(max_seq_len,B,src_voc_size) embed(src_voc_size,embedding_size) (max_seq_len,B,embedding_size) LSTM(embedding_size,hidden_size,nlayer,bidirectional=True) enc_out:(max_seq_len,B,hidden_size), enc_last_hidden:((2nlayer,B,hidden_size),(2nlayer,B,hidden_size))
Attention(enc_last_hidden,enc_out) attn_weights:(1,B,hidden_size) context = attn_weights*enc_out context:(1,B,hidden_size)
dec_input:(1,B,tgt_voc_size) embed(tgt_voc_size,embedding_size) word_emb:(1,B,embedding_size) (lstm)rnn_input:concat(word_emb,context) (lstm)rnn_input:(1,B,(hidden_size+embedding_size)) LSTM((hidden_size+embedding_size),hidden_size,nlayer,bidirectional=False) dec_out:(1,B,hidden_size), dec_last_hidden:((nlayer,B,hidden_size),(nlayer,2,hidden_size)) concat(dec_out, context) (1,B,hidden_size+hidden_size) linear(hidden_size+hidden_size,tgt_voc_size) (1,B,tgt_voc_size)
(1,B,tgt_voc_size)