def deeplab(
num_classes,
crop_size=[513, 513],
atrous_rates=[6, 12, 18],
output_stride=16,
fine_tune_batch_norm=False,
pretrained=True,
pretained_num_classes=21,
checkpoint_path='./pretrained_models/deeplabv3_pascal_trainval.pth'):
"""DeepLab v3+ for semantic segmentation."""
outputs_to_num_classes = {'semantic': num_classes}
model_options = common.ModelOptions(outputs_to_num_classes,
crop_size=crop_size,
atrous_rates=atrous_rates,
output_stride=output_stride)
feature_extractor = extractor.feature_extractor(
model_options.model_variant,
pretrained=False,
output_stride=model_options.output_stride)
model = DeepLab(feature_extractor, model_options, fine_tune_batch_norm)
if pretrained:
_load_state_dict(model, num_classes, pretained_num_classes,
checkpoint_path)
return model
def generate_data(sub_names, mov_names):
speech = []
noise = []
for sub_name, mov_name in zip(sub_names, mov_names):
fsub = open(sub_name)
intervels = []
while True:
if fsub.readline().strip() == '[Events]':
break
for line in fsub:
if line.startswith('Dialogue:'):
line = line.strip().split(',')
if line[3] != 'Default' or line[9].startswith('{'):
intervels.append(
(parse_time(line[1]), parse_time(line[2]), False))
else:
intervels.append(
(parse_time(line[1]), parse_time(line[2]), True))
intervels.sort(cmp=lambda x, y: cmp(x[0], y[0]))
i = 0
while i < len(intervels) - 1:
if intervels[i][1] > intervels[i + 1][0]:
intervels[i] = (intervels[i][0], intervels[i + 1][1],
intervels[i][2] and intervels[i + 1][2])
del intervels[i + 1]
else:
i = i + 1
dec = ffmpeg_decoder(mov_name, SAMPLE_RATE)
spec = spectrum(dec.ostream.get_handle(), squared=False)
feat = feature_extractor(spec.ostream.get_handle(), intervels, speech,
noise)
dec.start()
spec.start()
feat.start()
feat.join()
return speech, noise
def generate_data(sub_names, mov_names):
speech = []
noise = []
for sub_name, mov_name in zip(sub_names, mov_names):
fsub = open(sub_name)
intervels = []
while True:
if fsub.readline().strip() == '[Events]':
break
for line in fsub:
if line.startswith('Dialogue:'):
line = line.strip().split(',')
if line[3] != 'Default' or line[9].startswith('{'):
intervels.append( (parse_time(line[1]), parse_time(line[2]), False) )
else:
intervels.append( (parse_time(line[1]), parse_time(line[2]), True) )
intervels.sort(cmp=lambda x,y: cmp(x[0], y[0]))
i = 0
while i < len(intervels)-1:
if intervels[i][1] > intervels[i+1][0]:
intervels[i] = (intervels[i][0], intervels[i+1][1], intervels[i][2] and intervels[i+1][2])
del intervels[i+1]
else:
i = i + 1
dec = ffmpeg_decoder(mov_name, SAMPLE_RATE)
spec = spectrum(dec.ostream.get_handle(), squared = False)
feat = feature_extractor(spec.ostream.get_handle(), intervels, speech, noise)
dec.start()
spec.start()
feat.start()
feat.join()
return speech, noise
if line[3] != 'Default' or line[9].startswith('{\\an8'):
intervels.append( (parse_time(line[1]), parse_time(line[2]), False) )
else:
intervels.append( (parse_time(line[1]), parse_time(line[2]), True) )
intervels.sort(cmp=lambda x,y: cmp(x[0], y[0]))
i = 0
while i < len(intervels)-1:
if intervels[i][1] > intervels[i+1][0]:
intervels[i] = (intervels[i][0], intervels[i+1][1], intervels[i][2] and intervels[i+1][2])
del intervels[i+1]
else:
i = i + 1
dec = ffmpeg_decoder(mov_name, SAMPLE_RATE)
spec = spectrum(dec.ostream.get_handle(), squared = False)
feat = feature_extractor(spec.ostream.get_handle(), intervels, speech, noise)
dec.start()
spec.start()
feat.start()
feat.join()
random.shuffle(speech)
random.shuffle(noise)
for f in speech:
plt.imshow(np.log(abs(f)**2).reshape((5, 128)))
plt.show()
