def main(_):
tf.gfile.MkDir(args.output_dir)
data = ByteWavWholeReader(speaker_list=txt2list(args.speaker_list),
filenames=tf.gfile.Glob(args.file_pattern),
num_epoch=1)
XNOM = data.f[0]
XWAV = tf.expand_dims(mu_law_decode(data.x[0, :]), -1)
XBIN = tf.contrib.ffmpeg.encode_audio(XWAV, 'wav', 16000)
sess_config = tf.ConfigProto(allow_soft_placement=True,
gpu_options=tf.GPUOptions(allow_growth=True))
with tf.Session(config=sess_config) as sess:
sess.run(tf.tables_initializer())
sess.run(data.iterator.initializer)
csv = open('vctk.csv', 'w')
counter = 1
while True:
try:
fetch = {'xbin': XBIN, 'xwav': XWAV, 'wav_name': XNOM}
result = sess.run(fetch)
wav_name = result['wav_name'].decode('utf8')
print('\rFile {:05d}: Processing {}'.format(counter, wav_name),
end='')
csv.write('{}, {:d}\n'.format(wav_name, len(result['xwav'])))
filename = os.path.join(args.output_dir, wav_name) + '.wav'
with open(filename, 'wb') as fp:
fp.write(result['xbin'])
counter += 1
except tf.errors.OutOfRangeError:
print('\nEpoch complete')
break
print()
csv.close()
def main(unused_args):
'''
NOTE: the directory structure must be [args.dir_to_wav]/[Set]/[speakers]
'''
if not args.output_dir:
raise ValueError('`output_dir` (output dir) should be specified')
print('[WARNING] Protobuf is super slow (~7 examples per sec). \n'
'This could take 2 hours or more.')
reader = tf.WholeFileReader()
files = tf.gfile.Glob(args.file_pattern)
filename_queue = tf.train.string_input_producer(files,
num_epochs=1,
shuffle=False)
key, val = reader.read(filename_queue)
'''
wav = tf.contrib.ffmpeg.decode_audio(val, args.ext, args.fs, 1)
wav = tf.reshape(wav, [-1, ])
mulaw = mu_law_encode(wav)
'''
for s in txt2list(args.speaker_list):
tf.gfile.MakeDirs(join(args.output_dir, s))
counter = 1
N = len(files)
with tf.train.MonitoredSession() as sess:
while not sess.should_stop():
filename = sess.run(key).decode('utf8')
binary, _ = librosa.load(filename)
x_int = mu_law_encode(binary)
# TODO: remove this
#decoded = mu_law_decode(x_int)
#librosa.output.write_wav('testwav-{}.wav', decoded, _)
text = read_text(filename)
b, _ = splitext(filename)
_, b = split(b)
s = b.split('_')[0]
ex = make_mu_law_speaker_length(x_int, s, text, b)
fp = tf.python_io.TFRecordWriter(
join(args.output_dir, s, '{}.tfr'.format(b)))
fp.write(ex.SerializeToString())
fp.close()
print('\rFile {:5d}/{:5d}: {}'.format(counter, N, b), end='')
counter += 1
print()
def main(unused_args):
if args.logdir is None:
raise ValueError('Please specify the dir to the checkpoint')
arch = tf.gfile.Glob(join(args.logdir, 'arch*.json'))[0]
arch = json2dict(arch)
net = VQVAE(arch)
data = ByteWavWholeReader(speaker_list=txt2list(args.speaker_list),
filenames=tf.gfile.Glob(args.file_pattern))
ZH = net.encode(data.x, args.mode)
ema = tf.train.ExponentialMovingAverage(decay=0.995)
trg_vars = {ema.average_name(v): v for v in tf.trainable_variables()}
saver = tf.train.Saver(trg_vars)
sess_config = tf.ConfigProto(allow_soft_placement=True,
gpu_options=tf.GPUOptions(allow_growth=True))
with tf.Session(config=sess_config) as sess:
sess.run(tf.tables_initializer())
sess.run(data.iterator.initializer)
sess.run(tf.global_variables_initializer())
load(saver, sess, args.logdir, ckpt=args.ckpt)
hist = np.zeros([
arch['num_exemplar'],
], dtype=np.int64)
counter = 1
while True:
try:
z_ids = sess.run(ZH)
print('\rNum of processed files: {:d}'.format(counter), end='')
counter += 1
for i in z_ids[0]: # bz = 1
hist[i] += 1
except tf.errors.OutOfRangeError:
print()
break
with open('histogram.npf', 'wb') as fp:
hist.tofile(fp)
plt.figure(figsize=[10, 2])
plt.plot(np.log10(hist + 1), '.')
plt.xlim([0, arch['num_exemplar'] - 1])
plt.ylabel('log-frequency')
plt.xlabel('exemplar index')
plt.savefig('histogram.png')
plt.close()
def main(_):
speaker_list = txt2list(args.speaker_list)
dirs = validate_log_dirs(args)
arch = json2dict(args.arch)
arch.update(dirs)
arch.update({'ckpt': args.ckpt})
copy_arch_file(args.arch, arch['logdir'])
net = VQVAE(arch)
P = net.n_padding()
print('Receptive field: {} samples ({:.2f} sec)\n'.format(P, P / arch['fs']))
data = ByteWavReader(
speaker_list,
args.file_pattern,
T=arch['T'],
batch_size=arch['training']['batch_size'],
buffer_size=5000
)
net.train(data)
def main(_):
"""Train the model based on the command-line arguments."""
# Parse command-line arguments
speaker_list = txt2list(args.speaker_list)
dirs = validate_log_dirs(args)
arch = json2dict(args.arch)
arch.update(dirs)
arch.update({'ckpt': args.ckpt})
copy_arch_file(args.arch, arch['logdir'])
# Initialize the model
net = VQVAE(arch)
P = net.n_padding()
print('Receptive field: {} samples ({:.2f} sec)'.format(P, P / arch['fs']))
# Read the input data as specified by the command line arguments
data = ByteWavReader(speaker_list,
args.file_pattern,
T=arch['T'],
batch_size=arch['training']['batch_size'],
buffer_size=5000)
# Train the model on the input data
net.train(data)
def main(unused_args):
if args.logdir is None:
raise ValueError('Please specify the dir to the checkpoint')
speaker_list = txt2list(args.speaker_list)
arch = tf.gfile.Glob(os.path.join(args.logdir, 'arch*.json'))[0]
arch = json2dict(arch)
net = VQVAE(arch)
# they start roughly at the same position but end very differently (3 is longest)
filenames = [
'dataset/VCTK/tfr/p227/p227_363.tfr',
# 'dataset/VCTK/tfr/p240/p240_341.tfr',
# 'dataset/VCTK/tfr/p243/p243_359.tfr',
'dataset/VCTK/tfr/p225/p225_001.tfr'
]
data = ByteWavWholeReader(speaker_list, filenames)
X = tf.placeholder(dtype=tf.int64, shape=[None, None])
Y = tf.placeholder(dtype=tf.int64, shape=[
None,
])
ZH = net.encode(X, args.mode)
XH = net.generate(X, ZH, Y)
# XWAV = mu_law_decode(X)
# XBIN = tf.contrib.ffmpeg.encode_audio(XWAV, 'wav', arch['fs'])
ema = tf.train.ExponentialMovingAverage(decay=0.995)
trg_vars = {ema.average_name(v): v for v in tf.trainable_variables()}
saver = tf.train.Saver(trg_vars)
logdir = get_default_logdir(args.logdir)
tf.gfile.MkDir(logdir)
sess_config = tf.ConfigProto(allow_soft_placement=True,
gpu_options=tf.GPUOptions(allow_growth=True))
with tf.Session(config=sess_config) as sess:
sess.run(tf.tables_initializer())
sess.run(data.iterator.initializer)
results = []
for _ in filenames:
result = sess.run({'x': data.x, 'y': data.y})
results.append(result)
# results1 = sess.run({'x': data.x, 'y': data.y})
# results2 = sess.run({'x': data.x, 'y': data.y})
length_input = net.n_padding() + 1 # same as padding + 1
ini = 15149 - length_input
end = 42285
# x_source1 = results1['x'][:, ini: end]
# x_source2 = results2['x'][:, ini: end]
for i in range(len(results)):
x = results[i]['x']
if x.shape[-1] < end:
x = np.concatenate(
[x, x[0, 0] + np.zeros([1, end - x.shape[-1]])], -1)
results[i]['x'] = x[:, ini:end]
# from pdb import set_trace
# set_trace()
x_source = np.concatenate([
results[0]['x'], results[0]['x'], results[1]['x'], results[1]['x']
], 0)
B = x_source.shape[0]
y_input = np.concatenate([
results[0]['y'], results[1]['y'], results[1]['y'], results[0]['y']
], 0)
length_target = x_source.shape[1] - length_input
while True:
sess.run(tf.global_variables_initializer())
load(saver, sess, args.logdir, ckpt=args.ckpt)
z_blend = sess.run(ZH, feed_dict={X: x_source})
x_input = x_source[:, :length_input]
z_input = z_blend[:, :length_input, :]
# Generate
try:
x_gen = np.zeros([B, length_target],
dtype=np.int64) # + results['x'][0, 0]
for i in range(length_target):
xh = sess.run(XH,
feed_dict={
X: x_input,
ZH: z_input,
Y: y_input
})
z_input = z_blend[:, i + 1:i + 1 + length_input, :]
x_input[:, :-1] = x_input[:, 1:]
x_input[:, -1] = xh[:, -1]
x_gen[:, i] = xh[:, -1]
print('\rGenerating {:5d}/{:5d}... x={:3d}'.format(
i + 1, length_target, xh[0, -1]),
end='',
flush=True)
except KeyboardInterrupt:
print("Interrupted by the user.")
finally:
print()
x_wav = mu_law_decode(x_gen)
for i in range(x_wav.shape[0]):
x_1ch = np.expand_dims(x_wav[i], -1)
# x_bin = sess.run(XBIN, feed_dict={X: x_1ch})
librosa.output.write_wav('testwav-{}.wav'.format(i), x_1ch,
arch['fs'])
# with open(os.path.join(logdir, 'testwav-{}.wav'.format(i)), 'wb') as fp:
# fp.write(x_bin)
# For periodic gen.
if args.period > 0:
try:
print('Sleep for a while')
sleep(args.period * 60)
logdir = get_default_logdir(args.logdir)
tf.gfile.MkDir(logdir)
except KeyboardInterrupt:
print('Stop periodic gen.')
break
finally:
print('all finished')
else:
break