def __init__(self,
vocab: Vocabulary,
embedder: TextFieldEmbedder,
question_encoder: Seq2SeqEncoder,
passage_encoder: Seq2SeqEncoder,
r: float = 0.8,
dropout: float = 0.1,
initializer: InitializerApplicator = InitializerApplicator(),
regularizer: Optional[RegularizerApplicator] = None) -> None:
super(EvidenceExtraction, self).__init__(vocab, regularizer)
self._embedder = embedder
self._question_encoder = question_encoder
self._passage_encoder = passage_encoder
# size: 2H
encoding_dim = question_encoder.get_output_dim()
self._gru_cell = nn.GRUCell(2 * encoding_dim, encoding_dim)
self._gate = nn.Linear(2 * encoding_dim, 2 * encoding_dim)
self._match_layer_1 = nn.Linear(2 * encoding_dim, encoding_dim)
self._match_layer_2 = nn.Linear(encoding_dim, 1)
self._question_attention_for_passage = Attention(
NonlinearSimilarity(encoding_dim))
self._question_attention_for_question = Attention(
NonlinearSimilarity(encoding_dim))
self._passage_attention_for_answer = Attention(
NonlinearSimilarity(encoding_dim), normalize=False)
self._passage_attention_for_ranking = Attention(
NonlinearSimilarity(encoding_dim))
self._passage_self_attention = Attention(
NonlinearSimilarity(encoding_dim))
self._self_gru_cell = nn.GRUCell(2 * encoding_dim, encoding_dim)
self._self_gate = nn.Linear(2 * encoding_dim, encoding_dim)
self._answer_net = nn.GRUCell(encoding_dim, encoding_dim)
self._v_r_Q = nn.Parameter(torch.rand(encoding_dim))
self._r = r
self._span_start_accuracy = CategoricalAccuracy()
self._span_end_accuracy = CategoricalAccuracy()
self._span_accuracy = BooleanAccuracy()
self._squad_metrics = SquadEmAndF1()
if dropout > 0:
self._dropout = torch.nn.Dropout(p=dropout)
else:
self._dropout = lambda x: x
initializer(self)
def __init__(self,
vocab: Vocabulary,
embedder: TextFieldEmbedder,
question_encoder: Seq2SeqEncoder,
passage_encoder: Seq2SeqEncoder,
feed_forward: FeedForward,
dropout: float = 0.1,
num_decoding_steps: int = 40,
initializer: InitializerApplicator = InitializerApplicator(),
regularizer: Optional[RegularizerApplicator] = None) -> None:
super(AnswerSynthesis, self).__init__(vocab, regularizer)
self._vocab = vocab
self._vocab_size = vocab.get_vocab_size() # default: tokens
self._num_decoding_steps = num_decoding_steps
self._start_token_index = self._vocab.get_token_index(START_SYMBOL)
self._end_token_index = self._vocab.get_token_index(END_SYMBOL)
self._embedder = embedder
self._question_encoder = question_encoder
self._passage_encoder = passage_encoder
encoding_dim = question_encoder.get_output_dim()
embedding_dim = embedder.get_output_dim()
self._span_start_embedding = nn.Embedding(2, 50)
self._span_end_embedding = nn.Embedding(2, 50)
self._gru_decoder = nn.GRUCell(encoding_dim + embedding_dim,
encoding_dim)
self._feed_forward = feed_forward
self._attention = Attention(NonlinearSimilarity(encoding_dim))
self._W_r = nn.Linear(embedding_dim, encoding_dim, bias=False)
self._U_r = nn.Linear(encoding_dim, encoding_dim, bias=False)
self._V_r = nn.Linear(encoding_dim, encoding_dim, bias=False)
self._max_out = Maxout(encoding_dim,
num_layers=1,
output_dims=int(encoding_dim / 2),
pool_sizes=2)
self._W_o = nn.Linear(int(encoding_dim / 2),
self._vocab_size,
bias=False)
self._squad_metrics = SquadEmAndF1()
#self._predict_acc = CategoricalAccuracy()
if dropout > 0:
self._dropout = torch.nn.Dropout(p=dropout)
else:
self._dropout = lambda x: x
initializer(self)
self._num_iter = 0