def create_hparams(hparams_string=None, verbose=False, level=2):
"""Create model hyperparameters. Parse nondefault from given string."""
# hparams = tf.contrib.training.HParams(
hparams = Dict2Obj(
dict(
################################
# Experiment Parameters #
################################
epochs=50000,
iters_per_checkpoint=1000, # 500,
seed=1234,
dynamic_loss_scaling=True,
fp16_run=False,
distributed_run=False,
dist_backend="nccl",
dist_url="tcp://localhost:54321",
cudnn_enabled=True,
cudnn_benchmark=False,
ignore_layers=['speaker_embedding.weight'],
################################
# Data Parameters #
################################
train_mode='train-f04',
# f01:用基频；f02:用基频均值填充；f03:用零向量代替基频；f04:不用基频。
# f01,f02,f03的模式都把prenet_f0_dim设为1，f04把prenet_f0_dim设为0。
training_files=
r"F:\github\zhrtvc\data\SV2TTS\mellotron\samples_ssml\train.txt",
# 文件一行记录一个语音信息，每行的数据结构：数据文件夹名\t语音源文件\t文本\t说话人名称\n，样例如下：
# 000000 Aibao/005397.mp3 他走近钢琴并开始演奏“祖国从哪里开始”。 0
validation_files=
r"F:\github\zhrtvc\data\SV2TTS\mellotron\samples_ssml\validation.txt",
# 'filelists/ljs_audiopaths_text_sid_val_filelist.txt',
text_cleaners='ssml', # ['chinese_cleaners'],
p_arpabet=1.0,
cmudict_path=None, # "data/cmu_dictionary",
################################
# Audio Parameters #
################################
max_wav_value=32768.0,
sampling_rate=hparams_griffinlim.sample_rate, # 16000, # 22050,
filter_length=hparams_griffinlim.n_fft, # 1024,
hop_length=hparams_griffinlim.hop_size, # 256,
win_length=hparams_griffinlim.win_size, # 1024,
n_mel_channels=401, # 80,
mel_fmin=0.0,
mel_fmax=8000.0,
f0_min=80,
f0_max=880,
harm_thresh=0.25,
################################
# Model Parameters #
################################
n_symbols=145, # len(symbols),
symbols_embedding_dim=128 * level, # 512,
# Encoder parameters
encoder_kernel_size=5,
encoder_n_convolutions=3,
encoder_embedding_dim=128 * level, # 512,
# Decoder parameters
n_frames_per_step=1, # currently only 1 is supported
decoder_rnn_dim=256 * level, # 1024,
prenet_dim=64 * level, # 256,
prenet_f0_n_layers=1,
prenet_f0_dim=0, # 1, 如果不启用f0，则设置为0。
prenet_f0_kernel_size=1,
prenet_rms_dim=0,
prenet_rms_kernel_size=1,
max_decoder_steps=2000, # 1000,
gate_threshold=0.5,
p_attention_dropout=0.1,
p_decoder_dropout=0.1,
p_teacher_forcing=1.0,
# Attention parameters
attention_rnn_dim=256 * level, # 1024,
attention_dim=32 * level, # 128,
# Location Layer parameters
attention_location_n_filters=8 * level, # 32,
attention_location_kernel_size=31,
# Mel-post processing network parameters
postnet_embedding_dim=128 * level, # 512,
postnet_kernel_size=5,
postnet_n_convolutions=5,
# Speaker embedding
n_speakers=123,
speaker_embedding_dim=16 * level, # 32 * level, # 128,
# Reference encoder
with_gst=False, # True,
ref_enc_filters=[
8 * level, 8 * level, 16 * level, 16 * level, 32 * level,
32 * level
],
# [32, 32, 64, 64, 128, 128],
ref_enc_size=[3, 3],
ref_enc_strides=[2, 2],
ref_enc_pad=[1, 1],
ref_enc_gru_size=32 * level, # 128,
# Style Token Layer
token_embedding_size=
0, # 64 * level, # 256, # 如果with_gst=False，则手动改为0。
token_num=10,
num_heads=8,
################################
# Optimization Hyperparameters #
################################
use_saved_learning_rate=False,
learning_rate=1e-3,
learning_rate_min=1e-5,
learning_rate_anneal=50000,
weight_decay=1e-6,
grad_clip_thresh=1.0,
batch_size=32, # 32,
mask_padding=True, # set model's padded outputs to padded values
))
if hparams_string:
# tf.compat.v1.logging.info('Parsing command line hparams: %s', hparams_string)
hparams.parse(hparams_string)
# if verbose:
# tf.compat.v1.logging.info('Final parsed hparams: %s', hparams.values())
return hparams
def create_hparams(hparams_string=None, verbose=False, level=2):
"""Create model hyperparameters. Parse nondefault from given string."""
# hparams = tf.contrib.training.HParams(
hparams = Dict2Obj(
dict(
################################
# Experiment Parameters #
################################
epochs=50000,
iters_per_checkpoint=500,
seed=1234,
dynamic_loss_scaling=True,
fp16_run=False,
distributed_run=False,
dist_backend="nccl",
dist_url="tcp://localhost:54321",
cudnn_enabled=True,
cudnn_benchmark=False,
ignore_layers=['speaker_embedding.weight'],
################################
# Data Parameters #
################################
training_files=r"F:\github\zhrtvc\data\SV2TTS\mellotron\train.txt",
# 'filelists/ljs_audiopaths_text_sid_train_filelist.txt',
validation_files=
r"F:\github\zhrtvc\data\SV2TTS\mellotron\validation.txt",
# 'filelists/ljs_audiopaths_text_sid_val_filelist.txt',
text_cleaners=['english_cleaners'],
p_arpabet=1.0,
cmudict_path=None, # "data/cmu_dictionary",
################################
# Audio Parameters #
################################
max_wav_value=32768.0,
sampling_rate=22050, # 22050,
filter_length=1024,
hop_length=256,
win_length=1024,
n_mel_channels=80,
mel_fmin=0.0,
mel_fmax=8000.0,
f0_min=80,
f0_max=880,
harm_thresh=0.25,
################################
# Model Parameters #
################################
n_symbols=len(symbols),
symbols_embedding_dim=128 * level, # 512,
# Encoder parameters
encoder_kernel_size=5,
encoder_n_convolutions=3,
encoder_embedding_dim=128 * level, # 512,
# Decoder parameters
n_frames_per_step=1, # currently only 1 is supported
decoder_rnn_dim=256 * level, # 1024,
prenet_dim=64 * level, # 256,
prenet_f0_n_layers=1,
prenet_f0_dim=1,
prenet_f0_kernel_size=1,
prenet_rms_dim=0,
prenet_rms_kernel_size=1,
max_decoder_steps=1000,
gate_threshold=0.5,
p_attention_dropout=0.1,
p_decoder_dropout=0.1,
p_teacher_forcing=1.0,
# Attention parameters
attention_rnn_dim=256 * level, # 1024,
attention_dim=32 * level, # 128,
# Location Layer parameters
attention_location_n_filters=8 * level, # 32,
attention_location_kernel_size=31,
# Mel-post processing network parameters
postnet_embedding_dim=128 * level, # 512,
postnet_kernel_size=5,
postnet_n_convolutions=5,
# Speaker embedding
n_speakers=123,
speaker_embedding_dim=32 * level, # 128,
# Reference encoder
with_gst=True,
ref_enc_filters=[
8 * level, 8 * level, 16 * level, 16 * level, 32 * level,
32 * level
],
# [32, 32, 64, 64, 128, 128],
ref_enc_size=[3, 3],
ref_enc_strides=[2, 2],
ref_enc_pad=[1, 1],
ref_enc_gru_size=32 * level, # 128,
# Style Token Layer
token_embedding_size=64 * level, # 256,
token_num=10,
num_heads=8,
################################
# Optimization Hyperparameters #
################################
use_saved_learning_rate=False,
learning_rate=1e-3,
learning_rate_min=1e-5,
learning_rate_anneal=50000,
weight_decay=1e-6,
grad_clip_thresh=1.0,
batch_size=32, # 32,
mask_padding=True, # set model's padded outputs to padded values
))
# if hparams_string:
# tf.compat.v1.logging.info('Parsing command line hparams: %s', hparams_string)
# hparams.parse(hparams_string)
#
# if verbose:
# tf.compat.v1.logging.info('Final parsed hparams: %s', hparams.values())
return hparams
"win_size": 1024, # 800
"sample_rate": _sr, # 16000
"fmin": 0, # 55
"fmax": _sr // 2, # 7600
"preemphasize": False, # True
'symmetric_mels': True, # True
'signal_normalization': False, # True
'allow_clipping_in_normalization': False, # True
'ref_level_db': 0, # 20
'center': False, # True
'__file__': __file__
}
synthesizer_hparams = {k: v for k, v in default_hparams.items()}
synthesizer_hparams = {**synthesizer_hparams, **my_hp}
synthesizer_hparams = Dict2Obj(synthesizer_hparams)
def audio2mel_synthesizer(src):
"""
用aukit模块重现生成mel，和synthesizer的频谱适应。
:param src:
:return:
"""
_pad_len = (synthesizer_hparams.n_fft - synthesizer_hparams.hop_size) // 2
wavs = src.cpu().numpy()
mels = []
for wav in wavs:
wav = np.pad(wav.flatten(), (_pad_len, _pad_len), mode="reflect")
mel = mel_spectrogram(wav, synthesizer_hparams)
mel = mel / 20
def create_hparams(hparams_string=None, verbose=False, level=2):
"""Create model hyperparameters. Parse nondefault from given string."""
# hparams = tf.contrib.training.HParams(
hparams = Dict2Obj(
dict(
################################
# Experiment Parameters #
################################
dataloader_num_workers=10,
epochs=1000000,
iters_per_checkpoint=1000, # 500,
seed=1234,
dynamic_loss_scaling=True,
fp16_run=False,
distributed_run=False,
dist_backend="nccl",
dist_url="tcp://localhost:54321",
cudnn_enabled=True,
cudnn_benchmark=False,
ignore_layers=['speaker_embedding.weight'],
################################
# Data Parameters #
################################
train_mode='train-mspk',
# f01:用基频，prenet_f0_dim=1。
# f02:用基频均值填充，prenet_f0_dim=1。
# f03:用零向量代替基频，prenet_f0_dim=1。
# f04:不用基频，prenet_f0_dim=0。
# f05s02:用speaker_id等距分配代替基频，speaker_id用0表示，prenet_f0_dim=0。
# f06s02:用语音的embed向量代替，基频speaker_id用0表示，prenet_f0_dim=8。
# gst:用gst模式，把speaker_id用0表示，prenet_f0_dim=0, token_embedding_size=64 * level, with_gst=True。
# tacotron:用tacotron模式，把speaker_id用0表示，prenet_f0_dim=0, token_embedding_size=0, with_gst=False。
# mspk:multispeaker，快捷表示说话人，用speaker的md5的32位16进制数代表说话人，不用基频，encoder_model_fpath='mspk', speaker_embedding_dim=32, n_speakers=0, prenet_f0_dim=0。
# rtvc:利用语音编码向量的语音克隆，用GE2E模型把语音转为256维向量，作为speaker向量输入，不用基频，encoder_model_fpath='fpath', speaker_embedding_dim=32, n_speakers=256, prenet_f0_dim=0。
# training_files=r"../../data/SV2TTS/mellotron/samples_ssml/train.txt",
# 文件一行记录一个语音信息，每行的数据结构：数据文件夹名\t语音源文件\t文本\t说话人名称\n，样例如下：
# 000000 Aibao/005397.mp3 他走近钢琴并开始演奏“祖国从哪里开始”。 0
# validation_files=r"../../data/SV2TTS/mellotron/samples_ssml/validation.txt",
# 'filelists/ljs_audiopaths_text_sid_val_filelist.txt',
encoder_model_fpath=
r'/home/project/zhrtvc/models-gmw/models/encoder/saved_models/ge2e_pretrained.pt',
text_cleaners='hanzi', # ['chinese_cleaners'],
p_arpabet=1.0,
cmudict_path=None, # "data/cmu_dictionary",
################################
# Audio Parameters #
################################
max_wav_value=32768.0,
sampling_rate=
22050, # hparams_griffinlim.sample_rate, # 16000, # 22050,
filter_length=1024, # hparams_griffinlim.n_fft, # 1024,
hop_length=256, # hparams_griffinlim.hop_size, # 256,
win_length=1024, # hparams_griffinlim.win_size, # 1024,
n_mel_channels=80, # 401, # 80,
mel_fmin=0.0,
mel_fmax=8000.0, # 8000.0,
f0_min=80,
f0_max=880,
harm_thresh=0.25,
################################
# Model Parameters #
################################
n_symbols=145, # len(symbols),
symbols_embedding_dim=128 * level, # 512,
# Encoder parameters
encoder_kernel_size=5,
encoder_n_convolutions=3,
encoder_embedding_dim=128 * level, # 512,
# Decoder parameters
n_frames_per_step=1, # currently only 1 is supported
decoder_rnn_dim=256 * level, # 1024,
prenet_dim=64 * level, # 256,
prenet_f0_n_layers=1,
prenet_f0_dim=0, # 1, 如果不启用f0，则设置为0。
prenet_f0_kernel_size=1,
prenet_rms_dim=0,
prenet_rms_kernel_size=1,
max_decoder_steps=1000, # 1000,
gate_threshold=0.5,
p_attention_dropout=0.1,
p_decoder_dropout=0.1,
p_teacher_forcing=1.0,
# Attention parameters
attention_rnn_dim=256 * level, # 1024,
attention_dim=32 * level, # 128,
# Location Layer parameters
attention_location_n_filters=8 * level, # 32,
attention_location_kernel_size=31,
# Mel-post processing network parameters
postnet_embedding_dim=128 * level, # 512,
postnet_kernel_size=5,
postnet_n_convolutions=5,
# Speaker embedding
n_speakers=32, # 1000, # 123
speaker_embedding_dim=32, # 16 * level, # 32 * level, # 128,
# speaker_embedding_dim=64, # 16 * level, # 32 * level, # 128,
# Reference encoder
with_gst=False, # True,
ref_enc_filters=[
8 * level, 8 * level, 16 * level, 16 * level, 32 * level,
32 * level
],
# [32, 32, 64, 64, 128, 128],
ref_enc_size=[3, 3],
ref_enc_strides=[2, 2],
ref_enc_pad=[1, 1],
ref_enc_gru_size=32 * level, # 128,
# Style Token Layer
token_embedding_size=
0, # 64 * level, # 256, # 如果with_gst=False，则手动改为0。
token_num=10,
num_heads=8,
################################
# Optimization Hyperparameters #
################################
use_saved_learning_rate=False,
learning_rate=1e-3,
learning_rate_min=1e-5,
learning_rate_anneal=50000,
weight_decay=1e-6,
grad_clip_thresh=1.0,
batch_size=32, # 32,
mask_padding=True, # set model's padded outputs to padded values
))
if hparams_string:
# tf.compat.v1.logging.info('Parsing command line hparams: %s', hparams_string)
hparams.parse(hparams_string)
# if verbose:
# tf.compat.v1.logging.info('Final parsed hparams: %s', hparams.values())
return hparams
