Supported dtypes? about gnn HOT 1 CLOSED

Hi @DiTo97,

Interesting question, (and we do care about those poor bits). I would say that we already support different dtypes, but it does depend on how one uses them.

For instance, in the article you linked there's two concepts I would like to disambiguate:

• For node or edge features, the dtypes are generally limited by what can be stored inside a tf.train.Example. However if one really is interested at efficiency at this level, after deserializing I think smaller dtypes can be coerced. I'd need to see some exact errors to understand what's going on.

• tfgnn.Adjacency.from_indices. The arguments to this must be more than one bit each since they take the node ids (integers) to create the relationships. Supporting int16 for tiny graphs might allow a potential speedup.

from gnn.