In tf.Example, this is intrinsically supported by a feature with N entries in a list field. Think about what is the best way to support this in csv.
Couple of options come to mind:
- A ':' delimited set of numbers in a column. (pros: easy to produce; cons: requires more expensive string splitting)
- A sequence of bytes encoded as base64 into a column. (pros: more painful to produce; still requires base64 decoding but cheaper)
Scenarios
- Text run through a word2vec transform during data preparation.
- Image run through a image2vec (aka use inception model to extract features from image)