Facebook’s 2019 blueprint for mixing sparse embeddings and dense MLPs

What it does

DLRM is Facebook Research’s reference implementation of a deep learning recommendation model that predicts the probability of a click. It marries dense floating-point features, processed by bottom MLPs, with sparse categorical indices that pull vectors from embedding tables. Those vectors interact via explicit operators—Sum, Dot, or Cat—before a final top MLP emits the prediction.

The interesting bit

The repository treats the model as a benchmark artifact rather than a standalone product. It ships parallel PyTorch and Caffe2 implementations, and the README explicitly anchors follow-on research on system architecture and embedding compression to this exact code.

Key highlights

• Dual implementations in PyTorch and Caffe2, plus data loaders and benchmark scripts.
• Pre-trained weights available for the Criteo Terabyte dataset under a CC-BY-NC license.
• Built-in support for the Criteo Kaggle and Terabyte click-log datasets.
• Interaction operators (Sum, Dot, Cat) explicitly define how sparse embeddings mingle with dense features.
• Cited as the reference benchmark in multiple research papers on hardware implications and memory-efficient embeddings.

Caveats

• Documentation beyond quickstart commands and benchmarking scripts is thin; tuning guidance lives in the original paper.
• The pre-trained checkpoint is restricted to the Terabyte dataset and carries a non-commercial license.
• It is unclear from the README whether both backends receive equal maintenance attention.

Verdict

Researchers studying recommender-system architecture, embedding compression, or hardware implications should start here. If you need a modern, production-grade recommendation framework rather than a 2019 paper reference, this is not it.