[RMP] Merlin Models retrieval and ranking models profiling

Problem:

So far Merlin Models team have been focused on API design for modularity/flexibility and support to multiple retrieval and ranking neural architectures.
There has not been yet an active effort on profiling training and evaluation pipelines with Merlin Models (TF API), from the perspective of GPU compute and memory utilization.

Goal:

Profiling of Merlin Models training and evaluation pipelines of retrieval and ranking models, so that we might identify potential performance bottlenecks in TF/Merlin Model ops.

Constraints:

For such benchmarking it will be using assets from the research experiments:

• Scripts for training and evaluating retrieval and ranking models, which make advanced usage of Merlin Models API.
• Preprocessed public datasets suitable for each task, with a reasonable number of samples and features

Tasks

• https://github.com/NVIDIA-Merlin/models/issues/340