Possible improvements #47
Replies: 1 comment
-
Two-Stages
DataPreparator
Indexers DSSM/neural architecture with features Pandas as an input |
Beta Was this translation helpful? Give feedback.
-
Two-Stages
DataPreparator
Indexers DSSM/neural architecture with features Pandas as an input |
Beta Was this translation helpful? Give feedback.
-
Cold Scenario
Having a cold model inside every scenario is clunky and unneccessary, this should be extracted into a separate class similar to
Fallback
.Two Stages
The code scares me, it should be refactored into separate logical functions. I think one should be able to train second level reranking, having trained base models. (i.e. without providing splitters and training everything from scratch). This will also allow for separate retraining of first level and second level.
Data preparator
It currently provides questionable utility and can be safely removed. It is not clear why do we need it atm.
Indexer
There are a lot of conversions of indexers in every model. One way to solve this is to force user to convert indexes before training the models and convert back after getting predictions. We will have to store extra object, but model logic will be eased.
Ignite
We don't actually profit much from having ignite for neural models and we can probably benefit if we write our own training loop. This will probably make the code clearer.
Beta Was this translation helpful? Give feedback.
All reactions