LSTMwithCNN.py 文件源码

def __init__(self, word_index, embedding_matrix):
        embedding_layer_c = Embedding(len(word_index) + 1,
                                    EMBEDDING_DIM,
                                    weights=[embedding_matrix],
                                    input_length=MAX_SEQUENCE_LENGTH_C,
                                    trainable=False)
        embedding_layer_q = Embedding(len(word_index) + 1,
                                      EMBEDDING_DIM,
                                      weights=[embedding_matrix],
                                      input_length=MAX_SEQUENCE_LENGTH_Q,
                                      trainable=False)
        embedding_layer_a = Embedding(len(word_index) + 1,
                                      EMBEDDING_DIM,
                                      weights=[embedding_matrix],
                                      input_length=MAX_SEQUENCE_LENGTH_A,
                                      trainable=False)
        context = Input(shape=(MAX_SEQUENCE_LENGTH_C,), dtype='int32', name='context')
        question = Input(shape=(MAX_SEQUENCE_LENGTH_Q,), dtype='int32', name='question')
        answer = Input(shape=(MAX_SEQUENCE_LENGTH_A,), dtype='int32', name='answer')
        embedded_context = embedding_layer_c(context)
        embedded_question = embedding_layer_q(question)
        embedded_answer = embedding_layer_a(answer)

        l_lstm_c = Bidirectional(LSTM(60, return_sequences=True))(embedded_context)
        conv_blocksC = []
        for sz in [5,7]:
            conv = Convolution1D(filters=20,
                                 kernel_size=sz,
                                 padding="valid",
                                 activation="relu",
                                 strides=1)(l_lstm_c)
            conv = MaxPooling1D(pool_size=2)(conv)
            conv = Flatten()(conv)
            conv_blocksC.append(conv)

        l_lstm_q = Bidirectional(LSTM(60, return_sequences=True))(embedded_question)
        conv_blocksQ = []
        for sz in [3, 5]:
            conv = Convolution1D(filters=20,
                                 kernel_size=sz,
                                 padding="valid",
                                 activation="relu",
                                 strides=1)(l_lstm_q)
            conv = MaxPooling1D(pool_size=2)(conv)
            conv = Flatten()(conv)
            conv_blocksQ.append(conv)
        l_lstm_a = Bidirectional(LSTM(60))(embedded_answer)

        concat_c_q = concatenate([l_lstm_a] + conv_blocksQ + conv_blocksC , axis=1)
        relu_c_q_a = Dense(100, activation='relu')(concat_c_q)
        relu_c_q_a = Dropout(0.25)(relu_c_q_a)
        softmax_c_q_a = Dense(2, activation='softmax')(relu_c_q_a)
        self.model = Model([question, answer, context], softmax_c_q_a)
        opt = Nadam()
        self.model.compile(loss='categorical_crossentropy',
                      optimizer=opt,
                      metrics=['acc'])