| 107 | |
| 108 | |
| 109 | class RNNEncoder(nn.Module): |
| 110 | |
| 111 | def __init__(self, bidirectional, num_layers, input_size, |
| 112 | hidden_size, dropout=0.0): |
| 113 | super(RNNEncoder, self).__init__() |
| 114 | num_directions = 2 if bidirectional else 1 |
| 115 | assert hidden_size % num_directions == 0 |
| 116 | hidden_size = hidden_size // num_directions |
| 117 | |
| 118 | self.rnn = LayerNormLSTM( |
| 119 | input_size=input_size, |
| 120 | hidden_size=hidden_size, |
| 121 | num_layers=num_layers, |
| 122 | bidirectional=bidirectional) |
| 123 | |
| 124 | self.wo = nn.Linear(num_directions * hidden_size, 1, bias=True) |
| 125 | self.dropout = nn.Dropout(dropout) |
| 126 | self.sigmoid = nn.Sigmoid() |
| 127 | |
| 128 | def forward(self, x, mask): |
| 129 | """See :func:`EncoderBase.forward()`""" |
| 130 | x = torch.transpose(x, 1, 0) |
| 131 | memory_bank, _ = self.rnn(x) |
| 132 | memory_bank = self.dropout(memory_bank) + x |
| 133 | memory_bank = torch.transpose(memory_bank, 1, 0) |
| 134 | |
| 135 | sent_scores = self.sigmoid(self.wo(memory_bank)) |
| 136 | sent_scores = sent_scores.squeeze(-1) * mask.float() |
| 137 | return sent_scores |