cnn_dssm_th.py 文件源码

def model(sequence_length=None):
    graph = Graph()
    graph.add_input(name='input', input_shape=(sequence_length, embedding_dim))
    for fsz in filter_sizes:
        conv = Convolution1D(nb_filter=num_filters,
                             filter_length=fsz,
                             border_mode='valid',
                             activation='relu',
                             subsample_length=1,
                             input_dim=embedding_dim,
                             input_length=sequence_length)
        pool = MaxPooling1D(pool_length=sequence_length - fsz + 1)
        graph.add_node(conv, name='conv-%s' % fsz, input='input')
        graph.add_node(pool, name='maxpool-%s' % fsz, input='conv-%s' % fsz)
        graph.add_node(
            Flatten(),
            name='flatten-%s' %
            fsz,
            input='maxpool-%s' %
            fsz)

    if len(filter_sizes) > 1:
        graph.add_output(name='output',
                         inputs=['flatten-%s' % fsz for fsz in filter_sizes],
                         merge_mode='concat')
    else:
        graph.add_output(name='output', input='flatten-%s' % filter_sizes[0])

    # main sequential model
    model = Sequential()

    if conf.feature_level == "word":
        model.add(
            Embedding(
                vocab_size,
                embedding_dim,
                input_length=sequence_length,
                weights=[embedding_weights]))
    elif conf.feature_level == "char" or conf.feature_level == "ngram":
        model.add(
            Embedding(
                vocab_size,
                embedding_dim,
                input_length=sequence_length))


    model.add(
        Dropout(
            dropout_prob[0],
            input_shape=(
                sequence_length,
                embedding_dim)))
    model.add(graph)
    model.add(Dense(hidden_dims))
    # model.add(Dropout(dropout_prob[1]))
    model.add(Activation('relu'))
    return model

# Input Layer with all the query, similar and non similar documents.