def __init__(self, args, src_field, tgt_field):
super(Seq2Seq, self).__init__()
logger.info('[MODEL] Preparing the Standard Seq2Seq model')
self.beam_length_penalize = args.beam_length_penalize
# Build Embedding
if args.share_embedding:
self.src_embed = rnn_helper.build_embedding(
args.tgt_vocab_size_with_offsets, args.embed_size)
self.dec_embed = self.src_embed
assert args.share_vocab
enc_embed = self.src_embed
dec_embed = self.src_embed
else:
self.src_embed = rnn_helper.build_embedding(
args.src_vocab_size, args.embed_size)
self.dec_embed = rnn_helper.build_embedding(
args.tgt_vocab_size_with_offsets, args.embed_size)
dec_embed = self.dec_embed
# Build Encoder
encoder = RNNEncoder(hparams=args, embed=enc_embed)
self.encoder = encoder
# Build Bridge
if args.bridge == 'general':
bridge = LinearBridge(hparams=args)
elif args.bridge == 'none':
bridge = None
else:
raise NotImplementedError()
self.bridge = bridge
# Build Decoder
decoder = RNNDecoder(hparams=args, embed=dec_embed)
self.decoder = decoder
# Build Copy Attention
self.copy_mode = False
self.copy_coverage = False
if args.copy:
self.copy_mode = True
# Coverage
if args.copy_coverage > 0.0:
self.copy_coverage = True
# Special Tokens
self.tgt_sos_idx = tgt_field.vocab.stoi['<sos>']
self.tgt_eos_idx = tgt_field.vocab.stoi['<eos>']
self.tgt_unk_idx = tgt_field.vocab.stoi['<unk>']
self.tgt_pad_idx = tgt_field.vocab.stoi['<pad>']
def __init__(self, args, src_embed, concat=True, dropout=0.5):
super(LSTMFieldTableEncoder, self).__init__(args, src_embed)
self.concat = concat
self.hidden_size = args.hidden_size
self.field_size = args.field_key_embed_size
self.input_size = self.field_word_embed_size
assert args.field_input_tags == "none"
if concat:
self.word_encoder = LSTM(self.input_size + self.field_size, self.hidden_size)
else:
self.word_encoder = LSTM(self.input_size, self.hidden_size)
assert args.infobox_memory_bank_format == 'fwk_fwv_fk'
self.attribute_encoder = LSTM(self.hidden_size, self.hidden_size)
self.positional_embedding = build_embedding(500, 5)
self.output_projection = nn.Linear(self.hidden_size * 2, self.hidden_size, bias=False)
def __init__(self, method, dropout, vocab_size, word_embed_size, embed_size, down_scale=1.00):
super(T2STokenEncoder, self).__init__()
self.sub_embed = rnn_helper.build_embedding(vocab_size, word_embed_size)
self.scaled_embed_size = int(embed_size * down_scale)
if self.scaled_embed_size != word_embed_size:
self.down_scale_projection = torch.nn.Linear(word_embed_size, self.scaled_embed_size)
else:
self.down_scale_projection = None
if method == 'mean':
self.encoder = TokenMeanEncoder(dropout=dropout)
self.output_projection = torch.nn.Linear(word_embed_size + self.scaled_embed_size,
embed_size)
elif method == 'cnn':
self.encoder = TokenCNNEncoder(embedding_size=self.scaled_embed_size,
filters=[(1, 75), (2, 100), (3, 25)], dropout=dropout)
self.output_projection = torch.nn.Linear(word_embed_size + 200,
embed_size)
def __init__(self, args, src_field, tgt_field):
super(InfoSeq2Seq, self).__init__()
logger.info('[MODEL] Preparing the InfoSeq2Seq model')
self.drop_out = nn.Dropout(args.dropout)
self.beam_length_penalize = args.beam_length_penalize
self.hidden_size = args.hidden_size
self.teach_force_rate = args.teach_force_rate
# 如果打开 Table2Text模式，那么就只做Table2Text的生成,Text2Text 同理
self.table2text_mode = False
self.text2text_mode = False
if args.task_mode == 'text2text':
self.text2text_mode = True
elif args.task_mode == 'table2text':
self.table2text_mode = True
self.enable_query_encoder = True
self.enable_field_encoder = True
self.enabled_char_encoders = set(args.char_encoders.split(','))
if self.table2text_mode:
self.enable_query_encoder = False
assert args.copy is False
assert args.bridge == 'none' or args.bridge == 'general'
if self.text2text_mode:
self.enable_field_encoder = False
assert args.field_copy is False
assert args.bridge == 'none' or args.bridge == 'general'
# Word embedding for representing dialogue words
self.embed_size = args.embed_size
if args.share_embedding:
self.src_embed = rnn_helper.build_embedding(
args.tgt_vocab_size_with_offsets, args.embed_size)
self.dec_embed = self.src_embed
enc_embed = self.src_embed
dec_embed = self.src_embed
else:
self.src_embed = rnn_helper.build_embedding(
args.src_vocab_size, args.embed_size)
self.dec_embed = rnn_helper.build_embedding(
args.tgt_vocab_size_with_offsets, args.embed_size)
dec_embed = self.dec_embed
if 'src' in self.enabled_char_encoders:
self.sub_src_embed = rnn_helper.build_embedding(
args.sub_src_vocab_size, args.embed_size)
# 是否使用Dialogue的Pos Tag
if args.add_pos_tag_embedding and self.enable_query_encoder:
self.add_pos_tag_embedding = True
self.src_pos_tag_embed_size = args.field_tag_embedding_size
self.src_pos_tag_embed = rnn_helper.build_embedding(
args.src_tag_vocab_size, self.src_pos_tag_embed_size)
else:
self.src_pos_tag_embed_size = 0
self.add_pos_tag_embedding = False
self.src_pos_tag_embed = None
# Build Field Encoder
if self.enable_field_encoder:
if args.field_encoder == 'lstm':
field_encoder = FieldTableEncoder(args=args,
src_embed=self.src_embed)
elif args.field_encoder == 'transformer':
field_encoder = TransformerFieldTableEncoder(
args=args, src_embed=self.src_embed)
elif args.field_encoder == 'hierarchical_lstm':
field_encoder = LSTMFieldTableEncoder(args=args,
src_embed=self.src_embed)
elif args.field_encoder == 'hierarchical_field':
field_encoder = HierarchicalFieldTableEncoder(
args=args, src_embed=self.src_embed)
elif args.field_encoder == 'hierarchical_intra_field':
field_encoder = HierarchicalIntraFieldTableEncoder(
args=args, src_embed=self.src_embed)
elif args.field_encoder == 'hierarchical_infobox':
field_encoder = HierarchicalInfoboxEncoder(
args=args, src_embed=self.src_embed)
self.field_equivalent_input_size = field_encoder.get_field_equivalent_input_size(
)
self.field_encoder = field_encoder
else:
self.field_equivalent_input_size = None
self.field_encoder = None
if self.enable_query_encoder:
encoder = RNNEncoder(hparams=args,
embed=enc_embed,
tag_embed=self.src_pos_tag_embed)
self.encoder = encoder
if args.enc_birnn:
self.birnn_down_scale = torch.nn.Linear(
args.hidden_size * 2, args.hidden_size, False)
else:
self.birnn_down_scale = None
# Build Bridge
self.bridge_mode = args.bridge
if args.bridge == 'general' or args.bridge == 'fusion':
bridge = LinearBridge(args.bridge, args.rnn_type,
args.hidden_size, args.enc_layers,
args.dropout, args.enc_birnn)
elif args.bridge == 'attention' or args.bridge == 'field_attention' or args.bridge == 'fusion_attention':
bridge = AttentionBridge(args.bridge, args.hidden_size,
args.dropout, args.enc_birnn)
elif args.bridge == 'clue_attention' or args.bridge == 'clue_attention2':
bridge = AttentionBridge('attention', args.hidden_size,
args.dropout, args.enc_birnn)
self.posterior_bridge = AttentionBridge(
'attention', args.hidden_size, args.dropout,
args.enc_birnn)
if args.bridge == 'clue_attention' or args.bridge == 'clue_attention2':
self.clue_query_projection = torch.nn.Linear(
(int(args.enc_birnn) + 1) * args.hidden_size * 2,
args.hidden_size * (int(args.enc_birnn) + 1))
elif args.bridge == 'post_field_attention' or args.bridge == 'clue_field_attention' \
or args.bridge == 'clue_field_attention2':
bridge = AttentionBridge('field_attention', args.hidden_size,
args.dropout, args.enc_birnn)
self.posterior_bridge = AttentionBridge(
'field_attention', args.hidden_size, args.dropout,
args.enc_birnn)
if args.bridge == 'clue_field_attention' or args.bridge == 'clue_field_attention2':
self.clue_query_projection = torch.nn.Linear(
(int(args.enc_birnn) + 1) * args.hidden_size * 2,
args.hidden_size * (int(args.enc_birnn) + 1))
elif args.bridge == 'post_attn_fusion':
bridge = AttentionBridge('fusion_attention', args.hidden_size,
args.dropout, args.enc_birnn)
self.posterior_bridge = LinearBridge('general', args.rnn_type,
args.hidden_size,
args.enc_layers,
args.dropout)
elif args.bridge == 'post_fusion' or args.bridge == 'post_fusion2':
bridge = LinearBridge('fusion', args.rnn_type,
args.hidden_size, args.enc_layers,
args.dropout, args.enc_birnn)
post_bridge = LinearBridge('general', args.rnn_type,
args.hidden_size, args.enc_layers,
args.dropout, args.enc_birnn)
self.posterior_bridge = post_bridge
elif args.bridge == 'none':
bridge = None
else:
raise NotImplementedError()
self.bridge = bridge
else:
self.encoder = None
assert args.bridge == 'none', 'table2text mode does not support bridge'
# Build Decoder
decoder = TableAwareDecoder(hparams=args, embed=dec_embed)
self.decoder = decoder
# Build SRC-Copy Attention
self.src_copy_mode = False
self.copy_coverage = False
if args.copy:
assert self.table2text_mode is False
self.src_copy_mode = True
self.max_copy_token_num = args.max_copy_token_num
# State-to-Input Projection:
self.add_state_to_copy_token = args.add_state_to_copy_token
copy_state_input_dim = args.embed_size
if self.add_pos_tag_embedding:
copy_state_input_dim += self.src_pos_tag_embed_size
if self.add_state_to_copy_token:
copy_state_input_dim += self.hidden_size
if copy_state_input_dim > args.embed_size:
self.copied_state_to_input_embed_projection = nn.Linear(
copy_state_input_dim, args.embed_size, bias=False)
else:
self.copied_state_to_input_embed_projection = None
# Coverage
if args.copy_coverage > 0.0:
self.copy_coverage = True
# Build Field-SRC Copy
self.field_copy_mode = False
if args.field_copy:
self.field_copy_mode = True
self.max_kw_pairs_num = args.max_kw_pairs_num
# State-to-Input Projection
self.field_state_to_input_embed_projection = nn.Linear(
self.field_equivalent_input_size + self.hidden_size,
args.embed_size,
bias=False)
# Special Tokens
self.tgt_sos_idx = tgt_field.vocab.stoi['<sos>']
self.tgt_eos_idx = tgt_field.vocab.stoi['<eos>']
self.tgt_unk_idx = tgt_field.vocab.stoi['<unk>']
self.tgt_pad_idx = tgt_field.vocab.stoi['<pad>']
def __init__(self, args, src_embed):
super(InfoboxTableEncoder, self).__init__()
self.infobox_mode = 'standard'
self.src_embed = src_embed
self.field_key_embed_size = args.field_key_embed_size
self.field_all_pos_embed_size = 0
self.field_input_tags = set(args.field_input_tags.split(','))
assert len(self.field_input_tags - VALID_FIELD_POS_INPUTS) == 0
self.enabled_char_encoders = set(args.char_encoders.split(','))
# Word embedding for representing infobox words
self.field_key_embed = rnn_helper.build_embedding(
args.field_vocab_size, args.field_key_embed_size)
if 'field_key' in self.enabled_char_encoders:
self.sub_field_key_char_encoder = char_encoders.T2STokenEncoder(
args.char_encoder_type,
dropout=args.dropout,
vocab_size=args.sub_field_vocab_size,
word_embed_size=args.field_key_embed_size,
embed_size=args.field_key_embed_size,
)
else:
self.sub_field_key_char_encoder = None
if 'field_word' in self.enabled_char_encoders:
self.sub_field_word_char_encoder = char_encoders.T2STokenEncoder(
args.char_encoder_type,
dropout=args.dropout,
vocab_size=args.sub_field_word_vocab_size,
word_embed_size=args.embed_size,
embed_size=args.embed_size)
else:
self.sub_field_word_embed = None
self.sub_field_word_char_encoder = None
# Local Positions
if 'local_pos_fw' in self.field_input_tags:
self.local_pos_fw_embed = rnn_helper.build_embedding(
args.max_field_intra_word_num,
args.field_position_embedding_size)
self.field_all_pos_embed_size += args.field_position_embedding_size
else:
self.local_pos_fw_embed = None
if 'local_pos_bw' in self.field_input_tags:
self.local_pos_bw_embed = rnn_helper.build_embedding(
args.max_field_intra_word_num,
args.field_position_embedding_size)
self.field_all_pos_embed_size += args.field_position_embedding_size
else:
self.local_pos_bw_embed = None
# Global Positions
if 'field_kv_pos' in self.field_input_tags:
self.field_kv_pos_embed = rnn_helper.build_embedding(
args.max_kv_pairs_num, args.field_position_embedding_size)
self.field_all_pos_embed_size += args.field_position_embedding_size
else:
self.field_kv_pos_embed = None
if 'field_kw_pos' in self.field_input_tags:
self.field_kw_pos_embed = rnn_helper.build_embedding(
args.max_kw_pairs_num, args.field_position_embedding_size)
self.field_all_pos_embed_size += args.field_position_embedding_size
else:
self.field_kw_pos_embed = None
# POS Tag 的Embedding
if args.field_tag_usage == 'general':
self.field_tag_embed = rnn_helper.build_embedding(
args.field_pos_tag_vocab_size, args.field_tag_embedding_size)
self.field_tag_embed_size = args.field_tag_embedding_size
else:
self.field_tag_embed = None
self.field_tag_embed_size = 0
# 对于Filed Word是否同时使用Field Word和标准Word的Embedding
self.dual_field_word_embedding = args.dual_field_word_embedding
if args.field_word_vocab_path != 'none':
logger.info('[Model] Use a separate field word embedding ')
self.field_word_embedding = rnn_helper.build_embedding(
args.field_word_vocab_size, args.embed_size)
self.field_word_embed_size = args.embed_size
if args.dual_field_word_embedding:
self.field_word_embed_size = args.embed_size * 2
else:
assert args.dual_field_word_embedding is False, 'requires a separate field vocab'
self.field_word_embedding = self.src_embed
self.field_word_embed_size = args.embed_size
self.dual_attn = args.dual_attn != 'none'
if self.dual_attn:
self.dual_attn_projection = nn.Linear(
self.field_key_embed_size + self.field_all_pos_embed_size +
self.field_tag_embed_size,
args.hidden_size,
bias=False)