def __init__(self,
n_src_vocab=len(symbols) + 1,
len_max_seq=hp.max_seq_len,
d_word_vec=hp.encoder_hidden,
n_layers=hp.encoder_layer,
n_head=hp.encoder_head,
d_k=hp.encoder_hidden // hp.encoder_head,
d_v=hp.encoder_hidden // hp.encoder_head,
d_model=hp.encoder_hidden,
d_inner=hp.fft_conv1d_filter_size,
dropout=hp.encoder_dropout):
super(Encoder, self).__init__()
n_position = len_max_seq + 1
self.src_word_emb = nn.Embedding(n_src_vocab,
d_word_vec,
padding_idx=Constants.PAD)
self.position_enc = nn.Embedding.from_pretrained(
get_sinusoid_encoding_table(n_position, d_word_vec, padding_idx=0),
freeze=True)
self.layer_stack = nn.ModuleList([
FFTBlock(d_model, d_inner, n_head, d_k, d_v, dropout=dropout)
for _ in range(n_layers)
])
def __init__(self,
py_vocab_size,
hz_vocab_size=None,
len_max_seq=hp.max_seq_len,
d_word_vec=hp.encoder_hidden,
n_layers=hp.encoder_layer,
n_head=hp.encoder_head,
d_k=hp.encoder_hidden // hp.encoder_head,
d_v=hp.encoder_hidden // hp.encoder_head,
d_model=hp.encoder_hidden,
d_inner=hp.fft_conv1d_filter_size,
dropout=hp.encoder_dropout):
super(Encoder, self).__init__()
n_position = len_max_seq + 1
self.src_word_emb = nn.Embedding(
py_vocab_size, d_word_vec, padding_idx=Constants.PAD)
#self.hz_vocab = hz_vocab
self.d_word_vec =d_word_vec
if hz_vocab_size is not None:
self.cn_word_emb = nn.Embedding(
hz_vocab_size, hp.hz_emb_size, padding_idx=Constants.PAD)
else:
self.hz_word_emb = None
self.position_enc = nn.Parameter(
get_sinusoid_encoding_table(n_position, d_word_vec).unsqueeze(0), requires_grad=False)
self.layer_stack = nn.ModuleList([FFTBlock(
d_model, d_inner, n_head, d_k, d_v, dropout=dropout) for _ in range(n_layers)])
def __init__(self,
len_max_seq=hp.max_sep_len,
d_word_vec=hp.word_vec_dim,
n_layers=hp.decoder_n_layer,
n_head=hp.decoder_head,
d_k=64,
d_v=64,
d_model=hp.word_vec_dim,
d_inner=hp.decoder_conv1d_filter_size,
dropout=hp.dropout):
super(Decoder, self).__init__()
n_position = len_max_seq + 1
# self.src_word_emb = nn.Embedding(
# n_src_vocab, d_word_vec, padding_idx=Constants.PAD)
# self.encoder_prenet = EncoderPreNet()
self.position_enc = nn.Embedding.from_pretrained(
get_sinusoid_encoding_table(n_position, d_word_vec, padding_idx=0),
freeze=True)
self.layer_stack = nn.ModuleList([
FFTBlock(d_model, d_inner, n_head, d_k, d_v, dropout=dropout)
for _ in range(n_layers)
])
def __init__(self,
n_src_vocab=len(symbols) + 1,
len_max_seq=hp.max_seq_len,
d_word_vec=hp.encoder_hidden,
n_layers=hp.encoder_layer,
n_head=hp.encoder_head,
d_k=hp.encoder_hidden // hp.encoder_head,
d_v=hp.encoder_hidden // hp.encoder_head,
d_model=hp.encoder_hidden,
d_inner=hp.fft_conv1d_filter_size,
dropout=hp.encoder_dropout):
super(Encoder, self).__init__()
n_position = len_max_seq + 1
# <144, 256>, 0点的pad的embedding全是0
self.src_word_emb = nn.Embedding(n_src_vocab,
d_word_vec,
padding_idx=Constants.PAD)
# <1, 1001, 256>
self.position_enc = nn.Parameter(get_sinusoid_encoding_table(
n_position, d_word_vec).unsqueeze(0),
requires_grad=False)
self.layer_stack = nn.ModuleList([
FFTBlock(d_model, d_inner, n_head, d_k, d_v, dropout=dropout)
for _ in range(n_layers)
])
def __init__(self,
len_max_seq=hp.max_seq_len,
d_word_vec=hp.encoder_hidden,
n_layers=hp.decoder_layer,
n_head=hp.decoder_head,
d_k=hp.decoder_hidden // hp.decoder_head,
d_v=hp.decoder_hidden // hp.decoder_head,
d_model=hp.decoder_hidden,
d_inner=hp.fft_conv1d_filter_size,
dropout=hp.decoder_dropout):
super(Decoder, self).__init__()
n_position = len_max_seq + 1
self.speaker_fc = nn.Linear(512, 256, bias=False)
self.position_enc = nn.Parameter(get_sinusoid_encoding_table(
n_position, d_word_vec).unsqueeze(0),
requires_grad=False)
self.layer_stack = nn.ModuleList([
FFTBlock(d_model, d_inner, n_head, d_k, d_v, dropout=dropout)
for _ in range(n_layers)
])
def __init__(self, speaker_encoder_dim: int = 0):
super(Decoder, self).__init__()
d_word_vec = hp.encoder_hidden + speaker_encoder_dim
d_k = (hp.decoder_hidden + speaker_encoder_dim) // hp.decoder_head
d_v = (hp.decoder_hidden + speaker_encoder_dim) // hp.decoder_head
d_model = (hp.decoder_hidden + speaker_encoder_dim)
n_position = hp.max_seq_len + 1
self.position_enc = nn.Parameter(get_sinusoid_encoding_table(
n_position, d_word_vec).unsqueeze(0),
requires_grad=False)
self.layer_stack = nn.ModuleList([
FFTBlock(d_model,
hp.fft_conv1d_filter_size,
hp.decoder_head,
d_k,
d_v,
dropout=hp.decoder_dropout)
for _ in range(hp.decoder_layer)
])
def __init__(self,
len_max_seq=hp.max_seq_len,
n_layers=hp.decoder_n_layer,
n_head=hp.decoder_head,
d_k=hp.decoder_dim // hp.decoder_head,
d_v=hp.decoder_dim // hp.decoder_head,
d_model=hp.decoder_dim,
d_inner=hp.decoder_conv1d_filter_size,
dropout=hp.dropout):
super(Decoder, self).__init__()
n_position = len_max_seq + 1
self.position_enc = nn.Embedding.from_pretrained(
get_sinusoid_encoding_table(n_position, d_model, padding_idx=0),
freeze=True)
self.layer_stack = nn.ModuleList([
FFTBlock(d_model, d_inner, n_head, d_k, d_v, dropout=dropout)
for _ in range(n_layers)
])
