Open Oufattole opened 2 months ago
This is very interesting, but most challenging and potentially has more dependencies than other experimental questions (in other ways it is more independent as it is just a question of what you do with your output embeddings).
In particular, dependencies on
All make this potentially more challenging (and certainly more expensive) to explore -- that all being said, this is likely one of the highest impact questions that is "purely" about the learning process.
We can add additional classes that inherit the supervised model class For now the implementation will be Triplet tokenizaiton specific, but we should generalize it in the future. These are the methods: