def __init__(self, corpus, config):
super(ZeroShotLAPtrHRED, self).__init__(config)
self.vocab = corpus.vocab
self.rev_vocab = corpus.rev_vocab
self.vocab_size = len(self.vocab)
self.go_id = self.rev_vocab[BOS]
self.eos_id = self.rev_vocab[EOS]
self.pad_id = self.rev_vocab[PAD]
# build model here
self.embedding = nn.Embedding(self.vocab_size,
config.embed_size,
padding_idx=self.pad_id)
self.utt_encoder = RnnUttEncoder(
config.utt_cell_size,
config.dropout,
use_attn=config.utt_type == 'rnn_attn',
vocab_size=self.vocab_size,
embedding=self.embedding,
feat_size=1)
self.ctx_encoder = EncoderRNN(self.utt_encoder.output_size,
config.ctx_cell_size,
0.0,
config.dropout,
config.num_layer,
config.rnn_cell,
variable_lengths=False,
bidirection=config.bi_ctx_cell)
self.policy = nn.Linear(self.ctx_encoder.output_size,
config.dec_cell_size)
self.utt_policy = lambda x: x
combined_dec_init_state_size = config.dec_cell_size + corpus.laed_z_size
self.connector = nn_lib.LinearConnector(
combined_dec_init_state_size,
config.dec_cell_size,
is_lstm=config.rnn_cell == 'lstm')
self.attn_size = self.ctx_encoder.output_size
self.decoder = DecoderPointerGen(self.vocab_size,
config.max_dec_len,
config.embed_size,
config.dec_cell_size,
self.go_id,
self.eos_id,
n_layers=1,
rnn_cell=config.rnn_cell,
input_dropout_p=config.dropout,
dropout_p=config.dropout,
attn_size=self.attn_size,
attn_mode=config.attn_type,
use_gpu=config.use_gpu,
embedding=self.embedding)
self.nll_loss = criterions.NLLEntropy(self.pad_id, config)
self.l2_loss = criterions.L2Loss()
def __init__(self, corpus, config):
super(HRED, self).__init__(config)
self.vocab = corpus.vocab
self.rev_vocab = corpus.rev_vocab
self.vocab_size = len(self.vocab)
self.go_id = self.rev_vocab[BOS]
self.eos_id = self.rev_vocab[EOS]
self.pad_id = self.rev_vocab[PAD]
# build model here
self.utt_encoder = ElmoUttEncoder(
config,
config.utt_cell_size,
config.dropout,
use_attn=config.utt_type == 'attn_rnn',
feat_size=1)
self.ctx_encoder = EncoderRNN(self.utt_encoder.output_size,
config.ctx_cell_size,
0.0,
config.dropout,
config.num_layer,
config.rnn_cell,
variable_lengths=False,
bidirection=config.bi_ctx_cell)
if config.bi_ctx_cell or config.num_layer > 1:
self.connector = Bi2UniConnector(config.rnn_cell, config.num_layer,
config.ctx_cell_size,
config.dec_cell_size)
else:
self.connector = IdentityConnector()
self.decoder = ElmoDecoderRNN(self.vocab_size,
config.max_dec_len,
self.utt_encoder.embed_size,
config.dec_cell_size,
self.go_id,
self.eos_id,
self.vocab,
n_layers=1,
rnn_cell=config.rnn_cell,
input_dropout_p=config.dropout,
dropout_p=config.dropout,
use_attention=config.use_attn,
embedding=self.utt_encoder.embedding,
attn_size=self.ctx_encoder.output_size,
attn_mode=config.attn_type,
use_gpu=config.use_gpu)
self.nll = criterions.NLLEntropy(self.pad_id, config)
def __init__(self, corpus, config):
super(LAPtrHRED, self).__init__(config, corpus)
self.ctx_encoder = EncoderRNN(self.utt_encoder.output_size,
config.ctx_cell_size,
0.0,
config.dropout,
config.num_layer,
config.rnn_cell,
variable_lengths=False,
bidirection=config.bi_ctx_cell)
combined_dec_init_state_size = config.dec_cell_size + corpus.laed_z_size
if config.bi_ctx_cell or config.num_layer > 1:
self.connector = Bi2UniConnector(config.rnn_cell,
config.num_layer,
combined_dec_init_state_size,
config.dec_cell_size)
else:
self.connector = nn.Linear(combined_dec_init_state_size, config.dec_cell_size)
self.attn_size = self.ctx_encoder.output_size
self.plain_embedding = nn.Embedding(self.vocab_size, config.embed_size)
self.decoder = DecoderPointerGen(self.vocab_size,
config.max_dec_len,
config.embed_size,
config.dec_cell_size,
self.go_id,
self.eos_id,
n_layers=1,
rnn_cell=config.rnn_cell,
input_dropout_p=config.dropout,
dropout_p=config.dropout,
attn_size=self.attn_size,
attn_mode=config.attn_type,
use_gpu=config.use_gpu,
embedding=self.plain_embedding)
self.nll_loss = criterions.NLLEntropy(self.pad_id, config)
def __init__(self, corpus, config):
super(ZeroShotPtrHRED, self).__init__(config, corpus)
self.ctx_encoder = EncoderRNN(self.utt_encoder.output_size,
config.ctx_cell_size,
0.0,
config.dropout,
config.num_layer,
config.rnn_cell,
variable_lengths=False,
bidirection=config.bi_ctx_cell)
self.policy = nn.Linear(self.ctx_encoder.output_size,
config.dec_cell_size)
self.utt_policy = lambda x: x
self.connector = nn_lib.LinearConnector(
config.dec_cell_size,
config.dec_cell_size,
is_lstm=config.rnn_cell == 'lstm')
self.attn_size = self.ctx_encoder.output_size
self.decoder = DecoderPointerGen(self.vocab_size,
config.max_dec_len,
self.embedding.embedding_dim,
config.dec_cell_size,
self.go_id,
self.eos_id,
n_layers=1,
rnn_cell=config.rnn_cell,
input_dropout_p=config.dropout,
dropout_p=config.dropout,
attn_size=self.attn_size,
attn_mode=config.attn_type,
use_gpu=config.use_gpu,
embedding=self.embedding)
self.nll_loss = criterions.NLLEntropy(self.pad_id, config)
self.l2_loss = criterions.L2Loss()
def __init__(self,
config,
utt_cell_size,
dropout,
rnn_cell='gru',
bidirection=True,
use_attn=False,
embedding=None,
vocab_size=None,
embed_dim=None,
feat_size=0):
super(ElmoUttEncoder, self).__init__()
self.bidirection = bidirection
self.utt_cell_size = utt_cell_size
if embedding is None:
self.embed_size = embed_dim
self.embedding = ElmoEmbedding(config)
self.embed_size = self.embedding.embedding_dim
else:
self.embedding = embedding
self.embed_size = embedding.embedding_dim
self.rnn = EncoderRNN(self.embed_size + feat_size,
utt_cell_size,
0.0,
dropout,
rnn_cell=rnn_cell,
variable_lengths=False,
bidirection=bidirection)
self.multipler = 2 if bidirection else 1
self.output_size = self.utt_cell_size * self.multipler
self.use_attn = use_attn
self.feat_size = feat_size
if use_attn:
self.key_w = nn.Linear(self.utt_cell_size * self.multipler,
self.utt_cell_size)
self.query = nn.Linear(self.utt_cell_size, 1)