def get_tacotron2_alignment_test(text_seq):
hparams = hp_tacotron2.create_hparams()
hparams.sampling_rate = hp.sample_rate
checkpoint_path = os.path.join("Tacotron2",
os.path.join("outdir", "checkpoint_51000"))
tacotron2 = train_tacotron2.load_model(hparams)
tacotron2.load_state_dict(torch.load(checkpoint_path)["state_dict"])
_ = tacotron2.cuda().eval().half()
sequence = np.array(text_to_sequence(text_seq,
hp.hparams.text_cleaners))[None, :]
print("sequence size", np.shape(sequence))
sequence = torch.autograd.Variable(
torch.from_numpy(sequence)).cuda().long()
mel, mel_postnet, _, alignment = tacotron2.inference(sequence)
wav = audio.inv_mel_spectrogram(mel_postnet.float().data.cpu().numpy()[0])
file_name = text_seq.replace(" ", "_")
audio.save_wav(wav, "%s.wav" % file_name)
alignment = alignment.float().data.cpu().numpy()[0]
print("alignment size", np.shape(alignment))
get_D(alignment)
return alignment
def get_tacotron2_alignment_test(text_seq):
hparams = hp_tacotron2.create_hparams()
hparams.sampling_rate = hp.sample_rate
checkpoint_path = os.path.join("Tacotron2", os.path.join(
"pre_trained_model", "tacotron2_statedict.pt"))
tacotron2 = train_tacotron2.load_model(hparams)
tacotron2.load_state_dict(torch.load(checkpoint_path)["state_dict"])
_ = tacotron2.cuda().eval().half()
sequence = np.array(text_to_sequence(text_seq, hp.text_cleaners))[None, :]
print("sequence size", np.shape(sequence))
sequence = torch.autograd.Variable(
torch.from_numpy(sequence)).cuda().long()
mel, mel_postnet, _, alignment = tacotron2.inference(sequence)
plot_data((mel.float().data.cpu().numpy()[0],
mel_postnet.float().data.cpu().numpy()[0],
alignment.float().data.cpu().numpy()[0].T))
alignment = alignment.float().data.cpu().numpy()[0]
print("alignment size", np.shape(alignment))
get_D(alignment)
return alignment
def get_tacotron2():
hparams = hp_tacotron2.create_hparams()
hparams.sampling_rate = hp.sample_rate
checkpoint_path = os.path.join("Tacotron2", os.path.join(
"pre_trained_model", "tacotron2_statedict.pt"))
tacotron2 = train_tacotron2.load_model(hparams)
tacotron2.load_state_dict(torch.load(checkpoint_path)["state_dict"])
_ = tacotron2.cuda().eval().half()
return tacotron2
def get_tacotron2_alignment_test(text_seq):
hparams = hp_tacotron2.create_hparams()
hparams.sampling_rate = hp.sample_rate
checkpoint_path = os.path.join("Tacotron2", os.path.join(
"pre_trained_model", "tacotron2_statedict.pt"))
tacotron2 = train_tacotron2.load_model(hparams)
tacotron2.load_state_dict(torch.load(checkpoint_path)["state_dict"])
# print(tacotron2)
_ = tacotron2.cuda().eval().half()
# sequence = text_seq[None, :]
sequence = np.array(text_to_sequence(text_seq, hp.text_cleaners))[None, :]
# zeros_pad = np.zeros([1, 12])
# sequence = np.concatenate((sequence, zeros_pad), axis=1)
print("sequence size", np.shape(sequence))
sequence = torch.autograd.Variable(
torch.from_numpy(sequence)).cuda().long()
# print(sequence.size())
# sequence = torch.autograd.Variable(
# torch.from_numpy(text_seq).cuda().long())
# sequence = torch.autograd.Variable(text_seq)
mel, mel_postnet, _, alignment = tacotron2.inference(sequence)
# print(sequence.size())
# print(alignment.size())
# Test Mel
# print(mel_postnet.size())
# plot_data((mel.float().data.cpu().numpy()[0],
# mel_postnet.float().data.cpu().numpy()[0],
# alignment.float().data.cpu().numpy()[0].T))
# mel_postnet = mel_postnet.float().data.cpu().numpy()[0].T
# wav = audio.inv_mel_spectrogram(mel_postnet)
# audio.save_wav(wav, "test.wav")
# plot_data((mel_postnet))
# print(alignment.size())
alignment = alignment.float().data.cpu().numpy()[0]
print("alignment size", np.shape(alignment))
# print(alignment)
get_D(alignment)
return alignment
def loadTacotron2(self):
checkpoint_path = os.path.join("TrainedModels", "MelGenerator",
"tacotron2_cassie_lee_morris_80.pt")
hparams = create_hparams()
hparams.sampling_rate = self.tacotron2SampleRate
hparams.max_decoder_steps = 3000
hparams.gate_threshold = 0.25
self.tacotron2 = load_model(hparams)
self.tacotron2.load_state_dict(
torch.load(checkpoint_path,
map_location=self.deviceType)['state_dict'])
if (self.deviceType.type == "cpu"):
self.tacotron2.cpu().eval()
else:
self.tacotron2.cuda().eval()
help='number of gpus')
parser.add_argument('--rank',
type=int,
default=0,
required=False,
help='rank of current gpu')
parser.add_argument('--group_name',
type=str,
default='group_name',
required=False,
help='Distributed group name')
parser.add_argument('--hparams',
type=str,
required=False,
help='comma separated name=value pairs')
args = parser.parse_args()
hparams = create_hparams(args.hparams)
torch.backends.cudnn.enabled = hparams.cudnn_enabled
torch.backends.cudnn.benchmark = hparams.cudnn_benchmark
print("FP16 Run:", hparams.fp16_run)
print("Dynamic Loss Scaling:", hparams.dynamic_loss_scaling)
print("Distributed Run:", hparams.distributed_run)
print("cuDNN Enabled:", hparams.cudnn_enabled)
print("cuDNN Benchmark:", hparams.cudnn_benchmark)
train(args.output_directory, args.log_directory, args.checkpoint_path,
args.warm_start, args.n_gpus, args.rank, args.group_name, hparams)
help='number of gpus')
parser.add_argument('--rank',
type=int,
default=0,
required=False,
help='rank of current gpu')
parser.add_argument('--group_name',
type=str,
default='group_name',
required=False,
help='Distributed group name')
parser.add_argument('--hparams',
type=str,
required=False,
help='comma separated name=value pairs')
args = parser.parse_args()
hparams = create_hparams()
torch.backends.cudnn.enabled = hparams.cudnn_enabled
torch.backends.cudnn.benchmark = hparams.cudnn_benchmark
print("FP16 Run:", hparams.fp16_run)
print("Dynamic Loss Scaling:", hparams.dynamic_loss_scaling)
print("Distributed Run:", hparams.distributed_run)
print("cuDNN Enabled:", hparams.cudnn_enabled)
print("cuDNN Benchmark:", hparams.cudnn_benchmark)
train(args.output_directory, args.log_directory, args.checkpoint_path,
args.warm_start, args.n_gpus, args.rank, args.group_name, hparams)
