- [x] Make a decision around how much we want to do in the model vs how much we want to do in operation graphs? - [x] Come up with a consistent way to do padding (in the dataloader) so that we can use the same function for training, dataloading and serving - [x] Establish how we handle data transformations for small batches in a relatively standard way - [x] Dictionary/dataframe translation - [x] Optimizing Merlin operators for serving and dataloading (small batch) - [ ] NVTabular - [Simplify and standardize Categorify(), so that unique values parquet can be used for reverse categorify](https://github.com/NVIDIA-Merlin/NVTabular/issues/1748)