def log_audio(model: Tacotron2, iteration: int, logger: Tacotron2Logger, waveglow):
text = "Does it work yet?"
sequence = array(text_to_sequence(text, ['english_cleaners']))[None, :]
sequence = torch.autograd.Variable(
torch.from_numpy(sequence)).cuda().long()
mel_outputs, mel_outputs_postnet, _, alignments = model.inference(sequence)
with torch.no_grad():
audio = waveglow.infer(mel_outputs_postnet, sigma=0.666)
logger.add_audio(text, audio[0].data.cpu(), global_step=iteration, sample_rate=hparams.sampling_rate)
def prepare_directories_and_logger(hparams, args):
if args.rank == 0:
if not os.path.isdir(args.output_directory):
os.makedirs(args.output_directory)
os.chmod(args.output_directory, 0o775)
logger = Tacotron2Logger(os.path.join(args.output_directory, args.log_directory), hparams)
else:
logger = None
return logger
def prepare_directories_and_logger(output_directory, log_directory, rank):
if rank == 0:
# print("Is this getting executed?")
if not os.path.isdir(output_directory):
os.makedirs(output_directory)
os.chmod(output_directory, 0o775)
logger = Tacotron2Logger(log_directory)
else:
logger = None
return logger
def prepare_directories_and_logger(haparms, output_directory, log_directory,
rank):
if rank == 0:
if not os.path.isdir(output_directory):
os.makedirs(output_directory)
os.chmod(output_directory, 0o775)
logger = Tacotron2Logger(os.path.join(output_directory, log_directory), \
hparams.use_mmi, hparams.use_guided_attn_loss)
else:
logger = None
return logger
def prepare_directories_and_logger(output_directory, log_directory, rank,
run_name, prj_name, resume):
if rank == 0:
if not os.path.isdir(os.path.join(output_directory, prj_name,
run_name)):
os.makedirs(os.path.join(output_directory, prj_name, run_name))
os.chmod(os.path.join(output_directory, prj_name, run_name), 0o775)
logger = Tacotron2Logger(
run_name, prj_name, os.path.join(log_directory, prj_name,
run_name), resume)
else:
logger = None
return logger
def prepare_directories_and_logger(output_directory, log_directory, rank):
if rank == 0:
log_path = os.path.join(output_directory, log_directory)
try:
shutil.rmtree(log_path)
except:
pass
if not os.path.isdir(output_directory):
os.makedirs(output_directory)
os.chmod(output_directory, 0o777)
logger = Tacotron2Logger(log_path)
else:
logger = None
return logger
def log_audio(model: Tacotron2, iteration: int, logger: Tacotron2Logger,
waveglow, inference_batch, text_encoded, mel):
# load source data to obtain rhythm using tacotron 2 as a forced aligner
x, y = model.parse_batch(inference_batch)
with torch.no_grad():
# get rhythm (alignment map) using tacotron 2
mel_outputs, mel_outputs_postnet, gate_outputs, rhythm = model.forward(
x)
rhythm = rhythm.permute(1, 0, 2)
for emotion in range(4):
emotion_id = torch.LongTensor([emotion]).cuda()
with torch.no_grad():
mel_outputs, mel_outputs_postnet, gate_outputs, _ = model.inference_noattention(
(text_encoded, mel, emotion_id, rhythm))
audio = waveglow.infer(mel_outputs_postnet, sigma=0.8)
logger.add_audio(f"Emotion {str(emotion)}",
audio[0].data.cpu(),
global_step=iteration,
sample_rate=hparams.sampling_rate)
def prepare_directories_and_logger(output_directory, log_directory):
if not os.path.isdir(output_directory):
os.makedirs(output_directory)
# os.chmod(output_directory, 0o775)
logger = Tacotron2Logger(os.path.join(output_directory, log_directory))
return logger
