models.py 文件源码

def create_concat_model(self, emb_dim, emb_path, vocab_word,
                            vocab_word_size, word_maxlen, vocab_char_size,
                            char_maxlen):
        from aes.layers import Conv1DMask, MaxPooling1DMask
        from keras.layers import concatenate
        logger.info('Building concatenation model')
        input_char = Input(shape=(char_maxlen, ), name='input_char')
        char_emb = Embedding(
            vocab_char_size, emb_dim, mask_zero=True)(input_char)
        char_cnn = Conv1DMask(
            filters=emb_dim, kernel_size=3, padding='same')(char_emb)
        char_input = MaxPooling1DMask(
            pool_size=char_maxlen / word_maxlen, padding='same')(char_cnn)
        input_word = Input(shape=(word_maxlen, ), name='input_word')
        word_input = Embedding(
            vocab_word_size, emb_dim, mask_zero=True,
            name='word_emb')(input_word)
        merged = concatenate([char_input, word_input], axis=1)
        merged_dropped = Dropout(0.5)(merged)
        final_input = Dense(50)(merged_dropped)
        cnn = Conv1DMask(
            filters=emb_dim, kernel_size=3, padding='same')(final_input)
        dropped = Dropout(0.5)(cnn)
        mot = MeanOverTime(mask_zero=True)(dropped)
        densed = Dense(self.num_outputs, name='dense')(mot)
        output = Activation('sigmoid')(densed)
        model = Model(inputs=[input_char, input_word], outputs=output)
        model.get_layer('dense').bias.set_value(self.bias)
        if emb_path:
            from emb_reader import EmbReader as EmbReader
            logger.info('Initializing lookup table')
            emb_reader = EmbReader(emb_path, emb_dim=emb_dim)
            model.get_layer('word_emb').embeddings.set_value(
                emb_reader.get_emb_matrix_given_vocab(
                    vocab_word,
                    model.get_layer('word_emb').embeddings.get_value()))
        logger.info('  Done')
        return model