def test_get_model_symbols_count_uses_const(self):
ref = get_model_symbols_count(n_symbols=10,
n_accents=5,
accents_use_own_symbols=True,
shared_symbol_count=SHARED_SYMBOLS_COUNT)
res = get_model_symbols_count(n_symbols=10,
n_accents=5,
accents_use_own_symbols=True)
self.assertEqual(ref, res)
def __init__(self, hparams: HParams, logger: Logger):
super(Tacotron2, self).__init__()
model_symbols_count = get_model_symbols_count(
hparams.n_symbols, hparams.n_accents,
hparams.accents_use_own_symbols)
if hparams.accents_use_own_symbols:
logger.info(
f"All {hparams.n_accents} accent(s) use an own symbolset. The number of total symbols increased from {hparams.n_symbols} to {model_symbols_count}."
)
else:
logger.info(
f"All {hparams.n_accents} accent(s) share the same symbolset. The number of total symbols is {model_symbols_count}."
)
self.logger = logger
self.mask_padding = hparams.mask_padding
self.n_mel_channels = hparams.n_mel_channels
# TODO rename to symbol_embeddings but it will destroy all previous trained models
symbol_emb_weights = get_symbol_weights(hparams)
self.embedding = weights_to_embedding(symbol_emb_weights)
logger.debug(f"is cuda: {self.embedding.weight.is_cuda}")
speaker_emb_weights = get_speaker_weights(hparams)
self.speakers_embedding = weights_to_embedding(speaker_emb_weights)
#self.accent_embedding = nn.Embedding(hparams.n_accents, hparams.accents_embedding_dim)
# torch.nn.init.xavier_uniform_(self.accent_embedding.weight)
self.encoder = Encoder(hparams)
self.decoder = Decoder(hparams, logger)
self.postnet = Postnet(hparams)
def get_symbol_weights(hparams: HParams) -> torch.Tensor:
model_symbols_count = get_model_symbols_count(
hparams.n_symbols, hparams.n_accents, hparams.accents_use_own_symbols)
model_weights = get_uniform_weights(model_symbols_count,
hparams.symbols_embedding_dim)
return model_weights
def test_get_model_symbols_count_ns10_na5_returns_10(self):
res = get_model_symbols_count(n_symbols=10,
n_accents=5,
accents_use_own_symbols=False,
shared_symbol_count=1)
self.assertEqual(10, res)
def test_get_model_symbols_count_acc_shared2_ns2_na50_returns_1(self):
res = get_model_symbols_count(n_symbols=2,
n_accents=50,
accents_use_own_symbols=True,
shared_symbol_count=2)
self.assertEqual(2, res)
def test_get_model_symbols_count_acc_shared2_ns10_na5_returns_42(self):
res = get_model_symbols_count(n_symbols=10,
n_accents=5,
accents_use_own_symbols=True,
shared_symbol_count=2)
# 2 + 5 accents * 8 symbols
self.assertEqual(42, res)
def test_get_model_symbols_count_acc_shared1_ns10_na5_returns_46(self):
res = get_model_symbols_count(n_symbols=10,
n_accents=5,
accents_use_own_symbols=True,
shared_symbol_count=1)
# 1 + 5 accents * 9 symbols
self.assertEqual(46, res)