I use tensorflow in version 1.3 and use a Wide' n' Deep Estimator for my network. The API allows the definition of tf.feature_columns. Now I have single fields, which can have several entries. These are a free combination of words, i.e. a text where the appearance of the words is relevant to me. According to the documentation, there is only the possibility to convert the entire field into a hash---embedding_column with the bucket-size or categorial_column_with_hash_bucket---however, for me the individual words are relevant as input, not the text itself.
Example:
3 input of the same field: [hello world] [world is fine] [make software great again]
Categorial inputs I want: [hello, world, is, fine, make, software, great, again]
Is there a way to pass a set of elements in a column as input, so that tensorflow takes care of each word individually and categorically? In my opinion, the alternative is preprocessing, where I have to create 1000 of sparse columns and distribute the words to these columns individually.
Many thanks for your support