def __init__(self, config, input_emb_size, mix_speech_len, tgt_vocab_size, use_cuda, pretrain=None, score_fn=''):
super(seq2seq, self).__init__()
if pretrain is not None:
src_embedding = pretrain['src_emb']
tgt_embedding = pretrain['tgt_emb']
else:
src_embedding = None
tgt_embedding = None
self.encoder = models.rnn_encoder(config, input_emb_size, None, embedding=src_embedding)
if config.shared_vocab == False:
self.decoder = models.rnn_decoder(config, tgt_vocab_size, embedding=tgt_embedding, score_fn=score_fn)
else:
self.decoder = models.rnn_decoder(config, tgt_vocab_size, embedding=self.encoder.embedding,
score_fn=score_fn)
self.use_cuda = use_cuda
self.tgt_vocab_size = tgt_vocab_size
self.config = config
self.criterion = models.criterion(tgt_vocab_size, use_cuda,config.loss)
self.loss_for_ss = nn.MSELoss()
self.log_softmax = nn.LogSoftmax()
self.wav_loss = models.WaveLoss(dBscale=1, nfft=config.FRAME_LENGTH, hop_size=config.FRAME_SHIFT)
speech_fre = input_emb_size
num_labels = tgt_vocab_size
if config.use_tas:
self.ss_model = models.ConvTasNet()
else:
self.ss_model = models.SS(config, speech_fre, mix_speech_len, num_labels)
def __init__(self,
config,
input_emb_size,
mix_speech_len,
tgt_vocab_size,
use_cuda,
pretrain=None,
score_fn=None):
super(seq2seq, self).__init__()
if pretrain is not None:
src_embedding = pretrain['src_emb']
tgt_embedding = pretrain['tgt_emb']
else:
src_embedding = None
tgt_embedding = None
self.encoder = models.rnn_encoder(config,
input_emb_size,
None,
embedding=src_embedding)
if config.shared_vocab == False:
self.decoder = models.rnn_decoder(config,
tgt_vocab_size,
embedding=tgt_embedding,
score_fn=score_fn)
else:
self.decoder = models.rnn_decoder(config,
tgt_vocab_size,
embedding=self.encoder.embedding,
score_fn=score_fn)
self.use_cuda = use_cuda
self.tgt_vocab_size = tgt_vocab_size
self.config = config
self.criterion = models.criterion(tgt_vocab_size, use_cuda)
self.loss_for_ss = nn.MSELoss()
self.log_softmax = nn.LogSoftmax()
speech_fre = input_emb_size
num_labels = tgt_vocab_size
self.ss_model = models.SS(config, speech_fre, mix_speech_len,
num_labels)
def __init__(self, config, input_emb_size, mix_speech_len, tgt_vocab_size, use_cuda, pretrain=None, score_fn=''):
super(seq2seq_music, self).__init__()
if pretrain is not None:
src_embedding = pretrain['src_emb']
tgt_embedding = pretrain['tgt_emb']
else:
src_embedding = None
tgt_embedding = None
self.use_cuda = use_cuda
self.tgt_vocab_size = tgt_vocab_size
self.config = config
self.criterion = models.criterion(tgt_vocab_size, use_cuda,config.loss)
self.loss_for_ss = nn.MSELoss()
self.log_softmax = nn.LogSoftmax()
self.wav_loss = models.WaveLoss(dBscale=1, nfft=config.FRAME_LENGTH, hop_size=config.FRAME_SHIFT)
speech_fre = input_emb_size
num_labels = tgt_vocab_size
if config.use_tas:
if self.config.use_dprnn:
self.ss_model = models.FaSNet_base(config)
self.spk_lstm = nn.LSTMCell(self.ss_model.B + self.ss_model.N, self.ss_model.B) # LSTM over the speakers' step.
else:
self.ss_model = models.ConvTasNet_music(config)
if self.config.two_stage:
self.second_ss_model = models.ConvTasNet_2nd(config)
for p in self.encoder.parameters():
p.requires_grad = False
for p in self.decoder.parameters():
p.requires_grad = False
for p in self.ss_model.parameters():
p.requires_grad = False
self.spk_lstm = nn.LSTMCell(self.ss_model.B + self.ss_model.N, self.ss_model.B) # LSTM over the speakers' step.
else:
# self.ss_model = models.SS_att(config, speech_fre, mix_speech_len, num_labels)
self.ss_model = models.SS(config, speech_fre, mix_speech_len, num_labels)
def __init__(self,
config,
input_emb_size,
mix_speech_len,
tgt_vocab_size,
use_cuda,
pretrain=None,
score_fn=''):
super(seq2seq, self).__init__()
if pretrain is not None:
src_embedding = pretrain['src_emb']
tgt_embedding = pretrain['tgt_emb']
else:
src_embedding = None
tgt_embedding = None
# self.encoder = models.rnn_encoder(config, input_emb_size, None, embedding=src_embedding)
self.encoder = models.TransEncoder(config, input_emb_size)
self.decoder = models.TransDecoder(config,
sos_id=0,
eos_id=tgt_vocab_size - 1,
n_tgt_vocab=tgt_vocab_size)
# if config.shared_vocab == False:
# self.decoder = models.rnn_decoder(config, tgt_vocab_size, embedding=tgt_embedding, score_fn=score_fn)
# else:
# self.decoder = models.rnn_decoder(config, tgt_vocab_size, embedding=self.encoder.embedding,
# score_fn=score_fn)
self.use_cuda = use_cuda
self.tgt_vocab_size = tgt_vocab_size
self.config = config
self.criterion = models.criterion(tgt_vocab_size, use_cuda,
config.loss)
self.loss_for_ss = nn.MSELoss()
self.log_softmax = nn.LogSoftmax()
self.wav_loss = models.WaveLoss(dBscale=1,
nfft=config.FRAME_LENGTH,
hop_size=config.FRAME_SHIFT)
speech_fre = input_emb_size
num_labels = tgt_vocab_size
if config.use_tas:
if self.config.use_dprnn:
self.ss_model = models.FaSNet_base(config)
if self.config.two_stage:
self.second_ss_model = models.FaSNet_base_2nd(config)
for p in self.encoder.parameters():
p.requires_grad = False
for p in self.decoder.parameters():
p.requires_grad = False
for p in self.ss_model.parameters():
p.requires_grad = False
else:
self.ss_model = models.ConvTasNet(config)
if self.config.two_stage:
self.second_ss_model = models.ConvTasNet_2nd(config)
for p in self.encoder.parameters():
p.requires_grad = False
for p in self.decoder.parameters():
p.requires_grad = False
for p in self.ss_model.parameters():
p.requires_grad = False
else:
# self.ss_model = models.SS_att(config, speech_fre, mix_speech_len, num_labels)
self.ss_model = models.SS(config, speech_fre, mix_speech_len,
num_labels)