def main(mel_files, model_filename, output_dir, batch_size, implementation):
mel_files = utils.files_to_list(mel_files)
model = torch.load(model_filename)['model']
wavenet = nv_wavenet.NVWaveNet(**(model.export_weights()))
for files in chunker(mel_files, batch_size):
mels = []
for file_path in files:
print(file_path)
mel = torch.load(file_path)
mel = utils.to_gpu(mel)
mels.append(torch.unsqueeze(mel, 0))
cond_input = model.get_cond_input(torch.cat(mels, 0))
audio_data = wavenet.infer(cond_input, implementation)
for i, file_path in enumerate(files):
file_name = os.path.splitext(os.path.basename(file_path))[0]
audio = utils.mu_law_decode_numpy(audio_data[i, :].cpu().numpy(),
wavenet.A)
audio = utils.MAX_WAV_VALUE * audio
audio = 32768.0 * audio
wavdata = audio.astype('int16')
#wavdata = audio.astype('float16')
write("{}/{}.wav".format(output_dir, file_name), 16000, wavdata)
def main(audio_files, model_filename, output_dir, batch_size, speaker_id,
implementation):
audio_files = utils.files_to_list(audio_files)
model = torch.load(model_filename)['model']
model.eval()
wavenet = nv_wavenet.NVWaveNet(
**(model.decoders[speaker_id].export_weights()))
for files in chunker(audio_files, batch_size):
audio_ = []
for file_path in files:
print(file_path)
audio, sampling_rate = utils.load_wav_to_torch(file_path)
if sampling_rate != 16000:
raise ValueError("{} SR doesn't match target {} SR".format(
sampling_rate, 16000))
audio = utils.mu_law_encode(audio / utils.MAX_WAV_VALUE, 256)
audio = utils.to_gpu(audio)
audio_.append(torch.unsqueeze(audio, 0))
latent = model.get_latent_input(torch.cat(audio_, 0))
cond_input = model.decoders[speaker_id].get_cond_input(latent)
audio_data = wavenet.infer(cond_input, implementation)
for i, file_path in enumerate(files):
file_name = os.path.splitext(os.path.basename(file_path))[0]
audio = utils.mu_law_decode_numpy(audio_data[i, :].cpu().numpy(),
wavenet.A)
audio = utils.MAX_WAV_VALUE * audio
wavdata = audio.astype('int16')
write("{}/{}.wav".format(output_dir, file_name), 16000, wavdata)
def main(input_files,
model_dir,
output_dir,
batch_size,
implementation,
data_config,
audio_config,
preload_mels=False):
model_filename = get_latest_checkpoint(model_dir)
print("Model path: {}".format(model_filename))
model = torch.load(model_filename)['model']
wavenet = nv_wavenet.NVWaveNet(**(model.export_weights()))
print("Wavenet num layers: {}, max_dilation: {}".format(
wavenet.num_layers, wavenet.max_dilation))
writer = SummaryWriter(output_dir)
mel_extractor = Mel2SampOnehot(audio_config=audio_config, **data_config)
input_files = utils.files_to_list(input_files)
audio_processor = AudioProcessor(audio_config)
for j, files in enumerate(chunker(input_files, batch_size)):
mels = []
for i, file_path in enumerate(files):
if preload_mels:
mel = np.load(file_path[0]).T
mel = torch.from_numpy(mel)
mel = utils.to_gpu(mel)
else:
audio, _ = utils.load_wav_to_torch(file_path)
file_name = os.path.splitext(os.path.basename(file_path))[0]
writer.add_audio("eval_true/{}/{}".format(i, file_name),
audio / utils.MAX_WAV_VALUE, 0, 22050)
mel = mel_extractor.get_mel(audio)
mel = mel.t().cuda()
mels.append(torch.unsqueeze(mel, 0))
mels = torch.cat(mels, 0)
cond_input = model.get_cond_input(mels)
audio_data = wavenet.infer(cond_input, implementation)
for i, file_path in enumerate(files):
file_name = os.path.splitext(os.path.basename(file_path[0]))[0]
audio = utils.mu_law_decode_numpy(audio_data[i, :].cpu().numpy(),
256)
print("Range of {}.wav before deemphasis : {} to {}".format(
file_name, audio.min(), audio.max()))
if mel_extractor.apply_preemphasis:
audio = audio.astype("float32")
audio = audio_processor.deemphasis(audio[None, :])
audio = audio.numpy()[0]
print("Range of {}.wav after deemphasis : {} to {}".format(
file_name, audio.min(), audio.max()))
audio = np.tanh(audio)
output_filepath = "{}.wav".format(file_name)
output_filepath = os.path.join(output_dir, output_filepath)
assert audio.dtype in [np.float64, np.float32]
assert (np.abs(audio)).max() <= 1
writer.add_audio(output_filepath, audio, 0, 22050)
audio = (audio * 32767).astype("int16")
scipy.io.wavfile.write(output_filepath, 22050, audio)
def __init__(self, training_files, segment_length, mu_quantization,
filter_length, hop_length, win_length, sampling_rate):
audio_files = utils.files_to_list(training_files)
self.audio_files = audio_files
random.seed(1234)
random.shuffle(self.audio_files)
self.segment_length = segment_length
self.mu_quantization = mu_quantization
self.sampling_rate = sampling_rate
def __init__(self,
audio_files,
mu_quantization,
no_chunks,
audio_config,
segment_length,
use_tf=False,
use_lws=True,
load_mel=False,
verbose=False):
audio_files = utils.files_to_list(audio_files)
self.audio_files = audio_files
random.seed(1234)
random.shuffle(self.audio_files)
if not load_mel:
if use_tf:
audio_processor_cls = AudioProcessor
elif use_lws:
audio_processor_cls = LwsAudioProcessor
else:
raise ValueError(
"Mel spectrum can be calculated only with tf or lws!")
self.audio_processor = audio_processor_cls(audio_config)
self.mu_quantization = mu_quantization
self.segment_length = segment_length
audio_params = AudioProcessor._load_params(audio_config)
if verbose:
print("Audio params:")
pprint(audio_params)
self.audio_params = audio_params
self.window_size = audio_params["window_size"]
self.preemphasis_coeff = audio_params["preemphasis_coef"]
self.apply_preemphasis = audio_params["apply_preemphasis"]
self.window_step = audio_params["window_step"]
self.sample_rate = audio_params["sample_rate"]
self.mel_segment_length = int(
np.ceil((segment_length - self.window_size) / self.window_step))
self.num_mels = audio_params["num_mel_bins"]
self.use_tf = use_tf
self.load_mel = load_mel
self.no_chunks = no_chunks
self.use_lws = use_lws
def __init__(self, training_files, segment_length, mu_quantization,
filter_length, hop_length, win_length, sampling_rate,
mel_fmin, mel_fmax):
audio_files = utils.files_to_list(training_files)
self.audio_files = audio_files
random.seed(1234)
random.shuffle(self.audio_files)
mel_fmax = None if mel_fmax == -1 else mel_fmax
self.stft = TacotronSTFT(filter_length=filter_length,
hop_length=hop_length,
win_length=win_length,
sampling_rate=sampling_rate,
mel_fmin=mel_fmin,
mel_fmax=mel_fmax)
self.segment_length = segment_length
self.mu_quantization = mu_quantization
self.sampling_rate = sampling_rate
required=True,
type=str,
help='File containing list of wavefiles')
parser.add_argument('-o',
"--output_dir",
required=True,
type=str,
help='Directory to put Mel-Spectrogram Tensors')
parser.add_argument('-c',
'--config',
type=str,
help='JSON file for configuration')
args = parser.parse_args()
filepaths = utils.files_to_list(args.audio_list)
# Make directory if it doesn't exist
if not os.path.isdir(args.output_dir):
os.makedirs(args.output_dir)
os.chmod(args.output_dir, 0o775)
# Parse config. Only using data processing
with open(args.config) as f:
data = f.read()
config = json.loads(data)
data_config = config["data_config"]
mel_factory = Mel2SampOnehot(**data_config)
for filepath in filepaths:
audio, sampling_rate = utils.load_wav_to_torch(filepath)