[RMP] Potential New Examples and Example Tasks

This is a list of new example ideas and other example tasks, which are not all prioritized. If some are prioritized, we can slice them out into a smaller tickets + individual tickets per example/task.

Tasks:

• Standardize usage of tf memory allocation (+ add disclaimer)
• Cleaning NVTabular examples
• Use testbook for all notebook unit tests
• Capture metrics for ASV

WIP:

• Add an end-2-end POC example with multi-hot columns
• [Task] Example of serving LightFM in Merlin Systems NVIDIA-Merlin/models#138

Prioritized:

• How to deal with large models
• Multi-model data input (e.g. adding images, sound, text) (required pre-trained vectors)
• Benchmark Dataloaders : Merlin vs. native dataloaders
• Benchmark mult-GPU training: Scaling from 1 GPU, 2x GPUs, 4x GPUs, 8x GPUs

Non-Prioritized Potential Examples:

• Adding local predictions both for ranking and retrieval models
• Optimizing embeddings
• Use-case: Search (might part of pre-trained embeddings)
• Adding examples for different cloud environments
• Use-cases for different industries
• CPU-only examples (provide as documentation)
• POC (feature store, etc.) for large scale
• POC (feature store, etc.) on cloud

Done:

• [Task] Add XGBoost Serving Example NVIDIA-Merlin/models#136
• Sequential-based for Merlin Models (if possible with current Merlin Models)
• [Task] Add Implicit Training Example NVIDIA-Merlin/models#561
• [Task] Example of serving Implicit in Merlin Systems NVIDIA-Merlin/models#137
• [Task] Example of training LightFM in Merlin Models NVIDIA-Merlin/models#562
• Examples for RecSys challenges (previous and currently running)
• Hyperparameter Tuning
• -showing improvement overtime
• Adding new examples for NVTabular
• Using implicit algorithm (waiting for inference)
• Example for data-parallel training and impact (in progress library development)
• [Task] Add example for dataloaders for new dataloader repo NVIDIA-Merlin/dataloader#12
• Example for model-parallel training (in progress library development)
• Example Booking.com

This is an umbrella list of example ideas rather than a scoped implementation task, and it names no files, tests, or entry points. Start by reviewing the prioritized and unfinished items, then split one selected example or task into an individual issue with a defined scope. Done should be an agreed, independently actionable example or task with clear completion criteria.