Doc_CNN_LSTM.py 文件源码

def __init__(self, embedding_mat=None, maxlen_doc=7, maxlen_sent=50, filter_length=[3, 4, 5, 6],
                 nb_filters=200, n_vocab=10000, embedding_dims=300, hidden_gru=64, n_classes=5):
        if embedding_mat is not None:
            self.n_vocab, self.embedding_dims = embedding_mat.shape
        else:
            self.n_vocab = n_vocab
            self.embedding_dims = embedding_dims
        self.maxlen_doc = maxlen_doc
        self.maxlen_sent = maxlen_sent
        self.filter_length = filter_length
        self.nb_filters = nb_filters
        self.hidden_gru = hidden_gru

        print "Building the model"
        #graph model
        model=Graph()
        model.add_input(name='input', input_shape=(self.maxlen_doc*self.maxlen_sent,), dtype='int')

        #Model embedding layer, for word index-> word embedding transformation
        model.add_node(Embedding(self.n_vocab, self.embedding_dims, weights=[self.embedding_mat],
                                 input_length=self.maxlen_sent*self.maxlen_doc),
                       name='embedding', input='input')
        model.add_node(Reshape((self.maxlen_doc, 1, self.maxlen_sent, self.embedding_dims)),
                      name='reshape_5d', input='embedding')
        #define the different filters
        conv_layer = []
        for each_length in filter_length:
            model.add_node(TimeDistributedConvolution2D(self.nb_filters/len(filter_length),
                                                        each_length, self.embedding_dims, border_mode='valid',
                                               input_shape=(self.maxlen_doc,1,self.maxlen_sent, self.embedding_dims)),
                          name='conv_{}'.format(each_length), input='reshape_5d')
            model.add_node(Activation('relu'),
                          name='relu_conv_{}'.format(each_length), input='conv_{}'.format(each_length))

            model.add_node(TimeDistributedMaxPooling2D(pool_size=(int(self.maxlen_sent - each_length+1), 1),
                          border_mode='valid'),
                          name='pool_conv_{}'.format(each_length), input='relu_conv_{}'.format(each_length))
            model.add_node(TimeDistributedFlatten(),
                          name='flatten_conv_{}'.format(each_length), input='pool_conv_{}'.format(each_length))
            conv_layer.append('flatten_conv_{}'.format(each_length))
        # model.add_node(Activation('relu'), name='relu', inputs=conv_layer)
        print conv_layer
        model.add_node(GRU(self.hidden_gru), name='gru_forward', inputs=conv_layer)
        model.add_node(GRU(self.hidden_gru, go_backwards=True), name='gru_backward', inputs=conv_layer)
        model.add_node(Dropout(0.5), name='gru_outputs', inputs=['gru_forward', 'gru_backward'])
        model.add_node(Dense(n_classes), name='full_con', input='gru_outputs')
        model.add_node(Activation('softmax'), name='prob', input='full_con')
        model.add_output(name='pred', input='prob')

        model.compile('rmsprop', loss = {'pred': 'categorical_crossentropy'})