def enumerate_align_hash(self, video_list):
align_hash = {}
for video_path in video_list:
video_id = os.path.splitext(video_path)[0].split('/')[-1]
align_path = os.path.join(self.align_path, video_id)+".align"
align_hash[video_id] = Align(self.absolute_max_string_len, text_to_labels).from_file(align_path)
return align_hash
def enumerate_align_hash(self, video_list):
align_hash = {}
videos_to_remove = []
for video_path in video_list:
try:
video_id = os.path.splitext(video_path)[0].split('/')[-1]
align_path = os.path.join(self.align_path, video_id) + ".align"
align_hash[video_id] = Align(
self.absolute_max_string_len,
text_to_labels).from_file(align_path)
except Exception as e:
videos_to_remove.append(video_path)
return align_hash, videos_to_remove
def split_words(video, align):
video_aligns = []
for sub in align.align:
# Create new video
_video = Video(video.vtype, video.face_predictor_path)
_video.face = video.face[sub[0]:sub[1]]
_video.mouth = video.mouth[sub[0]:sub[1]]
_video.set_data(_video.mouth)
# Create new align
_align = Align(align.absolute_max_string_len, align.label_func).from_array([(0, sub[1]-sub[0], sub[2])])
# Append
video_aligns.append((_video, _align))
return video_aligns
def merge(video_aligns):
vsample = video_aligns[0][0]
asample = video_aligns[0][1]
video = Video(vsample.vtype, vsample.face_predictor_path)
video.face = np.ones((0, vsample.face.shape[1], vsample.face.shape[2], vsample.face.shape[3]), dtype=np.uint8)
video.mouth = np.ones((0, vsample.mouth.shape[1], vsample.mouth.shape[2], vsample.mouth.shape[3]), dtype=np.uint8)
align = []
inc = 0
for _video, _align in video_aligns:
video.face = np.concatenate((video.face, _video.face), 0)
video.mouth = np.concatenate((video.mouth, _video.mouth), 0)
for sub in _align.align:
_sub = (sub[0]+inc, sub[1]+inc, sub[2])
align.append(_sub)
inc = align[-1][1]
video.set_data(video.mouth)
align = Align(asample.absolute_max_string_len, asample.label_func).from_array(align)
return (video, align)
print _align.padded_label
print _align.label_length
print np.array_equiv(_align.sentence, align.sentence),
print np.array_equiv(_align.label, align.label),
print np.array_equiv(_align.padded_label, align.padded_label)
curriculum = Curriculum(rules)
video = Video(vtype='face',
face_predictor_path=
'evaluation/models/shape_predictor_68_face_landmarks.dat')
video.from_video('evaluation/samples/id2_vcd_swwp2s.mpg')
align = Align(
absolute_max_string_len=32,
label_func=text_to_labels).from_file('evaluation/samples/swwp2s.align')
print "=== TRAINING ==="
for i in range(6):
curriculum.update(i, train=True)
print curriculum
_video, _align, _ = curriculum.apply(video, align)
show_results(_video, _align, video, align)
print "=== VALIDATION/TEST ==="
for i in range(6):
curriculum.update(i, train=False)
print curriculum
_video, _align, _ = curriculum.apply(video, align)
show_results(_video, _align, video, align)
def get_align(align_url, video_length=75):
align = Align(video_length, text_to_labels).from_file(align_url)
return labels_to_text(align.padded_label.astype('int'))