def main():
logdir, ckpt = os.path.split(args.checkpoint)
arch = tf.gfile.Glob(os.path.join(
logdir, 'architecture*.json'))[0] # should only be 1 file
with open(arch) as fp:
arch = json.load(fp)
normalizer = Tanhize(
xmax=np.fromfile('./etc/xmax.npf'),
xmin=np.fromfile('./etc/xmin.npf'),
)
features = read_whole_features(args.file_pattern.format(args.src))
x = normalizer.forward_process(features['sp'])
x = nh_to_nchw(x)
y_s = features['speaker']
y_t_id = tf.placeholder(dtype=tf.int64, shape=[
1,
])
y_t = y_t_id * tf.ones(shape=[
tf.shape(x)[0],
], dtype=tf.int64)
machine = MODEL(arch)
z = machine.encode(x)
x_t = machine.decode(z, y_t) # NOTE: the API yields NHWC format
x_t = tf.squeeze(x_t)
x_t = normalizer.backward_process(x_t)
# For sanity check (validation)
x_s = machine.decode(z, y_s)
x_s = tf.squeeze(x_s)
x_s = normalizer.backward_process(x_s)
f0_s = features['f0']
f0_t = convert_f0(f0_s, args.src, args.trg)
output_dir = get_default_output(args.output_dir)
saver = tf.train.Saver()
sv = tf.train.Supervisor(logdir=output_dir)
with sv.managed_session() as sess:
load(saver, sess, logdir, ckpt=ckpt)
while True:
try:
feat, f0, sp = sess.run(
[features, f0_t, x_t],
feed_dict={y_t_id: np.asarray([SPEAKERS.index(args.trg)])})
feat.update({'sp': sp, 'f0': f0})
y = pw2wav(feat)
oFilename = make_output_wav_name(output_dir, feat['filename'])
sf.write(oFilename, y, FS)
except:
break
def main():
logdir, ckpt = os.path.split(args.checkpoint)
arch = tf.gfile.Glob(os.path.join(logdir, 'architecture*.json'))[0] # should only be 1 file
with open(arch) as fp:
arch = json.load(fp)
normalizer = Tanhize(
xmax=np.fromfile('./etc/xmax.npf'),
xmin=np.fromfile('./etc/xmin.npf'),
)
features = read_whole_features(args.file_pattern.format(args.src))
x = normalizer.forward_process(features['sp'])
x = nh_to_nchw(x)
y_s = features['speaker']
y_t_id = tf.placeholder(dtype=tf.int64, shape=[1,])
y_t = y_t_id * tf.ones(shape=[tf.shape(x)[0],], dtype=tf.int64)
machine = MODEL(arch)
z = machine.encode(x)
x_t = machine.decode(z, y_t) # NOTE: the API yields NHWC format
x_t = tf.squeeze(x_t)
x_t = normalizer.backward_process(x_t)
# For sanity check (validation)
x_s = machine.decode(z, y_s)
x_s = tf.squeeze(x_s)
x_s = normalizer.backward_process(x_s)
f0_s = features['f0']
f0_t = convert_f0(f0_s, args.src, args.trg)
output_dir = get_default_output(args.output_dir)
saver = tf.train.Saver()
sv = tf.train.Supervisor(logdir=output_dir)
with sv.managed_session() as sess:
load(saver, sess, logdir, ckpt=ckpt)
while True:
try:
feat, f0, sp = sess.run(
[features, f0_t, x_t],
feed_dict={y_t_id: np.asarray([SPEAKERS.index(args.trg)])}
)
feat.update({'sp': sp, 'f0': f0})
y = pw2wav(feat)
oFilename = make_output_wav_name(output_dir, feat['filename'])
sf.write(oFilename, y, FS)
except:
break
def sample(self):
with tf.name_scope("sample"):
normalizer = Tanhize(
xmax=np.fromfile('./etc/xmax.npf'),
xmin=np.fromfile('./etc/xmin.npf'),
)
features = read_whole_features(
self.args.file_pattern.format(self.args.src))
x = normalizer.forward_process(features['sp'])
x = nh_to_nchw(x)
#y_t_id = tf.placeholder(dtype=tf.int64, shape=[1,])
#y_t = y_t_id * tf.ones(shape=[tf.shape(x)[0],], dtype=tf.int64)
y_t = SPEAKERS.index(self.args.trg) * tf.ones(shape=[
tf.shape(x)[0],
],
dtype=tf.int64)
self.reuse = False
z, _ = self.encoder(x, False)
x_t = self.generator(z, y_t, False)
self.reuse = True
x_t = tf.transpose(x_t, [0, 2, 3, 1])
print "x_t shape:", x_t.get_shape().as_list()
x_t = tf.squeeze(x_t)
x_t = normalizer.backward_process(x_t)
f0_s = features['f0']
f0_t = convert_f0(f0_s, self.args.src, self.args.trg)
sample = dict()
sample['features'] = features
sample['x_t'] = x_t
sample['f0_t'] = f0_t
#sample['y_t'] = y_t_id
return sample
def sample(self):
with tf.name_scope("sample"):
normalizer = Tanhize(
xmax=np.fromfile('./etc/xmax.npf'),
xmin=np.fromfile('./etc/xmin.npf'),
)
FEAT_DIM = 1029
SP_DIM = 513
self.reues = True
#features = read_whole_features(self.args.file_pattern.format(self.args.src))
files = "./dataset/vcc2016/bin/Testing Set/SF1/200005.bin"
#filename_queue = tf.train.string_input_producer(files, num_epochs=1)
#reader = tf.WholeFileReader()
#key, value = reader.read(filename_queue)
key = tf.cast(os.path.split(files)[-1].split('.')[0], tf.string)
with open(files, 'rb') as f:
value = f.read()
value = tf.decode_raw(value, tf.float32)
value = tf.reshape(value, [-1, FEAT_DIM])
features = dict()
features['sp'] = value[:, :SP_DIM]
features['ap'] = value[:, SP_DIM:2 * SP_DIM]
features['f0'] = value[:, SP_DIM * 2]
features['en'] = value[:, SP_DIM * 2 + 1]
features['speaker'] = tf.cast(value[:, SP_DIM * 2 + 2], tf.int64)
features['filename'] = key
#x = normalizer.forward_process(features['sp'])
x = tf.clip_by_value(features['sp'], 0., 1.) * 2. - 1.
x = nh_to_nchw(x)
#y_t_id = tf.placeholder(dtype=tf.int64, shape=[1,])
#y_t = y_t_id * tf.ones(shape=[tf.shape(x)[0],], dtype=tf.int64)
y_t = SPEAKERS.index(self.args.trg) * tf.ones(shape=[
tf.shape(x)[0],
],
dtype=tf.int64)
with tf.variable_scope("encoder", reuse=True):
z, _ = self.encoder(x, False, True)
with tf.variable_scope("generator", reuse=True):
x_t = self.generator(z, y_t, False, True)
x_t = tf.transpose(x_t, [0, 2, 3, 1])
print "x_t shape:", x_t.get_shape().as_list()
x_t = tf.squeeze(x_t)
x_t = normalizer.backward_process(x_t)
f0_s = features['f0']
f0_t = convert_f0(f0_s, self.args.src, self.args.trg)
sample = dict()
sample['features'] = features
sample['x_t'] = x_t
sample['f0_t'] = f0_t
#sample['y_t'] = y_t_id
return sample
def main(unused_args=None):
# args(sys.argv)
if args.model is None:
raise ValueError(
'\n You MUST specify `model`.' +\
'\n Use `python convert.py --help` to see applicable options.'
)
module = import_module(args.module, package=None)
MODEL = getattr(module, args.model)
FS = 16000
with open(args.speaker_list) as fp:
SPEAKERS = [l.strip() for l in fp.readlines()]
logdir, ckpt = os.path.split(args.checkpoint)
if 'VAE' in logdir:
_path_to_arch, _ = os.path.split(logdir)
else:
_path_to_arch = logdir
arch = tf.gfile.Glob(os.path.join(_path_to_arch, 'architecture*.json'))
if len(arch) != 1:
print('WARNING: found more than 1 architecture files!')
arch = arch[0]
with open(arch) as fp:
arch = json.load(fp)
normalizer = Tanhize(
xmax=np.fromfile('./etc/{}_xmax.npf'.format(args.corpus_name)),
xmin=np.fromfile('./etc/{}_xmin.npf'.format(args.corpus_name)),
)
features = read_whole_features(args.file_pattern.format(args.src))
x = normalizer.forward_process(features['sp'])
x = nh_to_nhwc(x)
y_s = features['speaker']
y_t_id = tf.placeholder(dtype=tf.int64, shape=[
1,
])
y_t = y_t_id * tf.ones(shape=[
tf.shape(x)[0],
], dtype=tf.int64)
f0_t = features['f0']
# f0_t = convert_f0(f0_s, args.src, args.trg)
# f0_s_convert = tf.cast(f0_s,dtype=tf.int64)
f0_t_convert = tf.cast(f0_t, dtype=tf.int64)
machine = MODEL(arch, is_training=False)
z = machine.encode(x)
x_t = machine.decode(z, y_t,
f0_t_convert) # NOTE: the API yields NHWC format
x_t = tf.squeeze(x_t)
x_t = normalizer.backward_process(x_t)
output_dir = get_default_output(args.output_dir)
saver = tf.train.Saver()
sv = tf.train.Supervisor(logdir=output_dir)
with sv.managed_session() as sess:
load(saver, sess, logdir, ckpt=ckpt)
print()
while True:
try:
s_time = time.perf_counter()
feat, f0, sp = sess.run(
[features, f0_t, x_t],
feed_dict={y_t_id: np.asarray([SPEAKERS.index(args.trg)])})
feat.update({'sp': sp, 'f0': f0})
y = pw2wav(feat)
oFilename = make_output_wav_name(output_dir, feat['filename'])
print('\rProcessing {}'.format(oFilename), end='')
e_time = time.perf_counter()
print('\n')
print('Time_sp: {}'.format(e_time - s_time))
print('\n')
sf.write(oFilename, y, FS)
except KeyboardInterrupt:
break
finally:
pass
print()
def main():
checkpoint = torch.load(args.model_path)
encoder = Encoder()
generator = G()
encoder.load_state_dict(checkpoint['encoder_state_dict'])
generator.load_state_dict(checkpoint['generator_state_dict'])
encoder.cuda()
generator.cuda()
FS = 16000
SPEAKERS = list()
with open(args.speaker_list) as fp:
SPEAKERS = [l.strip() for l in fp.readlines()]
normalizer = Tanhize(
xmax=np.fromfile('./etc/{}_xmax.npf'.format(args.corpus_name)),
xmin=np.fromfile('./etc/{}_xmin.npf'.format(args.corpus_name)),
)
total_sp_speaker = []
total_speaker = []
total_features = read_whole_features(args.file_pattern.format(args.src))
for features in total_features:
x = normalizer.forward_process(features['sp'])
x = nh_to_nchw(x)
y_s = features['speaker']
#print('????',SPEAKERS.index(args.trg))
#y_t_id = tf.placeholder(dtype=tf.int64, shape=[1,])
#y_t = y_t_id * torch.ones(shape=[x.shape[0],], dtype=torch.int64)
#print(y_t)
x = Variable(torch.FloatTensor(x).cuda(), requires_grad=False)
y_t = torch.ones((x.shape[0])).view(-1, 1) * (SPEAKERS.index(args.trg))
z, _ = encoder(x)
x_t, _ = generator(z, y_t) # NOTE: the API yields NHWC format
x_t = torch.squeeze(x_t)
#print('x_t.shape',x_t.shape)
x_t = normalizer.backward_process(x_t)
#print('backward_process.finish')
x_s, _ = generator(z, y_s)
x_s = torch.squeeze(x_s)
x_s = normalizer.backward_process(x_s)
f0_s = features['f0']
#print(f0_s.shape)
f0_t = convert_f0(f0_s, args.src, args.trg)
#output_dir = get_default_output(args.output_dir)
output_dir = args.output_dir
features['sp'] = x_t.cpu().data.numpy()
features['f0'] = f0_t
#print('=-=-=-=-=-=')
y = pw2wav(features)
oFilename = make_output_wav_name(output_dir, features['filename'])
print(f'\r Processing {oFilename}', end=' ')
if not os.path.exists(os.path.dirname(oFilename)):
try:
os.makedirs(os.path.dirname(oFilename))
except OSError as exc: # Guard against race condition
print('error')
pass
sf.write(oFilename, y, FS)
#print('2: ',features['sp'].shape)
#print('3: ',features['f0'].shape)
print('\n==finish==')
def main():
logdir, ckpt = os.path.split(args.checkpoint)
print('logdir:', logdir)
print('ckpt:', ckpt)
arch = args.arch
with open(arch) as fp:
arch = json.load(fp)
normalizer = Tanhize(
xmax=np.fromfile('./etc/xmax.npf'),
xmin=np.fromfile('./etc/xmin.npf'),
)
features = read_whole_features(args.file_pattern.format(args.src))
print "features shape:", features['sp'].shape
x = normalizer.forward_process(features['sp'])
x = nh_to_nchw(x)
y_s = features['speaker']
y_t_id = tf.placeholder(dtype=tf.int64, shape=[
1,
])
y_t = y_t_id * tf.ones(shape=[
tf.shape(x)[0],
], dtype=tf.int64)
machine = MODEL(arch, args, False, False)
with tf.variable_scope("encoder"):
z, _ = machine.encoder(x, True)
with tf.variable_scope("generator"):
x_t = machine.generator(z, y_t,
True) # NOTE: the API yields NHWC format
x_t = tf.transpose(x, [0, 2, 3, 1])
print "x_t shape:", x_t.get_shape().as_list()
x_t = tf.squeeze(x_t)
x_t = normalizer.backward_process(x_t)
f0_s = features['f0']
f0_t = convert_f0(f0_s, args.src, args.trg)
output_dir = get_default_output(args.output_dir)
print("output_dir########:", output_dir)
saver = tf.train.Saver()
sv = tf.train.Supervisor(logdir=output_dir)
print "logdir:", logdir
with sv.managed_session() as sess:
#sess.run(tf.global_variables_initializer())
#load(saver, sess, logdir, ckpt=ckpt)
ckpt = os.path.join(logdir, ckpt)
saver.restore(sess, ckpt)
while True:
try:
print("Excuting")
feat, f0, sp = sess.run(
[features, f0_t, x_t],
feed_dict={y_t_id: np.asarray([SPEAKERS.index(args.trg)])})
print("Excuting.")
feat.update({'sp': sp, 'f0': f0})
y = pw2wav(feat)
print("Excuting..")
sf.write(
os.path.join(
output_dir,
os.path.splitext(os.path.split(
feat['filename'])[-1])[0] + '.wav'), y, FS)
print("Excuted")
except:
break