As discussed, I have added some training_<key> where key refers to a namespace key with a production rule - things like base_dist etc.
Since we don't actually have options for base or target I just removed the explicit node stuff there. Even if there was a choice, we don't technically need it because we could just have the logic that chose the action inside that production rule.
I then have changed the model to load from the training_context. This didn't break much apart from the tests, because when you use the API there is no training output but you may want to declare a model - in this instance there is a new function called explicit_model which should be called. This might mean that some examples or docs are out of date (already!) but we can worry about that after we agree this approach is what we want.
The train, report, and benchmark all seem to be working fine with this. Let me know what you think.
As discussed, I have added some
training_<key>where key refers to a namespace key with a production rule - things likebase_distetc.Since we don't actually have options for base or target I just removed the explicit node stuff there. Even if there was a choice, we don't technically need it because we could just have the logic that chose the action inside that production rule.
I then have changed the model to load from the training_context. This didn't break much apart from the tests, because when you use the API there is no training output but you may want to declare a model - in this instance there is a new function called
explicit_modelwhich should be called. This might mean that some examples or docs are out of date (already!) but we can worry about that after we agree this approach is what we want.The train, report, and benchmark all seem to be working fine with this. Let me know what you think.
closes #62