Multiple embedding layers in keras

问题

With pretrained embeddings, we can specify them as weights in keras' embedding layer. To use multiple embeddings, would specifying multiple embedding layer be suitable? i.e.

embedding_layer1 = Embedding(len(word_index) + 1,
                        EMBEDDING_DIM,
                        weights=[embedding_matrix_1],
                        input_length=MAX_SEQUENCE_LENGTH,
                        trainable=False)

 embedding_layer2 = Embedding(len(word_index) + 1,
                        EMBEDDING_DIM,
                        weights=[embedding_matrix_2],
                        input_length=MAX_SEQUENCE_LENGTH,
                        trainable=False)

 model.add(embedding_layer1)
 model.add(embedding_layer2)

This suggests to sum them up and represent them into a single layer, which is not what I am after.

回答1:

Here is an example of using multiple embedding layers through multiple inputs by leveraging Keras' functional API. This is for a Kaggle competition so you'll have to read through the code. They feed the network a dictionary with a key for each data input. It's quite clever and I was able to build a separate model using this framework that performed well.

deep-learning-support-9663

来源：https://stackoverflow.com/questions/49805424/multiple-embedding-layers-in-keras