[RMP] Using session-based models as query encoders (for downstream models or ANN search)

Prerequisites:

• Sequential models to generate embeddings (MM supports this, T4Rec Doesn't)
• ANN lookup Op vs TopK lookup in the Model (MM supports this for retrieval; needs to be generalized to all models, Sara has TopK layer for PyTorch which can be added to T4Rec)

Tasks:

• Generation of Embeddings in Transformers4Rec Merlin Models PyTorch API
• Generalize TopK to all model types doing multiclass classification in Merlin Models ( P2 )

Examples

• NVIDIA-Merlin/models#1097
• Output of the the model being fed into an ANN search
• Output of the model being saved as a pre trained embedding and loaded into a later model using our pretrained embedding functionality in the dataloader.

Notes from our discussions on this topic:

• We need to decide whether we should extend T4Rec to support these use cases (ADR)
• Starting point: This spreadsheet provides an overview for multi-stage session-based recommendation use cases, requirements, support in existing libraries and tasks.