Convolutional Network for Sentence Classification (Keras)

ngram_cnn.py

from keras.models import Graph
from keras.layers import containers
from keras.layers.core import Dense, Dropout, Activation, Reshape, Flatten
from keras.layers.embeddings import Embedding
from keras.layers.convolutional import Convolution2D, MaxPooling2D

def ngram_cnn(n_vocab, max_length, embedding_size, ngram_filters=[2, 3, 4, 5], n_feature_maps=100, dropout=0.5, n_hidden=15):
    """A single-layer convolutional network using different n-gram filters.

    Parameters
    ----------
    n_vocab: int
        Number of words in the corpus vocabulary.
    max_length: int
        Maximum sentence length in the corpus.
    embedding_size: int
        Size of the dense embedding layer.
    ngram_filters: iterable
        N-gram filter sizes for the convolutional layers.
    n_feature_maps: int
        The number of feature maps used for each filter.
    dropout: float
        Dropout probability for the dropout layer after the conv layers.
    n_hidden: int
        Number of hidden units used in the fully-connected layer

    References
    ----------
    A Sensitivity Analysis of Convolutional Neural Networks for Sentence Classification.
         Ye Zhang, Byron C. Wallace. <http://arxiv.org/pdf/1510.03820v3.pdf>
    Convolutional Neural Networks for Sentence Classification.
         Yoon Kim. <http://arxiv.org/pdf/1408.5882v2.pdf>
    """
    model = Graph()
    model.add_input(name='data', input_shape=(max_length,), dtype='int32')

    model.add_node(Embedding(n_vocab, embedding_size, input_length=max_length),
                             name='embedding', input='data')
    model.add_node(Reshape((1, max_length, embedding_size)),
                   name='reshape_embedding', input='embedding')

    # convolutional n-grams
    for n_gram in ngram_filters:
        sequential = containers.Sequential()
        sequential.add(Convolution2D(n_feature_maps, n_gram, embedding_size, input_shape=(1, max_length, embedding_size)))
        sequential.add(Activation('relu'))
        sequential.add(MaxPooling2D(pool_size=(max_length - n_gram + 1, 1)))
        sequential.add(Flatten())  # this generates a single number from the n-gram level convolution

        model.add_node(sequential, name=('unit_' + str(n_gram)), input='reshape_embedding')

    model.add_node(Dropout(dropout), name='dropout', inputs=['unit_' + str(n) for n in ngram_filters])

    # fully-connected module
    fc = containers.Sequential()
    fc.add(Dense(n_hidden, input_shape=(n_feature_maps * len(ngram_filters),)))
    fc.add(Activation('relu'))
    fc.add(Dense(1))
    fc.add(Activation('sigmoid'))

    model.add_node(fc, name='fully_connected', input='dropout')
    model.add_output(name='output', input='fully_connected')

    model.compile(loss={'output': 'binary_crossentropy'},
                  optimizer='adam')

    return model