def forward()

in drqa/reader/rnn_reader.py [0:0]

• 24 lines of code
• 8 McCabe index (conditional complexity)

    def forward(self, x1, x1_f, x1_mask, x2, x2_mask):
        """Inputs:
        x1 = document word indices             [batch * len_d]
        x1_f = document word features indices  [batch * len_d * nfeat]
        x1_mask = document padding mask        [batch * len_d]
        x2 = question word indices             [batch * len_q]
        x2_mask = question padding mask        [batch * len_q]
        """
        # Embed both document and question
        x1_emb = self.embedding(x1)
        x2_emb = self.embedding(x2)

        # Dropout on embeddings
        if self.args.dropout_emb > 0:
            x1_emb = nn.functional.dropout(x1_emb, p=self.args.dropout_emb,
                                           training=self.training)
            x2_emb = nn.functional.dropout(x2_emb, p=self.args.dropout_emb,
                                           training=self.training)

        # Form document encoding inputs
        drnn_input = [x1_emb]

        # Add attention-weighted question representation
        if self.args.use_qemb:
            x2_weighted_emb = self.qemb_match(x1_emb, x2_emb, x2_mask)
            drnn_input.append(x2_weighted_emb)

        # Add manual features
        if self.args.num_features > 0:
            drnn_input.append(x1_f)

        # Encode document with RNN
        doc_hiddens = self.doc_rnn(torch.cat(drnn_input, 2), x1_mask)

        # Encode question with RNN + merge hiddens
        question_hiddens = self.question_rnn(x2_emb, x2_mask)
        if self.args.question_merge == 'avg':
            q_merge_weights = layers.uniform_weights(question_hiddens, x2_mask)
        elif self.args.question_merge == 'self_attn':
            q_merge_weights = self.self_attn(question_hiddens, x2_mask)
        question_hidden = layers.weighted_avg(question_hiddens, q_merge_weights)

        # Predict start and end positions
        start_scores = self.start_attn(doc_hiddens, question_hidden, x1_mask)
        end_scores = self.end_attn(doc_hiddens, question_hidden, x1_mask)
        return start_scores, end_scores