def __init__(self,
src_vocab_size,
trg_vocab_size,
embed_size,
hidden_size,
num_layers,
dpt=0.2):
super(NSeq2Seq, self).__init__()
self.encoder = Encoder(src_vocab_size, embed_size, hidden_size,
num_layers, dpt)
# self.decoder = BasicDecoder(trg_vocab_size, embed_size, hidden_size, num_layers, dpt)
# self.decoder = BasicAttentionDecoder(trg_vocab_size, embed_size, 2 * hidden_size, num_layers, dpt)
# self.decoder = BasicBahdanauAttnDecoder(trg_vocab_size, embed_size, hidden_size, num_layers, dpt)
# Store hyperparameters
self.h_dim = hidden_size
# self.vocab_size_trg, self.emb_dim_trg = embedding_trg.size()
# self.bos_token = tokens_bos_eos_pad_unk[0]
# self.eos_token = tokens_bos_eos_pad_unk[1]
# self.pad_token = tokens_bos_eos_pad_unk[2]
# self.unk_token = tokens_bos_eos_pad_unk[3]
# self.reverse_input = reverse_input
# Create encoder, decoder, attention
# self.encoder = EncoderLSTM(embedding_src, h_dim, num_layers, dropout_p=dropout_p, bidirectional=bi)
embedding = nn.Embedding(trg_vocab_size, embed_size)
self.decoder = Decoder.DecoderLSTM(embedding,
hidden_size,
num_layers * 2,
dropout_p=dpt)
self.attention = Attention.Attention(bidirectional=True,
h_dim=self.h_dim)
# Create linear layers to combine context and hidden state
self.linear1 = nn.Linear(2 * self.h_dim, self.emb_dim_trg)
self.tanh = nn.Tanh()
self.dropout = nn.Dropout(dpt)
self.linear2 = nn.Linear(self.emb_dim_trg, self.vocab_size_trg)
# Tie weights of decoder embedding and output
# if True and self.decoder.embedding.weight.size() == self.linear2.weight.size():
# print('Weight tying!')
self.linear2.weight = self.decoder.embedding.weight
