CATEGORICAL_ONLY v. NUMERICAL_ONLY v. CATEGORICAL_AND_NUMERICAL is confusing.

[mmcdermott]

This has two disparate impacts.

First, in embedding, if embeddings are split by categorical and numerical, then this really means you only embed the indicated things. If the embeddings aren't split, then NUMERICAL_ONLY and CATEGORICAL_AND_NUMERICAL are equivalent at the input layer, as there is no way to not embed the categorical data alongside the numerical. In the joint case, CATEGORICAL ONLY gives all values a value of 1, which is odd.

Then, at generation time, it indicates what is generated / predicted in each given layer of the dep graph.