model.py 文件源码

def create_model(self,):
        """
        RNN model creation
        Layers include Embedding Layer, 3 LSTM stacked,
        Simple Context layer (manually defined),
        Time Distributed Layer
        """
        length_vocab, embedding_size = self.word2vec.shape
        print ("shape of word2vec matrix ", self.word2vec.shape)

        model = Sequential()

        # TODO: look at mask zero flag
        model.add(
                Embedding(
                        length_vocab, embedding_size,
                        input_length=max_length,
                        weights=[self.word2vec], mask_zero=True,
                        name='embedding_layer'
                )
        )

        for i in range(rnn_layers):
            lstm = LSTM(rnn_size, return_sequences=True,
                name='lstm_layer_%d' % (i + 1)
            )

            model.add(lstm)
            # No drop out added !

        model.add(Lambda(self.simple_context,
                     mask=lambda inputs, mask: mask[:, max_len_desc:],
                     output_shape=self.output_shape_simple_context_layer,
                     name='simple_context_layer'))

        vocab_size = self.word2vec.shape[0]
        model.add(TimeDistributed(Dense(vocab_size,
                                name='time_distributed_layer')))

        model.add(Activation('softmax', name='activation_layer'))

        model.compile(loss='categorical_crossentropy', optimizer='adam')
        K.set_value(model.optimizer.lr, np.float32(learning_rate))
        print (model.summary())
        return model