def test():
# ==== Test: batch mixer (conclusion: capacity should be larger to make sure good mixing) ====
x, y = read('./dataset/vcc2016/bin/*/*/1*001.bin', 32, min_after_dequeue=1024, capacity=2048)
sv = tf.train.Supervisor()
with sv.managed_session() as sess:
for _ in range(200):
x_, y_ = sess.run([x, y])
print(y_)
# ===== Read binary ====
features = read_whole_features('./dataset/vcc2016/bin/Training Set/SF1/*001.bin')
sv = tf.train.Supervisor()
with sv.managed_session() as sess:
features = sess.run(features)
y = pw2wav(features)
librosa.output.write_wav('test1.wav', y, 16000) # TODO fs should be specified externally.
# ==== Direct read =====
f = './dataset/vcc2016/bin/Training Set/SF1/100001.bin'
features = np.fromfile(f, np.float32)
features = np.reshape(features, [-1, 513*2 + 1 + 1 + 1]) # f0, en, spk
y = pw2wav(features)
librosa.output.write_wav('test2.wav', y, 16000)
def test():
# ==== Test: batch mixer (conclusion: capacity should be larger to make sure good mixing) ====
#x, y = read('./dataset/savee/bin/*/*/*.bin', 1, min_after_dequeue=1024, capacity=2048)
#sv = tf.train.Supervisor()
#with sv.managed_session() as sess:
# for _ in range(200):
# x_, y_ = sess.run([x, y])
# print(y_)
# ===== Read binary ====
features = read_whole_features('./dataset/savee/bin/Training Set/SNeu/*.bin')
sv = tf.train.Supervisor()
with sv.managed_session() as sess:
features = sess.run(features)
y = pw2wav(features)
sf.write('test1.wav', y, 22050) # TODO fs should be specified externally.
# ==== Direct read =====
f = './dataset/savee/bin/Training Set/SNeu/n01.bin'
features = np.fromfile(f, np.float32)
features = np.reshape(features, [-1, 513*2 + 1 + 1 + 1]) # f0, en, spk
y = pw2wav(features)
sf.write('test2.wav', y, 22050)
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 main():
tf.gfile.MkDir('./etc')
# ==== Save max and min value ====
x = read_whole_features(
args.train_file_pattern) # TODO: use it as a obj and keep `n_files`
x_all = list()
y_all = list()
f0_all = list()
sv = tf.train.Supervisor()
with sv.managed_session() as sess:
counter = 1
while True: # TODO: read according to speaker instead of all speakers
try:
features = sess.run(x)
print('\rProcessing {}: {}'.format(counter,
features['filename']),
end='')
x_all.append(features['sp'])
y_all.append(features['speaker'])
f0_all.append(features['f0'])
counter += 1
finally:
pass
print()
x_all = np.concatenate(x_all, axis=0)
y_all = np.concatenate(y_all, axis=0)
f0_all = np.concatenate(f0_all, axis=0)
xmin = x_all.min()
xmax = x_all.max()
with open(args.speaker_list) as fp:
SPEAKERS = [l.strip() for l in fp.readlines()]
# ==== F0 stats ====
for s in SPEAKERS:
print('Speaker {}'.format(s), flush=True)
f0 = f0_all[SPEAKERS.index(s) == y_all]
print(' len: {}'.format(len(f0)))
f0 = f0[f0 > 2.]
f0 = np.log(f0)
mu, std = f0.mean(), f0.std()
# Save as `float32`
with open('./etc/{}.npf'.format(s), 'wb') as fp:
fp.write(np.asarray([mu, std]).tostring())
# ==== Min/Max value ====
# mu = x_all.mean(0)
# std = x_all.std(0)
q005 = np.percentile(x_all, 0.5, axis=0)
q995 = np.percentile(x_all, 99.5, axis=0)
# Save as `float32`
with open('./etc/{}_xmin.npf'.format(args.corpus_name), 'wb') as fp:
fp.write(q005.tostring())
with open('./etc/{}_xmax.npf'.format(args.corpus_name), 'wb') as fp:
fp.write(q995.tostring())
def main():
try:
os.makedirs('./etc')
except OSError as e:
if e.errno != errno.EEXIST:
raise
# ==== Save max and min value ====
x = read_whole_features(
args.train_file_pattern) # TODO: use it as a obj and keep `n_files`
x_all = list()
y_all = list()
f0_all = list()
counter = 1
for features in x: # TODO: read according to speaker instead of all speakers
#print(features)
print('\rProcessing {}: {}'.format(counter, features['filename']),
end='')
x_all.append(features['sp'])
y_all.append(features['speaker'])
f0_all.append(features['f0'])
counter += 1
print()
x_all = np.concatenate(x_all, axis=0)
y_all = np.concatenate(y_all, axis=0)
f0_all = np.concatenate(f0_all, axis=0)
with open(args.speaker_list) as fp:
SPEAKERS = [l.strip() for l in fp.readlines()]
# ==== F0 stats ====
for s in SPEAKERS:
print('Speaker {}'.format(s), flush=True)
f0 = f0_all[SPEAKERS.index(s) == y_all]
print(' len: {}'.format(len(f0)))
f0 = f0[f0 > 2.]
f0 = np.log(f0)
mu, std = f0.mean(), f0.std()
# Save as `float32`
with open('./etc/{}.npf'.format(s), 'wb') as fp:
fp.write(np.asarray([mu, std]).tostring())
# ==== Min/Max value ====
# mu = x_all.mean(0)
# std = x_all.std(0)
q005 = np.percentile(x_all, 0.5, axis=0)
q995 = np.percentile(x_all, 99.5, axis=0)
# Save as `float32`
with open('./etc/{}_xmin.npf'.format(args.corpus_name), 'wb') as fp:
fp.write(q005.tostring())
with open('./etc/{}_xmax.npf'.format(args.corpus_name), 'wb') as fp:
fp.write(q995.tostring())
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():
tf.gfile.MkDir('./etc')
# ==== Save max and min value ====
x = read_whole_features(args.train_file_pattern)
x_all = list()
y_all = list()
f0_all = list()
sv = tf.train.Supervisor()
with sv.managed_session() as sess:
while True:
try:
features = sess.run(x)
print('Processing {}'.format(features['filename']))
x_all.append(features['sp'])
y_all.append(features['speaker'])
f0_all.append(features['f0'])
finally:
pass
x_all = np.concatenate(x_all, axis=0)
y_all = np.concatenate(y_all, axis=0)
f0_all = np.concatenate(f0_all, axis=0)
# ==== F0 stats ====
for s in SPEAKERS:
print('Speaker {}'.format(s), flush=True)
f0 = f0_all[SPEAKERS.index(s) == y_all]
print(' len: {}'.format(len(f0)))
f0 = f0[f0 > 2.]
f0 = np.log(f0)
mu, std = f0.mean(), f0.std()
# Save as `float32`
with open('./etc/{}.npf'.format(s), 'wb') as fp:
fp.write(np.asarray([mu, std]).tostring())
# ==== Min/Max value ====
# mu = x_all.mean(0)
# std = x_all.std(0)
q005 = np.percentile(x_all, 0.5, axis=0)
q995 = np.percentile(x_all, 99.5, axis=0)
# Save as `float32`
with open('./etc/xmin.npf', 'wb') as fp:
fp.write(q005.tostring())
with open('./etc/xmax.npf', 'wb') as fp:
fp.write(q995.tostring())
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():
tf.gfile.MkDir('./etc')
# ==== Save max and min value ====
x = read_whole_features(args.train_file_pattern)
x_all = list()
y_all = list()
f0_all = list()
sv = tf.train.Supervisor()
with sv.managed_session() as sess:
while True:
try:
features = sess.run(x)
print('Processing {}'.format(features['filename']))
x_all.append(features['sp'])
y_all.append(features['speaker'])
f0_all.append(features['f0'])
finally:
pass
x_all = np.concatenate(x_all, axis=0)
y_all = np.concatenate(y_all, axis=0)
f0_all = np.concatenate(f0_all, axis=0)
# ==== F0 stats ====
for s in SPEAKERS:
print('Speaker {}'.format(s), flush=True)
f0 = f0_all[SPEAKERS.index(s) == y_all]
print(' len: {}'.format(len(f0)))
f0 = f0[f0 > 2.]
f0 = np.log(f0)
mu, std = f0.mean(), f0.std()
# Save as `float32`
with open('./etc/{}.npf'.format(s), 'wb') as fp:
fp.write(np.asarray([mu, std]).tostring())
# ==== Min/Max value ====
# mu = x_all.mean(0)
# std = x_all.std(0)
q005 = np.percentile(x_all, 0.5, axis=0)
q995 = np.percentile(x_all, 99.5, axis=0)
# Save as `float32`
with open('./etc/xmin.npf', 'wb') as fp:
fp.write(q005.tostring())
with open('./etc/xmax.npf', 'wb') as fp:
fp.write(q995.tostring())
def main():
x = read_whole_features(DATA_DIR + '*a*/*bin')
x_all = list()
y_all = list()
f0_all = list()
sv = tf.train.Supervisor()
with sv.managed_session() as sess:
while True:
try:
features = sess.run(x)
print('Processing {}'.format(features['filename']))
x_all.append(features['sp'])
y_all.append(features['speaker'])
f0_all.append(features['f0'])
finally:
pass
x_all = np.concatenate(x_all, axis=0)
y_all = np.concatenate(y_all, axis=0)
f0_all = np.concatenate(f0_all, axis=0)
for s in SPEAKERS:
print('Speaker {}'.format(s))
f0 = f0_all[np.int64(s) == y_all]
print(' len: {}'.format(len(f0)))
f0 = f0[f0 > 2.]
f0 = np.log(f0)
mu, std = f0.mean(), f0.std()
print mu, std
# Save as `float32`
with open(DATA_DIR + '/etc/{}.npf'.format(s), 'wb') as fp:
fp.write(np.asarray([mu, std]).tostring())
# ==== Min/Max value ====
# mu = x_all.mean(0)
# std = x_all.std(0)
q005 = np.percentile(x_all, 0.5, axis=0)
q995 = np.percentile(x_all, 99.5, axis=0)
# Save as `float32`
with open(DATA_DIR + '/etc/xmin.npf', 'wb') as fp:
fp.write(q005.tostring())
with open(DATA_DIR + '/etc/xmax.npf', 'wb') as fp:
fp.write(q995.tostring())
def test():
# ==== Test: batch mixer (conclusion: capacity should be larger to make sure good mixing) ====
#x, y = read('./dataset/savee/bin/*/*/*.bin', 1, min_after_dequeue=1024, capacity=2048)
#sv = tf.train.Supervisor()
#with sv.managed_session() as sess:
# for _ in range(200):
# x_, y_ = sess.run([x, y])
# print(y_)
# ===== Read binary ====
features = read_whole_features('./dataset/savee/bin/Test Set/SNeu/DC16.bin')
sv = tf.train.Supervisor()
with sv.managed_session() as sess:
features = sess.run(features)
y,sp = pw2wav(features)
sf.write('origin1.wav', y, 22050) # TODO fs should be specified externally.
with open('01.bin', 'wb') as fp:
fp.write(sp)
"""
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
