You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This is still an open question to how we will implement this.
To me, the easiest way to define a decoder-only or encoder-only domain would be to not implement the encode or decode of the domain (default implementation raises NoImplementedError).
Then the encode_domains would try to encode and skip if NoImplementedError is raised.
The question is how do we handle data. We certainly want to do translations toward decoder-only domains so raw representations should be available, but then we can't compare with the GW representation because this is not available. Do we need a special case for those where the error happens after decoding of the domain?
In this case it also raises the question of how data is given to the model. For now everything is given at once because everything can be inputs or outputs. We might need to separate them input "input" domains and "target" domains.
An alternative would be to infer this in the data based on the domain (if it's decoder only, data is target).
Now every DomainModule has to encode and decode representations.
We should have a way to tell to the modules whether they are encoders, decoders or both.
The text was updated successfully, but these errors were encountered: