Customized Feature Column

Customized Feature Column

TF2.0 users generally define model using Keras. tf.keras.layers.Embedding is the native embedding layer in Keras. It can only handle the dense input. For sparse inputs, user often use tf.feature_column.embedding_columns to convert them to dense representation to feed to a DNN.

For ElasticDL, user define the model using keras. And we have provide the elastic.layers.Embedding to interact with the ElasticDL parameter server and partition the embedding table among multiple PS instances. It can replace the native keras embedding layer but can't replace the embedding_column.

In this doc, we are focuing on how to write a customized feature column to interact with the parameter server and how to replace the native feature column with ours.

How to replace feature column

How to write a customzied feature column