sst1_cnn_rnn.py 文件源码

def build_model():
    main_input = Input(shape=(maxlen, ), dtype='int32', name='main_input')
    embedding  = Embedding(max_features, embedding_dims,
                  weights=[np.matrix(W)], input_length=maxlen,
                  name='embedding')(main_input)

    embedding = Dropout(0.50)(embedding)

    conv4 = Convolution1D(nb_filter=nb_filter,
                          filter_length=4,
                          border_mode='valid',
                          activation='relu',
                          subsample_length=1,
                          name='conv4')(embedding)
    maxConv4 = MaxPooling1D(pool_length=2,
                             name='maxConv4')(conv4)

    conv5 = Convolution1D(nb_filter=nb_filter,
                          filter_length=5,
                          border_mode='valid',
                          activation='relu',
                          subsample_length=1,
                          name='conv5')(embedding)
    maxConv5 = MaxPooling1D(pool_length=2,
                            name='maxConv5')(conv5)

    x = merge([maxConv4, maxConv5], mode='concat')

    x = Dropout(0.15)(x)

    x = RNN(rnn_output_size)(x)

    x = Dense(hidden_dims, activation='relu', init='he_normal',
              W_constraint = maxnorm(3), b_constraint=maxnorm(3),
              name='mlp')(x)

    x = Dropout(0.10, name='drop')(x)

    output = Dense(nb_classes, init='he_normal',
                   activation='softmax', name='output')(x)

    model = Model(input=main_input, output=output)
    model.compile(loss={'output':'categorical_crossentropy'},
                optimizer=Adadelta(lr=0.95, epsilon=1e-06),
                metrics=["accuracy"])
    return model