Explore different numerical Value embedding strategies
Opened this issue · 2 comments
Oufattole commented
We currently only support Continuous value embeddings (a one to many FFN). We should try other things, like supporting quantizing.
Oufattole commented
follow the hydra target schema:
meds-torch/configs/model/ocp.yaml
Line 7 in 3533dfa
mmcdermott commented
priority tag specifically commenting on comparing ways to turn a coninuous value (code-specific or code-independent) into a vector embedding (e.g., scaling a code-independent embedding vector, scaling a code-specific embedding vector, using an FFN, etc).