def train(args):
"""train."""
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if torch.cuda.is_available() and args.auto_select_gpu is True:
cvd = use_single_gpu()
logging.info(f"GPU {cvd} is used")
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# torch.backends.cudnn.enabled = False
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
elif torch.cuda.is_available() and args.auto_select_gpu is False:
torch.cuda.set_device(args.gpu_id)
logging.info(f"GPU {args.gpu_id} is used")
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# torch.backends.cudnn.enabled = False
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
else:
device = torch.device("cpu")
logging.info("Warning: CPU is used")
train_set = SVSDataset(
align_root_path=args.train_align,
pitch_beat_root_path=args.train_pitch,
wav_root_path=args.train_wav,
char_max_len=args.char_max_len,
max_len=args.num_frames,
sr=args.sampling_rate,
preemphasis=args.preemphasis,
nfft=args.nfft,
frame_shift=args.frame_shift,
frame_length=args.frame_length,
n_mels=args.n_mels,
power=args.power,
max_db=args.max_db,
ref_db=args.ref_db,
sing_quality=args.sing_quality,
standard=args.standard,
db_joint=args.db_joint,
Hz2semitone=args.Hz2semitone,
semitone_min=args.semitone_min,
semitone_max=args.semitone_max,
phone_shift_size=args.phone_shift_size,
semitone_shift=args.semitone_shift,
)
dev_set = SVSDataset(
align_root_path=args.val_align,
pitch_beat_root_path=args.val_pitch,
wav_root_path=args.val_wav,
char_max_len=args.char_max_len,
max_len=args.num_frames,
sr=args.sampling_rate,
preemphasis=args.preemphasis,
nfft=args.nfft,
frame_shift=args.frame_shift,
frame_length=args.frame_length,
n_mels=args.n_mels,
power=args.power,
max_db=args.max_db,
ref_db=args.ref_db,
sing_quality=args.sing_quality,
standard=args.standard,
db_joint=args.db_joint,
Hz2semitone=args.Hz2semitone,
semitone_min=args.semitone_min,
semitone_max=args.semitone_max,
phone_shift_size=-1,
semitone_shift=False,
)
collate_fn_svs_train = SVSCollator(
args.num_frames,
args.char_max_len,
args.use_asr_post,
args.phone_size,
args.n_mels,
args.db_joint,
args.random_crop,
args.crop_min_length,
args.Hz2semitone,
)
collate_fn_svs_val = SVSCollator(
args.num_frames,
args.char_max_len,
args.use_asr_post,
args.phone_size,
args.n_mels,
args.db_joint,
False, # random crop
-1, # crop_min_length
args.Hz2semitone,
)
train_loader = torch.utils.data.DataLoader(
dataset=train_set,
batch_size=args.batchsize,
shuffle=True,
num_workers=args.num_workers,
collate_fn=collate_fn_svs_train,
pin_memory=True,
)
dev_loader = torch.utils.data.DataLoader(
dataset=dev_set,
batch_size=args.batchsize,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn_svs_val,
pin_memory=True,
)
assert (args.feat_dim == dev_set[0]["spec"].shape[1]
or args.feat_dim == dev_set[0]["mel"].shape[1])
if args.collect_stats:
collect_stats(train_loader, args)
logging.info("collect_stats finished !")
quit()
# prepare model
if args.model_type == "GLU_Transformer":
if args.db_joint:
model = GLU_TransformerSVS_combine(
phone_size=args.phone_size,
singer_size=args.singer_size,
embed_size=args.embedding_size,
hidden_size=args.hidden_size,
glu_num_layers=args.glu_num_layers,
dropout=args.dropout,
output_dim=args.feat_dim,
dec_nhead=args.dec_nhead,
dec_num_block=args.dec_num_block,
n_mels=args.n_mels,
double_mel_loss=args.double_mel_loss,
local_gaussian=args.local_gaussian,
Hz2semitone=args.Hz2semitone,
semitone_size=args.semitone_size,
device=device,
)
else:
model = GLU_TransformerSVS(
phone_size=args.phone_size,
embed_size=args.embedding_size,
hidden_size=args.hidden_size,
glu_num_layers=args.glu_num_layers,
dropout=args.dropout,
output_dim=args.feat_dim,
dec_nhead=args.dec_nhead,
dec_num_block=args.dec_num_block,
n_mels=args.n_mels,
double_mel_loss=args.double_mel_loss,
local_gaussian=args.local_gaussian,
Hz2semitone=args.Hz2semitone,
semitone_size=args.semitone_size,
device=device,
)
elif args.model_type == "LSTM":
if args.db_joint:
model = LSTMSVS_combine(
phone_size=args.phone_size,
singer_size=args.singer_size,
embed_size=args.embedding_size,
d_model=args.hidden_size,
num_layers=args.num_rnn_layers,
dropout=args.dropout,
d_output=args.feat_dim,
n_mels=args.n_mels,
double_mel_loss=args.double_mel_loss,
Hz2semitone=args.Hz2semitone,
semitone_size=args.semitone_size,
device=device,
use_asr_post=args.use_asr_post,
)
else:
model = LSTMSVS(
phone_size=args.phone_size,
embed_size=args.embedding_size,
d_model=args.hidden_size,
num_layers=args.num_rnn_layers,
dropout=args.dropout,
d_output=args.feat_dim,
n_mels=args.n_mels,
double_mel_loss=args.double_mel_loss,
Hz2semitone=args.Hz2semitone,
semitone_size=args.semitone_size,
device=device,
use_asr_post=args.use_asr_post,
)
elif args.model_type == "GRU_gs":
model = GRUSVS_gs(
phone_size=args.phone_size,
embed_size=args.embedding_size,
d_model=args.hidden_size,
num_layers=args.num_rnn_layers,
dropout=args.dropout,
d_output=args.feat_dim,
n_mels=args.n_mels,
device=device,
use_asr_post=args.use_asr_post,
)
elif args.model_type == "PureTransformer":
model = TransformerSVS(
phone_size=args.phone_size,
embed_size=args.embedding_size,
hidden_size=args.hidden_size,
glu_num_layers=args.glu_num_layers,
dropout=args.dropout,
output_dim=args.feat_dim,
dec_nhead=args.dec_nhead,
dec_num_block=args.dec_num_block,
n_mels=args.n_mels,
double_mel_loss=args.double_mel_loss,
local_gaussian=args.local_gaussian,
device=device,
)
elif args.model_type == "Conformer":
model = ConformerSVS(
phone_size=args.phone_size,
embed_size=args.embedding_size,
enc_attention_dim=args.enc_attention_dim,
enc_attention_heads=args.enc_attention_heads,
enc_linear_units=args.enc_linear_units,
enc_num_blocks=args.enc_num_blocks,
enc_dropout_rate=args.enc_dropout_rate,
enc_positional_dropout_rate=args.enc_positional_dropout_rate,
enc_attention_dropout_rate=args.enc_attention_dropout_rate,
enc_input_layer=args.enc_input_layer,
enc_normalize_before=args.enc_normalize_before,
enc_concat_after=args.enc_concat_after,
enc_positionwise_layer_type=args.enc_positionwise_layer_type,
enc_positionwise_conv_kernel_size=(
args.enc_positionwise_conv_kernel_size),
enc_macaron_style=args.enc_macaron_style,
enc_pos_enc_layer_type=args.enc_pos_enc_layer_type,
enc_selfattention_layer_type=args.enc_selfattention_layer_type,
enc_activation_type=args.enc_activation_type,
enc_use_cnn_module=args.enc_use_cnn_module,
enc_cnn_module_kernel=args.enc_cnn_module_kernel,
enc_padding_idx=args.enc_padding_idx,
output_dim=args.feat_dim,
dec_nhead=args.dec_nhead,
dec_num_block=args.dec_num_block,
n_mels=args.n_mels,
double_mel_loss=args.double_mel_loss,
local_gaussian=args.local_gaussian,
dec_dropout=args.dec_dropout,
Hz2semitone=args.Hz2semitone,
semitone_size=args.semitone_size,
device=device,
)
elif args.model_type == "Comformer_full":
if args.db_joint:
model = ConformerSVS_FULL_combine(
phone_size=args.phone_size,
singer_size=args.singer_size,
embed_size=args.embedding_size,
output_dim=args.feat_dim,
n_mels=args.n_mels,
enc_attention_dim=args.enc_attention_dim,
enc_attention_heads=args.enc_attention_heads,
enc_linear_units=args.enc_linear_units,
enc_num_blocks=args.enc_num_blocks,
enc_dropout_rate=args.enc_dropout_rate,
enc_positional_dropout_rate=args.enc_positional_dropout_rate,
enc_attention_dropout_rate=args.enc_attention_dropout_rate,
enc_input_layer=args.enc_input_layer,
enc_normalize_before=args.enc_normalize_before,
enc_concat_after=args.enc_concat_after,
enc_positionwise_layer_type=args.enc_positionwise_layer_type,
enc_positionwise_conv_kernel_size=(
args.enc_positionwise_conv_kernel_size),
enc_macaron_style=args.enc_macaron_style,
enc_pos_enc_layer_type=args.enc_pos_enc_layer_type,
enc_selfattention_layer_type=args.enc_selfattention_layer_type,
enc_activation_type=args.enc_activation_type,
enc_use_cnn_module=args.enc_use_cnn_module,
enc_cnn_module_kernel=args.enc_cnn_module_kernel,
enc_padding_idx=args.enc_padding_idx,
dec_attention_dim=args.dec_attention_dim,
dec_attention_heads=args.dec_attention_heads,
dec_linear_units=args.dec_linear_units,
dec_num_blocks=args.dec_num_blocks,
dec_dropout_rate=args.dec_dropout_rate,
dec_positional_dropout_rate=args.dec_positional_dropout_rate,
dec_attention_dropout_rate=args.dec_attention_dropout_rate,
dec_input_layer=args.dec_input_layer,
dec_normalize_before=args.dec_normalize_before,
dec_concat_after=args.dec_concat_after,
dec_positionwise_layer_type=args.dec_positionwise_layer_type,
dec_positionwise_conv_kernel_size=(
args.dec_positionwise_conv_kernel_size),
dec_macaron_style=args.dec_macaron_style,
dec_pos_enc_layer_type=args.dec_pos_enc_layer_type,
dec_selfattention_layer_type=args.dec_selfattention_layer_type,
dec_activation_type=args.dec_activation_type,
dec_use_cnn_module=args.dec_use_cnn_module,
dec_cnn_module_kernel=args.dec_cnn_module_kernel,
dec_padding_idx=args.dec_padding_idx,
Hz2semitone=args.Hz2semitone,
semitone_size=args.semitone_size,
device=device,
)
else:
model = ConformerSVS_FULL(
phone_size=args.phone_size,
embed_size=args.embedding_size,
output_dim=args.feat_dim,
n_mels=args.n_mels,
enc_attention_dim=args.enc_attention_dim,
enc_attention_heads=args.enc_attention_heads,
enc_linear_units=args.enc_linear_units,
enc_num_blocks=args.enc_num_blocks,
enc_dropout_rate=args.enc_dropout_rate,
enc_positional_dropout_rate=args.enc_positional_dropout_rate,
enc_attention_dropout_rate=args.enc_attention_dropout_rate,
enc_input_layer=args.enc_input_layer,
enc_normalize_before=args.enc_normalize_before,
enc_concat_after=args.enc_concat_after,
enc_positionwise_layer_type=args.enc_positionwise_layer_type,
enc_positionwise_conv_kernel_size=(
args.enc_positionwise_conv_kernel_size),
enc_macaron_style=args.enc_macaron_style,
enc_pos_enc_layer_type=args.enc_pos_enc_layer_type,
enc_selfattention_layer_type=args.enc_selfattention_layer_type,
enc_activation_type=args.enc_activation_type,
enc_use_cnn_module=args.enc_use_cnn_module,
enc_cnn_module_kernel=args.enc_cnn_module_kernel,
enc_padding_idx=args.enc_padding_idx,
dec_attention_dim=args.dec_attention_dim,
dec_attention_heads=args.dec_attention_heads,
dec_linear_units=args.dec_linear_units,
dec_num_blocks=args.dec_num_blocks,
dec_dropout_rate=args.dec_dropout_rate,
dec_positional_dropout_rate=args.dec_positional_dropout_rate,
dec_attention_dropout_rate=args.dec_attention_dropout_rate,
dec_input_layer=args.dec_input_layer,
dec_normalize_before=args.dec_normalize_before,
dec_concat_after=args.dec_concat_after,
dec_positionwise_layer_type=args.dec_positionwise_layer_type,
dec_positionwise_conv_kernel_size=(
args.dec_positionwise_conv_kernel_size),
dec_macaron_style=args.dec_macaron_style,
dec_pos_enc_layer_type=args.dec_pos_enc_layer_type,
dec_selfattention_layer_type=args.dec_selfattention_layer_type,
dec_activation_type=args.dec_activation_type,
dec_use_cnn_module=args.dec_use_cnn_module,
dec_cnn_module_kernel=args.dec_cnn_module_kernel,
dec_padding_idx=args.dec_padding_idx,
Hz2semitone=args.Hz2semitone,
semitone_size=args.semitone_size,
device=device,
)
elif args.model_type == "USTC_DAR":
model = USTC_SVS(
phone_size=args.phone_size,
embed_size=args.embedding_size,
middle_dim_fc=args.middle_dim_fc,
output_dim=args.feat_dim,
multi_history_num=args.multi_history_num,
middle_dim_prenet=args.middle_dim_prenet,
n_blocks_prenet=args.n_blocks_prenet,
n_heads_prenet=args.n_heads_prenet,
kernel_size_prenet=args.kernel_size_prenet,
bi_d_model=args.bi_d_model,
bi_num_layers=args.bi_num_layers,
uni_d_model=args.uni_d_model,
uni_num_layers=args.uni_num_layers,
dropout=args.dropout,
feedbackLink_drop_rate=args.feedbackLink_drop_rate,
device=device,
)
else:
raise ValueError("Not Support Model Type %s" % args.model_type)
logging.info(f"{model}")
model = model.to(device)
logging.info(
f"The model has {count_parameters(model):,} trainable parameters")
model_load_dir = ""
pretrain_encoder_dir = ""
start_epoch = 1 # FIX ME
if args.pretrain_encoder != "":
pretrain_encoder_dir = args.pretrain_encoder
if args.initmodel != "":
model_load_dir = args.initmodel
if args.resume and os.path.exists(args.model_save_dir):
checks = os.listdir(args.model_save_dir)
start_epoch = max(
list(
map(
lambda x: int(x.split(".")[0].split("_")[-1])
if x.endswith("pth.tar") else -1,
checks,
)))
model_temp_load_dir = "{}/epoch_loss_{}.pth.tar".format(
args.model_save_dir, start_epoch)
if start_epoch < 0:
model_load_dir = ""
elif os.path.isfile(model_temp_load_dir):
model_load_dir = model_temp_load_dir
else:
model_load_dir = "{}/epoch_spec_loss_{}.pth.tar".format(
args.model_save_dir, start_epoch)
# load encoder parm from Transformer-TTS
if pretrain_encoder_dir != "":
pretrain = torch.load(pretrain_encoder_dir, map_location=device)
pretrain_dict = pretrain["model"]
model_dict = model.state_dict()
state_dict_new = {}
para_list = []
i = 0
for k, v in pretrain_dict.items():
k_new = k[7:]
if (k_new in model_dict
and model_dict[k_new].size() == pretrain_dict[k].size()):
i += 1
state_dict_new[k_new] = v
model_dict.update(state_dict_new)
model.load_state_dict(model_dict)
logging.info(f"Load {i} layers total. Load pretrain encoder success !")
# load weights for pre-trained model
if model_load_dir != "":
logging.info(f"Model Start to Load, dir: {model_load_dir}")
model_load = torch.load(model_load_dir, map_location=device)
loading_dict = model_load["state_dict"]
model_dict = model.state_dict()
state_dict_new = {}
para_list = []
for k, v in loading_dict.items():
# assert k in model_dict
if (k == "normalizer.mean" or k == "normalizer.std"
or k == "mel_normalizer.mean"
or k == "mel_normalizer.std"):
continue
if model_dict[k].size() == loading_dict[k].size():
state_dict_new[k] = v
else:
para_list.append(k)
logging.info(f"Total {len(loading_dict)} parameter sets, "
f"Loaded {len(state_dict_new)} parameter sets")
if len(para_list) > 0:
logging.warning("Not loading {} because of different sizes".format(
", ".join(para_list)))
model_dict.update(state_dict_new)
model.load_state_dict(model_dict)
logging.info(f"Loaded checkpoint {args.initmodel}")
# setup optimizer
if args.optimizer == "noam":
optimizer = ScheduledOptim(
torch.optim.Adam(model.parameters(),
lr=args.lr,
betas=(0.9, 0.98),
eps=1e-09),
args.hidden_size,
args.noam_warmup_steps,
args.noam_scale,
)
elif args.optimizer == "adam":
optimizer = torch.optim.Adam(model.parameters(),
lr=args.lr,
betas=(0.9, 0.98),
eps=1e-09)
if args.scheduler == "OneCycleLR":
scheduler = torch.optim.lr_scheduler.OneCycleLR(
optimizer,
max_lr=args.lr,
steps_per_epoch=len(train_loader),
epochs=args.max_epochs,
)
elif args.scheduler == "ReduceLROnPlateau":
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
optimizer, "min", verbose=True, patience=10, factor=0.5)
elif args.scheduler == "ExponentialLR":
scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer,
verbose=True,
gamma=0.9886)
else:
raise ValueError("Not Support Optimizer")
# Setup tensorborad logger
if args.use_tfboard:
from tensorboardX import SummaryWriter
logger = SummaryWriter("{}/log".format(args.model_save_dir))
else:
logger = None
if args.loss == "l1":
loss = MaskedLoss("l1", mask_free=args.mask_free)
elif args.loss == "mse":
loss = MaskedLoss("mse", mask_free=args.mask_free)
else:
raise ValueError("Not Support Loss Type")
if args.perceptual_loss > 0:
win_length = int(args.sampling_rate * args.frame_length)
psd_dict, bark_num = cal_psd2bark_dict(fs=args.sampling_rate,
win_len=win_length)
sf = cal_spread_function(bark_num)
loss_perceptual_entropy = PerceptualEntropy(bark_num, sf,
args.sampling_rate,
win_length, psd_dict)
else:
loss_perceptual_entropy = None
# Training
total_loss_epoch_to_save = {}
total_loss_counter = 0
spec_loss_epoch_to_save = {}
spec_loss_counter = 0
total_learning_step = 0
# args.num_saved_model = 5
# preload vocoder model
voc_model = []
if args.vocoder_category == "wavernn":
logging.info("load voc_model from {}".format(args.wavernn_voc_model))
voc_model = WaveRNN(
rnn_dims=args.voc_rnn_dims,
fc_dims=args.voc_fc_dims,
bits=args.voc_bits,
pad=args.voc_pad,
upsample_factors=(
args.voc_upsample_factors_0,
args.voc_upsample_factors_1,
args.voc_upsample_factors_2,
),
feat_dims=args.n_mels,
compute_dims=args.voc_compute_dims,
res_out_dims=args.voc_res_out_dims,
res_blocks=args.voc_res_blocks,
hop_length=args.hop_length,
sample_rate=args.sampling_rate,
mode=args.voc_mode,
).to(device)
voc_model.load(args.wavernn_voc_model)
for epoch in range(start_epoch + 1, 1 + args.max_epochs):
"""Train Stage"""
start_t_train = time.time()
train_info = train_one_epoch(
train_loader,
model,
device,
optimizer,
loss,
loss_perceptual_entropy,
epoch,
args,
voc_model,
)
end_t_train = time.time()
out_log = "Train epoch: {:04d}, ".format(epoch)
if args.optimizer == "noam":
out_log += "lr: {:.6f}, ".format(
optimizer._optimizer.param_groups[0]["lr"])
elif args.optimizer == "adam":
out_log += "lr: {:.6f}, ".format(optimizer.param_groups[0]["lr"])
if args.vocoder_category == "wavernn":
out_log += "loss: {:.4f} ".format(train_info["loss"])
else:
out_log += "loss: {:.4f}, spec_loss: {:.4f} ".format(
train_info["loss"], train_info["spec_loss"])
if args.n_mels > 0:
out_log += "mel_loss: {:.4f}, ".format(train_info["mel_loss"])
if args.perceptual_loss > 0:
out_log += "pe_loss: {:.4f}, ".format(train_info["pe_loss"])
logging.info("{} time: {:.2f}s".format(out_log,
end_t_train - start_t_train))
"""Dev Stage"""
torch.backends.cudnn.enabled = False # 莫名的bug，关掉才可以跑
# start_t_dev = time.time()
dev_info = validate(
dev_loader,
model,
device,
loss,
loss_perceptual_entropy,
epoch,
args,
voc_model,
)
end_t_dev = time.time()
dev_log = "Dev epoch: {:04d}, loss: {:.4f}, spec_loss: {:.4f}, ".format(
epoch, dev_info["loss"], dev_info["spec_loss"])
dev_log += "mcd_value: {:.4f}, ".format(dev_info["mcd_value"])
if args.n_mels > 0:
dev_log += "mel_loss: {:.4f}, ".format(dev_info["mel_loss"])
if args.perceptual_loss > 0:
dev_log += "pe_loss: {:.4f}, ".format(dev_info["pe_loss"])
logging.info("{} time: {:.2f}s".format(dev_log,
end_t_dev - start_t_train))
sys.stdout.flush()
torch.backends.cudnn.enabled = True
if args.scheduler == "OneCycleLR":
scheduler.step()
elif args.scheduler == "ReduceLROnPlateau":
scheduler.step(dev_info["loss"])
elif args.scheduler == "ExponentialLR":
before = total_learning_step // args.lr_decay_learning_steps
total_learning_step += len(train_loader)
after = total_learning_step // args.lr_decay_learning_steps
if after > before: # decay per 250 learning steps
scheduler.step()
"""Save model Stage"""
if not os.path.exists(args.model_save_dir):
os.makedirs(args.model_save_dir)
(total_loss_counter, total_loss_epoch_to_save) = Auto_save_model(
args,
epoch,
model,
optimizer,
train_info,
dev_info,
logger,
total_loss_counter,
total_loss_epoch_to_save,
save_loss_select="loss",
)
if (dev_info["spec_loss"] !=
0): # spec_loss 有意义时再存模型，比如 USTC DAR model 不需要计算线性谱spec loss
(spec_loss_counter, spec_loss_epoch_to_save) = Auto_save_model(
args,
epoch,
model,
optimizer,
train_info,
dev_info,
logger,
spec_loss_counter,
spec_loss_epoch_to_save,
save_loss_select="spec_loss",
)
if args.use_tfboard:
logger.close()
def train_joint(args):
"""train_joint."""
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if torch.cuda.is_available() and args.auto_select_gpu is True:
cvd = use_single_gpu()
logging.info(f"GPU {cvd} is used")
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# torch.backends.cudnn.enabled = False
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
elif torch.cuda.is_available() and args.auto_select_gpu is False:
torch.cuda.set_device(args.gpu_id)
logging.info(f"GPU {args.gpu_id} is used")
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# torch.backends.cudnn.enabled = False
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
else:
device = torch.device("cpu")
logging.info("Warning: CPU is used")
train_set = SVSDataset(
align_root_path=args.train_align,
pitch_beat_root_path=args.train_pitch,
wav_root_path=args.train_wav,
char_max_len=args.char_max_len,
max_len=args.num_frames,
sr=args.sampling_rate,
preemphasis=args.preemphasis,
nfft=args.nfft,
frame_shift=args.frame_shift,
frame_length=args.frame_length,
n_mels=args.n_mels,
power=args.power,
max_db=args.max_db,
ref_db=args.ref_db,
sing_quality=args.sing_quality,
standard=args.standard,
db_joint=args.db_joint,
Hz2semitone=args.Hz2semitone,
semitone_min=args.semitone_min,
semitone_max=args.semitone_max,
phone_shift_size=-1,
semitone_shift=False,
)
dev_set = SVSDataset(
align_root_path=args.val_align,
pitch_beat_root_path=args.val_pitch,
wav_root_path=args.val_wav,
char_max_len=args.char_max_len,
max_len=args.num_frames,
sr=args.sampling_rate,
preemphasis=args.preemphasis,
nfft=args.nfft,
frame_shift=args.frame_shift,
frame_length=args.frame_length,
n_mels=args.n_mels,
power=args.power,
max_db=args.max_db,
ref_db=args.ref_db,
sing_quality=args.sing_quality,
standard=args.standard,
db_joint=args.db_joint,
Hz2semitone=args.Hz2semitone,
semitone_min=args.semitone_min,
semitone_max=args.semitone_max,
phone_shift_size=-1,
semitone_shift=False,
)
collate_fn_svs_train = SVSCollator(
args.num_frames,
args.char_max_len,
args.use_asr_post,
args.phone_size,
args.n_mels,
args.db_joint,
args.random_crop,
args.crop_min_length,
args.Hz2semitone,
)
collate_fn_svs_val = SVSCollator(
args.num_frames,
args.char_max_len,
args.use_asr_post,
args.phone_size,
args.n_mels,
args.db_joint,
False, # random crop
-1, # crop_min_length
args.Hz2semitone,
)
train_loader = torch.utils.data.DataLoader(
dataset=train_set,
batch_size=args.batchsize,
shuffle=True,
num_workers=args.num_workers,
collate_fn=collate_fn_svs_train,
pin_memory=True,
)
dev_loader = torch.utils.data.DataLoader(
dataset=dev_set,
batch_size=args.batchsize,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn_svs_val,
pin_memory=True,
)
assert (args.feat_dim == dev_set[0]["spec"].shape[1]
or args.feat_dim == dev_set[0]["mel"].shape[1])
if args.collect_stats:
collect_stats(train_loader, args)
logging.info("collect_stats finished !")
quit()
# init model_generate
if args.model_type == "GLU_Transformer":
if args.db_joint:
model_generate = GLU_TransformerSVS_combine(
phone_size=args.phone_size,
singer_size=args.singer_size,
embed_size=args.embedding_size,
hidden_size=args.hidden_size,
glu_num_layers=args.glu_num_layers,
dropout=args.dropout,
output_dim=args.feat_dim,
dec_nhead=args.dec_nhead,
dec_num_block=args.dec_num_block,
n_mels=args.n_mels,
double_mel_loss=args.double_mel_loss,
local_gaussian=args.local_gaussian,
Hz2semitone=args.Hz2semitone,
semitone_size=args.semitone_size,
device=device,
)
else:
model_generate = GLU_TransformerSVS(
phone_size=args.phone_size,
embed_size=args.embedding_size,
hidden_size=args.hidden_size,
glu_num_layers=args.glu_num_layers,
dropout=args.dropout,
output_dim=args.feat_dim,
dec_nhead=args.dec_nhead,
dec_num_block=args.dec_num_block,
n_mels=args.n_mels,
double_mel_loss=args.double_mel_loss,
local_gaussian=args.local_gaussian,
Hz2semitone=args.Hz2semitone,
semitone_size=args.semitone_size,
device=device,
)
elif args.model_type == "LSTM":
if args.db_joint:
model_generate = LSTMSVS_combine(
phone_size=args.phone_size,
singer_size=args.singer_size,
embed_size=args.embedding_size,
d_model=args.hidden_size,
num_layers=args.num_rnn_layers,
dropout=args.dropout,
d_output=args.feat_dim,
n_mels=args.n_mels,
double_mel_loss=args.double_mel_loss,
Hz2semitone=args.Hz2semitone,
semitone_size=args.semitone_size,
device=device,
use_asr_post=args.use_asr_post,
)
else:
model_generate = LSTMSVS(
phone_size=args.phone_size,
embed_size=args.embedding_size,
d_model=args.hidden_size,
num_layers=args.num_rnn_layers,
dropout=args.dropout,
d_output=args.feat_dim,
n_mels=args.n_mels,
double_mel_loss=args.double_mel_loss,
Hz2semitone=args.Hz2semitone,
semitone_size=args.semitone_size,
device=device,
use_asr_post=args.use_asr_post,
)
# init model_predict
model_predict = RNN_Discriminator(
embed_size=128,
d_model=128,
hidden_size=128,
num_layers=2,
n_specs=1025,
singer_size=7,
phone_size=43,
simitone_size=59,
dropout=0.1,
bidirectional=True,
device=device,
)
logging.info(f"*********** model_generate ***********")
logging.info(f"{model_generate}")
logging.info(
f"The model has {count_parameters(model_generate):,} trainable parameters"
)
logging.info(f"*********** model_predict ***********")
logging.info(f"{model_predict}")
logging.info(
f"The model has {count_parameters(model_predict):,} trainable parameters"
)
model_generate = load_model_weights(args.initmodel_generator,
model_generate, device)
model_predict = load_model_weights(args.initmodel_predictor, model_predict,
device)
model = Joint_generator_predictor(model_generate, model_predict)
# setup optimizer
if args.optimizer == "noam":
optimizer = ScheduledOptim(
torch.optim.Adam(model.parameters(),
lr=args.lr,
betas=(0.9, 0.98),
eps=1e-09),
args.hidden_size,
args.noam_warmup_steps,
args.noam_scale,
)
elif args.optimizer == "adam":
optimizer = torch.optim.Adam(model.parameters(),
lr=args.lr,
betas=(0.9, 0.98),
eps=1e-09)
if args.scheduler == "OneCycleLR":
scheduler = torch.optim.lr_scheduler.OneCycleLR(
optimizer,
max_lr=args.lr,
steps_per_epoch=len(train_loader),
epochs=args.max_epochs,
)
elif args.scheduler == "ReduceLROnPlateau":
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
optimizer, "min", verbose=True, patience=10, factor=0.5)
elif args.scheduler == "ExponentialLR":
scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer,
verbose=True,
gamma=0.9886)
else:
raise ValueError("Not Support Optimizer")
# setup loss function
loss_predict = nn.CrossEntropyLoss(reduction="sum")
if args.loss == "l1":
loss_generate = MaskedLoss("l1", mask_free=args.mask_free)
elif args.loss == "mse":
loss_generate = MaskedLoss("mse", mask_free=args.mask_free)
else:
raise ValueError("Not Support Loss Type")
if args.perceptual_loss > 0:
win_length = int(args.sampling_rate * args.frame_length)
psd_dict, bark_num = cal_psd2bark_dict(fs=args.sampling_rate,
win_len=win_length)
sf = cal_spread_function(bark_num)
loss_perceptual_entropy = PerceptualEntropy(bark_num, sf,
args.sampling_rate,
win_length, psd_dict)
else:
loss_perceptual_entropy = None
# Training
generator_loss_epoch_to_save = {}
generator_loss_counter = 0
spec_loss_epoch_to_save = {}
spec_loss_counter = 0
predictor_loss_epoch_to_save = {}
predictor_loss_counter = 0
for epoch in range(0, 1 + args.max_epochs):
"""Train Stage"""
start_t_train = time.time()
train_info = train_one_epoch_joint(
train_loader,
model,
device,
optimizer,
loss_generate,
loss_predict,
loss_perceptual_entropy,
epoch,
args,
)
end_t_train = time.time()
# Print Total info
out_log = "Train epoch: {:04d} ".format(epoch)
if args.optimizer == "noam":
out_log += "lr: {:.6f}, \n\t".format(
optimizer._optimizer.param_groups[0]["lr"])
elif args.optimizer == "adam":
out_log += "lr: {:.6f}, \n\t".format(
optimizer.param_groups[0]["lr"])
out_log += "total_loss: {:.4f} \n\t".format(train_info["loss"])
# Print Generator info
if args.vocoder_category == "wavernn":
out_log += "generator_loss: {:.4f} ".format(
train_info["generator_loss"])
else:
out_log += "generator_loss: {:.4f}, spec_loss: {:.4f} ".format(
train_info["generator_loss"], train_info["spec_loss"])
if args.n_mels > 0:
out_log += "mel_loss: {:.4f}, ".format(train_info["mel_loss"])
if args.perceptual_loss > 0:
out_log += "pe_loss: {:.4f}\n\t".format(train_info["pe_loss"])
# Print Predictor info
out_log += "predictor_loss: {:.4f}, singer_loss: {:.4f}, ".format(
train_info["predictor_loss"],
train_info["singer_loss"],
)
out_log += "phone_loss: {:.4f}, semitone_loss: {:.4f} \n\t\t".format(
train_info["phone_loss"],
train_info["semitone_loss"],
)
out_log += "singer_accuracy: {:.4f}%, ".format(
train_info["singer_accuracy"] * 100, )
out_log += "phone_accuracy: {:.4f}%, semitone_accuracy: {:.4f}% ".format(
train_info["phone_accuracy"] * 100,
train_info["semitone_accuracy"] * 100,
)
logging.info("{} time: {:.2f}s".format(out_log,
end_t_train - start_t_train))
"""Dev Stage"""
torch.backends.cudnn.enabled = False # 莫名的bug，关掉才可以跑
# start_t_dev = time.time()
dev_info = validate_one_epoch_joint(
dev_loader,
model,
device,
optimizer,
loss_generate,
loss_predict,
loss_perceptual_entropy,
epoch,
args,
)
end_t_dev = time.time()
# Print Total info
dev_log = "Dev epoch: {:04d} ".format(epoch)
if args.optimizer == "noam":
dev_log += "lr: {:.6f}, \n\t".format(
optimizer._optimizer.param_groups[0]["lr"])
elif args.optimizer == "adam":
dev_log += "lr: {:.6f}, \n\t".format(
optimizer.param_groups[0]["lr"])
dev_log += "total_loss: {:.4f} \n\t".format(dev_info["loss"])
# Print Generator info
if args.vocoder_category == "wavernn":
dev_log += "generator_loss: {:.4f} ".format(
dev_info["generator_loss"])
else:
dev_log += "generator_loss: {:.4f}, spec_loss: {:.4f} ".format(
dev_info["generator_loss"], dev_info["spec_loss"])
if args.n_mels > 0:
dev_log += "mel_loss: {:.4f}, ".format(dev_info["mel_loss"])
if args.perceptual_loss > 0:
dev_log += "pe_loss: {:.4f}\n\t".format(dev_info["pe_loss"])
# Print Predictor info
dev_log += "predictor_loss: {:.4f}, singer_loss: {:.4f}, ".format(
dev_info["predictor_loss"],
dev_info["singer_loss"],
)
dev_log += "phone_loss: {:.4f}, semitone_loss: {:.4f} \n\t\t".format(
dev_info["phone_loss"],
dev_info["semitone_loss"],
)
dev_log += "singer_accuracy: {:.4f}%, ".format(
dev_info["singer_accuracy"] * 100, )
dev_log += "phone_accuracy: {:.4f}%, semitone_accuracy: {:.4f}% ".format(
dev_info["phone_accuracy"] * 100,
dev_info["semitone_accuracy"] * 100,
)
logging.info("{} time: {:.2f}s".format(dev_log,
end_t_dev - start_t_train))
sys.stdout.flush()
torch.backends.cudnn.enabled = True
"""Save model Stage"""
if not os.path.exists(args.model_save_dir):
os.makedirs(args.model_save_dir)
(generator_loss_counter,
generator_loss_epoch_to_save) = Auto_save_model(
args,
epoch,
model,
optimizer,
train_info,
dev_info,
None, # logger
generator_loss_counter,
generator_loss_epoch_to_save,
save_loss_select="generator_loss",
)
(spec_loss_counter, spec_loss_epoch_to_save) = Auto_save_model(
args,
epoch,
model,
optimizer,
train_info,
dev_info,
None, # logger
spec_loss_counter,
spec_loss_epoch_to_save,
save_loss_select="spec_loss",
)
(predictor_loss_counter,
predictor_loss_epoch_to_save) = Auto_save_model(
args,
epoch,
model,
optimizer,
train_info,
dev_info,
None, # logger
predictor_loss_counter,
predictor_loss_epoch_to_save,
save_loss_select="predictor_loss",
)