def __init__(self, use_postnet=True, n_spkers=1):
super(FastSpeech2, self).__init__()
### Speaker Embedding Table ###
self.use_spk_embed = hp.use_spk_embed
if self.use_spk_embed:
self.n_spkers = n_spkers
self.spk_embed_dim = hp.spk_embed_dim
self.spk_embed_weight_std = hp.spk_embed_weight_std
self.embed_speakers = Embedding(n_spkers, self.spk_embed_dim, padding_idx=None, std=self.spk_embed_weight_std)
self.use_emo_embed = hp.use_emo_embed
if self.use_emo_embed:
self.n_emotes = n_emotes
self.emo_embed_dim = hp.emo_embed_dim
self.emo_embed_weight_std = hp.emo_embed_weight_std
self.embed_emotions = Embedding(n_emotes, self.emo_embed_dim, padding_idx=None, std=self.emo_embed_weight_std)
### Encoder, Speaker Integrator, Variance Adaptor, Deocder, Postnet ###
self.encoder = Encoder()
if self.use_spk_embed:
self.speaker_integrator = SpeakerIntegrator()
self.variance_adaptor = VarianceAdaptor()
self.decoder = Decoder()
self.mel_linear = nn.Linear(hp.decoder_hidden, hp.n_mel_channels)
self.use_postnet = use_postnet
if self.use_postnet:
self.postnet = PostNet()
def __init__(self, opt):
self.opt = opt
self.tt = torch.cuda if opt.cuda else torch
checkpoint = torch.load(opt.model)
model_opt = checkpoint['settings']
self.model_opt = model_opt
model = Decoder(model_opt.user_size,
d_k=model_opt.d_k,
d_v=model_opt.d_v,
d_model=model_opt.d_model,
d_word_vec=model_opt.d_word_vec,
kernel_size=model_opt.window_size,
finit=model_opt.finit,
d_inner_hid=model_opt.d_inner_hid,
n_head=model_opt.n_head,
dropout=model_opt.dropout)
prob_projection = nn.Softmax()
model.load_state_dict(checkpoint['model'])
print('[Info] Trained model state loaded.')
if opt.cuda:
model.cuda()
prob_projection.cuda()
else:
model.cpu()
prob_projection.cpu()
model.prob_projection = prob_projection
self.model = model
self.model.eval()
def __init__(self):
super(FastSpeech2, self).__init__()
self.encoder = Encoder()
self.variance_adaptor = VarianceAdaptor()
self.decoder = Decoder()
self.mel_linear = Linear(hp.decoder_hidden, hp.n_mel_channels)
self.postnet = PostNet()
def __init__(self):
super(FastSpeech, self).__init__()
self.encoder = Encoder()
self.length_regulator = LengthRegulator()
self.decoder = Decoder()
self.mel_linear = Linear(hp.decoder_output_size, hp.num_mels)
self.postnet = PostNet()
def __init__(self):
super(FastSpeech, self).__init__()
self.encoder = Encoder()
self.length_regulator = LengthRegulator()
self.decoder = Decoder()
self.mel_linear = Linear(hp.decoder_dim, hp.num_mels)
self.postnet = CBHG(hp.num_mels, K=8, projections=[256, hp.num_mels])
self.last_linear = Linear(hp.num_mels * 2, hp.num_mels)
def __init__(self, use_postnet=True):
super(STYLER, self).__init__()
self.style_modeling = StyleModeling()
self.decoder = Decoder()
self.mel_linear = nn.Linear(hp.decoder_hidden, hp.n_mel_channels)
self.use_postnet = use_postnet
if self.use_postnet:
self.postnet = PostNet()
encoder_output = None
def __init__(self, use_postnet=True):
super(FastSpeech2, self).__init__()
self.encoder = Encoder()
self.variance_adaptor = TacotronDuration()
self.decoder = Decoder()
self.mel_linear = nn.Linear(hp.decoder_hidden, hp.n_mel_channels)
self.use_postnet = use_postnet
if self.use_postnet:
self.postnet = PostNet()
def __init__(self, py_vocab_size,hz_vocab_size=None, use_postnet=True):
super(FastSpeech2, self).__init__()
self.encoder = Encoder(py_vocab_size, hz_vocab_size = hz_vocab_size)
self.variance_adaptor = VarianceAdaptor()
self.decoder = Decoder()
self.mel_linear = nn.Linear(hp.decoder_hidden, hp.n_mel_channels)
self.use_postnet = use_postnet
if self.use_postnet:
self.postnet = UNet(scale=8)
def __init__(self, n_layers=hp.variance_predictor_layer):
super(VariancePredictor, self).__init__()
self.decoder = Decoder(
len_max_seq=hp.max_seq_len,
d_word_vec=hp.variance_predictor_hidden,
n_layers=n_layers,
n_head=hp.variance_predictor_head,
d_k=hp.variance_predictor_hidden // hp.variance_predictor_head,
d_v=hp.variance_predictor_hidden // hp.variance_predictor_head,
d_model=hp.variance_predictor_hidden,
d_inner=hp.fft_conv1d_filter_size,
dropout=hp.variance_predictor_dropout)
self.linear_layer = nn.Linear(hp.variance_predictor_hidden, 1)
def __init__(self):
super(LengthPredictor, self).__init__()
self.decoder = Decoder(
len_max_seq=hp.max_seq_len,
d_word_vec=hp.length_predictor_hidden,
n_layers=hp.length_predictor_layer,
n_head=hp.length_predictor_head,
d_k=hp.length_predictor_hidden // hp.length_predictor_head,
d_v=hp.length_predictor_hidden // hp.length_predictor_head,
d_model=hp.length_predictor_hidden,
d_inner=hp.fft_conv1d_filter_size,
dropout=hp.length_predictor_dropout)
self.linear_layer = nn.Linear(hp.length_predictor_hidden, 1)
self.tanh = nn.Tanh()
def __init__(self, use_postnet=True):
super(FastSpeech2, self).__init__()
# self.gst = GST()
self.encoder = Encoder()
self.variance_adaptor = VarianceAdaptor()
self.decoder = Decoder()
if hp.vocoder=='WORLD':
# self.f0_decoder= Decoder()
self.ap_linear = nn.Linear(hp.decoder_hidden, hp.n_ap_channels)
self.sp_linear = nn.Linear(hp.decoder_hidden, hp.n_sp_channels)
else:
self.mel_linear = nn.Linear(hp.decoder_hidden, hp.n_mel_channels)
self.use_postnet = use_postnet
if self.use_postnet:
self.postnet = PostNet()
def __init__(self,
src_vocab_size,
tgt_vocab_size,
n_layer=6,
d_model=512,
d_ff=2048,
n_head=8,
dropout=0.1):
super(Transformer, self).__init__()
self.src_embed = nn.Sequential(Embeddings(d_model, src_vocab_size),
PositionalEncoding(d_model, dropout))
self.tgt_embed = nn.Sequential(Embeddings(d_model, tgt_vocab_size),
PositionalEncoding(d_model, dropout))
self.encoder = Encoder(n_head, d_model, d_ff, dropout, n_layer)
self.decoder = Decoder(n_head, d_model, d_ff, dropout, n_layer)
self.generator = Generator(d_model, tgt_vocab_size)
# Initialize parameters with Glorot / fan_avg.
for p in self.parameters():
if p.dim() > 1:
nn.init.xavier_uniform_(p)
def __init__(self,
bert_model_path,
n_tgt_vocab,
len_max_seq,
d_word_vec=768,
d_model=768,
d_inner=3072,
n_layers=12,
n_head=12,
d_k=64,
d_v=64,
dropout=0.1):
super().__init__()
self.bert = BertModel.from_pretrained(bert_model_path)
self.config = BertConfig(bert_model_path + 'bert_config.json')
self.decoder = Decoder(n_tgt_vocab=n_tgt_vocab,
len_max_seq=len_max_seq,
d_word_vec=d_word_vec,
d_model=d_model,
d_inner=d_inner,
n_layers=n_layers,
n_head=n_head,
d_k=d_k,
d_v=d_v,
dropout=dropout)
self.tgt_word_prj = nn.Linear(d_model, n_tgt_vocab, bias=False)
nn.init.xavier_normal_(self.tgt_word_prj.weight)
self.tgt_word_prj.weight = self.decoder.tgt_word_emb.weight
self.x_logit_scale = (d_model**-0.5)
self.o_l = nn.Linear(d_model, 512, bias=False)
self.h_l = nn.Linear(512, 1, bias=True)
nn.init.xavier_normal_(self.o_l.weight)
nn.init.xavier_normal_(self.h_l.weight)
self.a_l_1 = nn.Linear(d_model, 512, bias=False)
self.a_l_2 = nn.Linear(d_model, 512, bias=False)
nn.init.xavier_normal_(self.a_l_1.weight)
nn.init.xavier_normal_(self.a_l_2.weight)
def __init__(self,n_layers):
super(VariancePredictor, self).__init__()
self.decoder = Decoder(n_layers)
self.linear_layer = nn.Linear(hp.decoder_hidden, 1)
def __init__(self,n_layers):
super(LengthPredictor, self).__init__()
self.decoder = Decoder(n_layers)
self.linear_layer = nn.Linear(hp.decoder_hidden, 1)
self.tanh=nn.Tanh()
def main():
''' Main function'''
parser = argparse.ArgumentParser()
parser.add_argument('-epoch', type=int, default=20)
parser.add_argument('-batch_size', type=int, default=32)
parser.add_argument('-d_model', type=int, default=64)
parser.add_argument('-d_inner_hid', type=int, default=64)
parser.add_argument('-d_k', type=int, default=64)
parser.add_argument('-d_v', type=int, default=64)
parser.add_argument('-window_size', type=int, default=3)
parser.add_argument('-finit', type=int, default=0)
parser.add_argument('-n_head', type=int, default=8)
parser.add_argument('-n_warmup_steps', type=int, default=1000)
parser.add_argument('-dropout', type=float, default=0.1)
parser.add_argument('-embs_share_weight', action='store_true')
parser.add_argument('-proj_share_weight', action='store_true')
parser.add_argument('-save_model', default='Lastfm_test')
parser.add_argument('-no_cuda', action='store_true')
opt = parser.parse_args()
opt.cuda = not opt.no_cuda
opt.d_word_vec = opt.d_model
#========= Preparing DataLoader =========#
train_data = DataLoader(use_valid=False,
load_dict=True,
batch_size=opt.batch_size,
cuda=opt.cuda)
opt.user_size = train_data.user_size
#========= Preparing Model =========#
decoder = Decoder(opt.user_size,
d_k=opt.d_k,
d_v=opt.d_v,
d_model=opt.d_model,
d_word_vec=opt.d_word_vec,
d_inner_hid=opt.d_inner_hid,
n_head=opt.n_head,
kernel_size=opt.window_size,
dropout=opt.dropout)
optimizer = ScheduledOptim(
optim.Adam(decoder.parameters(), betas=(0.9, 0.98), eps=1e-09),
opt.d_model, opt.n_warmup_steps)
def get_criterion(user_size):
''' With PAD token zero weight '''
weight = torch.ones(user_size)
weight[Constants.PAD] = 0
weight[Constants.EOS] = 0
return nn.CrossEntropyLoss(weight, size_average=False)
crit = get_criterion(train_data.user_size)
if opt.cuda:
decoder = decoder.cuda()
crit = crit.cuda()
train(decoder, train_data, crit, optimizer, opt)
else:
length_regulator_output, decoder_pos = self.length_regulator(
encoder_output, encoder_mask, alpha=alpha)
decoder_output = self.decoder(length_regulator_output, decoder_pos)
mel_output = self.mel_linear(decoder_output)
mel_output_postnet = self.postnet(mel_output) + mel_output
return mel_output, mel_output_postnet
if __name__ == "__main__":
# Test
test_encoder = Encoder()
test_decoder = Decoder()
# print(test_encoder)
# print(test_decoder)
test_src = torch.stack([
torch.Tensor([1, 2, 4, 3, 2, 5, 0, 0]),
torch.Tensor([3, 4, 2, 6, 7, 1, 2, 3])
]).long()
test_pos = torch.stack([
torch.Tensor([1, 2, 3, 4, 5, 6, 0, 0]),
torch.Tensor([1, 2, 3, 4, 5, 6, 7, 8])
]).long()
test_target = torch.stack([
torch.Tensor([0, 2, 3, 0, 3, 2, 1, 0]),
torch.Tensor([1, 2, 3, 2, 2, 0, 3, 6])
])