir + '/externals/')
from Capsule import Capsule_main as df_Capsule
self.pointer = df_Capsule
self.vgg_ext, self.capnet = self.pointer.init_model()
self.facelib = Flib()
self.facelib.set_face_detector()
def preproc(self, im):
# im is face area
im = self.pointer.preprocess_image(im)
return im
def crop_face(self, im):
face_detector = self.facelib._face_detector
# Image size
height, width = im.shape[:2]
gray = cv2.cvtColor(im, cv2.COLOR_BGR2GRAY)
faces = face_detector(gray, 1)
if len(faces):
face = faces[0]
x, y, size = self.pointer.get_boundingbox(face, width, height)
cropped_face = im[y:y + size, x:x + size]
return cropped_face
else:
return None
def get_softlabel(self, im):
# Model prediction
output = self.pointer.predict(im, self.vgg_ext, self.capnet, False)
conf = output.detach().cpu().numpy()[0][1]
return conf
def get_hardlabel(self, im):
conf = self.get_softlabel(im)
label = np.argmax(conf)
return label
def run(self, im):
cropped_face = self.crop_face(im)
if cropped_face is not None:
preproced_face = self.preproc(cropped_face)
conf = self.get_softlabel(preproced_face)
return conf
else:
return 0.5
def __init__(self, mode='c23'):
"""
mode: c23, c40, raw
"""
super(XceptionNet, self).__init__()
# Set up env
pwd = os.path.dirname(__file__)
root_dir = pwd + '/../'
sys.path.append(root_dir + '/externals/')
from xceptionnet.classification import xception_main as df_xception
self.pointer = df_xception
self.model = self.pointer.init_model(mode)
self.facelib = Flib()
self.facelib.set_face_detector()
class MesoNet(DeepForCls):
def __init__(self, mode='meso4'):
"""
mode = meso4, mesoinception4
"""
super(MesoNet, self).__init__()
# Set up env
pwd = os.path.dirname(__file__)
root_dir = pwd + '/../'
sys.path.append(root_dir + '/externals/')
from MesoNet.classifiers import Meso4, MesoInception4
import MesoNet.utils as pointer
self.pointer = pointer
if mode == 'meso4':
self.net = Meso4()
model_path = root_dir + '/externals/MesoNet/weights/Meso4_DF'
elif mode == 'mesoinception4':
self.net = MesoInception4()
model_path = root_dir + '/externals/MesoNet/weights/MesoInception_DF'
else:
raise ValueError('name should be meso4 or mesoinception4')
self.net.load(model_path)
self.facelib = Flib()
self.facelib.set_face_detector()
self.facelib.set_landmarks_predictor(68)
def preproc(self, im):
im = self.pointer.preprocess(im)
return im
def crop_face(self, im):
points = self.facelib.get_face_landmarks(im)[0]
cropped_face = self.pointer.crop_face(im, points)
return cropped_face
def get_softlabel(self, im):
# Model prediction
conf = self.net.predict(im)[0][0] # conf of real
return 1 - conf
def run(self, im):
cropped_face = self.crop_face(im)
preproced_face = self.preproc(cropped_face)
conf = self.get_softlabel(preproced_face)
return conf
class ClassNSeg(DeepForCls):
def __init__(self):
super(ClassNSeg, self).__init__()
# Set up env
pwd = os.path.dirname(__file__)
root_dir = pwd + '/../'
sys.path.append(root_dir + '/externals/')
from ClassNSeg import ClassNSeg_main as df_classnseg
self.pointer = df_classnseg
self.model = self.pointer.init_model()
self.facelib = Flib()
self.facelib.set_face_detector()
def preproc(self, im):
# im is face area
im = self.pointer.preprocess_image(im)
return im
def crop_face(self, im):
face_detector = self.facelib._face_detector
bbox, flag = self.pointer.get_bbox(im, face_detector)
if bbox is None or (not flag):
return None
altered_cropped, cors = self.pointer.extract_face(im, bbox)
return altered_cropped
def get_softlabel(self, im):
# Model prediction
output = self.pointer.predict(im, self.model)
conf = output.detach().numpy()[1]
return conf
def get_hardlabel(self, im):
conf = self.get_softlabel(im)
label = np.argmax(conf)
return label
def run(self, im):
cropped_face = self.crop_face(im)
if cropped_face is not None:
preproced_face = self.preproc(cropped_face)
conf = self.get_softlabel(preproced_face)
return conf
else:
return 0.5
def __init__(self, mode='LR_CelebA'):
super(Upconv, self).__init__()
# Set up env
pwd = os.path.dirname(__file__)
root_dir = pwd + '/../'
sys.path.append(root_dir + '/externals/')
from Upconv import Upconv_main as df_Upconv
self.pointer = df_Upconv
if mode == 'SVM_FacesHQ':
model_path = root_dir + '/externals/Upconv/models/SVM_FacesHQ.pickle'
self.praH = 722
elif mode == 'LR_FacesHQ':
model_path = root_dir + '/externals/Upconv/models/LR_FacesHQ.pickle'
self.praH = 722
elif mode == 'SVM_FF':
model_path = root_dir + '/externals/Upconv/models/SVM_FF.pickle'
self.praH = 300
elif mode == 'LR_FF':
model_path = root_dir + '/externals/Upconv/models/LR_FF.pickle'
self.praH = 300
elif mode == 'SVM_CelebA':
model_path = root_dir + '/externals/Upconv/models/SVM_CelebA.pickle'
self.praH = 80
elif mode == 'SVM_r_CelebA':
model_path = root_dir + '/externals/Upconv/models/SVM_r_CelebA.pickle'
self.praH = 80
elif mode == 'SVM_p_CelebA':
model_path = root_dir + '/externals/Upconv/models/SVM_p_CelebA.pickle'
self.praH = 80
elif mode == 'LR_CelebA':
model_path = root_dir + '/externals/Upconv/models/LR_CelebA.pickle'
self.praH = 80
else:
raise ValueError(
'name should be in [SVM_FacesHQ, LR_FacesHQ, SVM_FF, LR_FF, SVM_CelebA, SVM_r_CelebA, SVM_p_CelebA, LR_CelebA]'
)
self.model = self.pointer.init_model(model_path)
self.facelib = Flib()
self.facelib.set_face_detector()
self.facelib.set_landmarks_predictor(68)
class VA(DeepForCls):
def __init__(self):
super(VA, self).__init__()
# Set up env
pwd = os.path.dirname(__file__)
root_dir = pwd + '/../'
sys.path.append(root_dir + '/externals/')
from VA import VA_main as df_VA
self.pointer = df_VA
self.model = self.pointer.init_model('deepfake')
self.facelib = Flib()
self.facelib.set_face_detector()
self.facelib.set_landmarks_predictor(68)
def preproc(self, im):
# im is face area
im = self.pointer.preprocess_image(im)
return im
def get_softlabel(self, im):
# Model prediction
scores = self.pointer.predict(im, 'deepfake', self.model,
self.facelib._face_detector,
self.facelib._lmark_predictor)
# conf = {'Score_MLP': scores[0][0], 'Score_LogReg': scores[0][1]}
if scores is not None:
return scores[0][0]
else:
return 0.5
def get_hardlabel(self, im):
conf = self.get_softlabel(im)
label = np.argmax(conf)
return label
def run(self, im):
# cropped_face = self.crop_face(im, True)
preproced_face = self.preproc(im)
conf = self.get_softlabel(preproced_face)
return conf
class Upconv(DeepForCls):
def __init__(self, mode='LR_CelebA'):
super(Upconv, self).__init__()
# Set up env
pwd = os.path.dirname(__file__)
root_dir = pwd + '/../'
sys.path.append(root_dir + '/externals/')
from Upconv import Upconv_main as df_Upconv
self.pointer = df_Upconv
if mode == 'SVM_FacesHQ':
model_path = root_dir + '/externals/Upconv/models/SVM_FacesHQ.pickle'
self.praH = 722
elif mode == 'LR_FacesHQ':
model_path = root_dir + '/externals/Upconv/models/LR_FacesHQ.pickle'
self.praH = 722
elif mode == 'SVM_FF':
model_path = root_dir + '/externals/Upconv/models/SVM_FF.pickle'
self.praH = 300
elif mode == 'LR_FF':
model_path = root_dir + '/externals/Upconv/models/LR_FF.pickle'
self.praH = 300
elif mode == 'SVM_CelebA':
model_path = root_dir + '/externals/Upconv/models/SVM_CelebA.pickle'
self.praH = 80
elif mode == 'SVM_r_CelebA':
model_path = root_dir + '/externals/Upconv/models/SVM_r_CelebA.pickle'
self.praH = 80
elif mode == 'SVM_p_CelebA':
model_path = root_dir + '/externals/Upconv/models/SVM_p_CelebA.pickle'
self.praH = 80
elif mode == 'LR_CelebA':
model_path = root_dir + '/externals/Upconv/models/LR_CelebA.pickle'
self.praH = 80
else:
raise ValueError(
'name should be in [SVM_FacesHQ, LR_FacesHQ, SVM_FF, LR_FF, SVM_CelebA, SVM_r_CelebA, SVM_p_CelebA, LR_CelebA]'
)
self.model = self.pointer.init_model(model_path)
self.facelib = Flib()
self.facelib.set_face_detector()
self.facelib.set_landmarks_predictor(68)
def crop_face(self, im):
face_detector = self.facelib._face_detector
# Image size
height, width = im.shape[:2]
gray = cv2.cvtColor(im, cv2.COLOR_BGR2GRAY)
faces = face_detector(gray, 1)
if len(faces):
# For now only take biggest face
face = faces[0]
x, y, size = self.pointer.get_boundingbox(face, width, height)
cropped_face = im[y:y + size, x:x + size]
return cropped_face, [x, y, x + size, y + size]
else:
return None, None
def preproc(self, im):
im = self.pointer.preprocess_image(im)
return im
def get_softlabel(self, im):
conf = self.pointer.predict(im, self.model, self.praH)[0]
return 1 - conf
def get_hardlabel(self, im):
conf = self.get_softlabel(im)
label = np.argmax(conf)
return label
def run(self, im):
cropped_face, boxes = self.crop_face(im)
if cropped_face is not None:
preproced_face = self.preproc(cropped_face)
conf = self.get_softlabel(preproced_face)
else:
conf = 0.5
return conf