def __init__(self, corpus, config):
super(ContEncoder, 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.config = config
self.embedding = nn.Embedding(self.vocab_size, config.embed_size,
padding_idx=self.rev_vocab[PAD])
self.utt_encoder = RnnUttEncoder(config.utt_cell_size, 0.0,
bidirection=False,
# bidirection=True in the original code
use_attn=config.utt_type == 'attn_rnn',
vocab_size=self.vocab_size,
embedding=self.embedding)
self.ctx_encoder = EncoderRNN(self.utt_encoder.output_size,
config.ctx_cell_size,
0.0,
0.0,
config.num_layer,
config.rnn_cell,
variable_lengths=self.config.fix_batch)
def __init__(self, corpus, config):
super(StED, 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]
if not hasattr(config, "freeze_step"):
config.freeze_step = 6000
# build model here
# word embeddings
self.x_embedding = nn.Embedding(self.vocab_size, config.embed_size)
# latent action learned
self.x_encoder = EncoderRNN(config.embed_size, config.dec_cell_size,
dropout_p=config.dropout,
rnn_cell=config.rnn_cell,
variable_lengths=False)
self.q_y = nn.Linear(config.dec_cell_size, config.y_size * config.k)
self.x_init_connector = nn_lib.LinearConnector(config.y_size * config.k,
config.dec_cell_size,
config.rnn_cell == 'lstm')
# decoder
self.prev_decoder = DecoderRNN(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,
use_attention=False,
use_gpu=config.use_gpu,
embedding=self.x_embedding)
self.next_decoder = DecoderRNN(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,
use_attention=False,
use_gpu=config.use_gpu,
embedding=self.x_embedding)
# Encoder-Decoder STARTS here
self.embedding = nn.Embedding(self.vocab_size, config.embed_size,
padding_idx=self.rev_vocab[PAD])
self.utt_encoder = RnnUttEncoder(config.utt_cell_size, config.dropout,
bidirection=False,
# bidirection=True in the original code
use_attn=config.utt_type == 'attn_rnn',
vocab_size=self.vocab_size,
embedding=self.embedding)
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=config.fix_batch)
# FNN to get Y
self.p_fc1 = nn.Linear(config.ctx_cell_size, config.ctx_cell_size)
self.p_y = nn.Linear(config.ctx_cell_size, config.y_size * config.k)
# connector
self.c_init_connector = nn_lib.LinearConnector(config.y_size * config.k,
config.dec_cell_size,
config.rnn_cell == 'lstm')
# decoder
self.decoder = DecoderRNN(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,
use_attention=config.use_attn,
attn_size=config.dec_cell_size,
attn_mode=config.attn_type,
use_gpu=config.use_gpu,
embedding=self.embedding)
# force G(z,c) has z
if config.use_attribute:
self.attribute_loss = criterions.NLLEntropy(-100, config)
self.cat_connector = nn_lib.GumbelConnector(config.use_gpu)
self.greedy_cat_connector = nn_lib.GreedyConnector()
self.nll_loss = criterions.NLLEntropy(self.rev_vocab[PAD], self.config)
self.cat_kl_loss = criterions.CatKLLoss()
self.log_uniform_y = Variable(torch.log(torch.ones(1) / config.k))
self.entropy_loss = criterions.Entropy()
if self.use_gpu:
self.log_uniform_y = self.log_uniform_y.cuda()
self.kl_w = 0.0