Currently, the extrapolator network is only trained to handle one-hot class vectors. However, for the extrapolation, it may be useful if the extrapolator could either ...
... handle ambiguous input classes as part of it's training, and/or ...
... output completions for different classes, not just the most probable of the last prefix step.
Currently, the extrapolator network is only trained to handle one-hot class vectors. However, for the extrapolation, it may be useful if the extrapolator could either ...