def __init__()

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

• 50 lines of code
• 9 McCabe index (conditional complexity)

    def __init__(self, args, normalize=True):
        super(RnnDocReader, self).__init__()
        # Store config
        self.args = args

        # Word embeddings (+1 for padding)
        self.embedding = nn.Embedding(args.vocab_size,
                                      args.embedding_dim,
                                      padding_idx=0)

        # Projection for attention weighted question
        if args.use_qemb:
            self.qemb_match = layers.SeqAttnMatch(args.embedding_dim)

        # Input size to RNN: word emb + question emb + manual features
        doc_input_size = args.embedding_dim + args.num_features
        if args.use_qemb:
            doc_input_size += args.embedding_dim

        # RNN document encoder
        self.doc_rnn = layers.StackedBRNN(
            input_size=doc_input_size,
            hidden_size=args.hidden_size,
            num_layers=args.doc_layers,
            dropout_rate=args.dropout_rnn,
            dropout_output=args.dropout_rnn_output,
            concat_layers=args.concat_rnn_layers,
            rnn_type=self.RNN_TYPES[args.rnn_type],
            padding=args.rnn_padding,
        )

        # RNN question encoder
        self.question_rnn = layers.StackedBRNN(
            input_size=args.embedding_dim,
            hidden_size=args.hidden_size,
            num_layers=args.question_layers,
            dropout_rate=args.dropout_rnn,
            dropout_output=args.dropout_rnn_output,
            concat_layers=args.concat_rnn_layers,
            rnn_type=self.RNN_TYPES[args.rnn_type],
            padding=args.rnn_padding,
        )

        # Output sizes of rnn encoders
        doc_hidden_size = 2 * args.hidden_size
        question_hidden_size = 2 * args.hidden_size
        if args.concat_rnn_layers:
            doc_hidden_size *= args.doc_layers
            question_hidden_size *= args.question_layers

        # Question merging
        if args.question_merge not in ['avg', 'self_attn']:
            raise NotImplementedError('merge_mode = %s' % args.merge_mode)
        if args.question_merge == 'self_attn':
            self.self_attn = layers.LinearSeqAttn(question_hidden_size)

        # Bilinear attention for span start/end
        self.start_attn = layers.BilinearSeqAttn(
            doc_hidden_size,
            question_hidden_size,
            normalize=normalize,
        )
        self.end_attn = layers.BilinearSeqAttn(
            doc_hidden_size,
            question_hidden_size,
            normalize=normalize,
        )