How nvlabs/neural_rx works
Project type and purpose
• This is a research‐driven machine‐learning pipeline and inference framework for 5G NR multiuser MIMO receivers.
• It consists of:
• A training and evaluation pipeline (scripts + configs + utils + Sionna/TF)
• A model repository (pretrained weights + ONNX/TensorRT export)
• Notebooks for tutorials and result visualization.
• Results and weights storage.
Break down the main components
a. Configuration Layer
• Directory: config/
• Purpose: parameterize experiments (channel models, modulation, MCS, site‐specific data).
b. Data & Channel Simulation
• Provided by NVIDIA Sionna (external library).
• Local channel‐realization datasets generated at runtime or from data/ (ray‐tracing).
c. Model Definition & Utilities
• Directory: utils/
• Contains architecture building blocks: graph‐CNN modules, impairment models, baseline receivers, ONNX helpers.
d. Training & Evaluation Pipeline
• Directory: scripts/
• Scripts:
• train_neural_rx.py – loads config, builds model (utils), trains in TF.
• evaluate.py – runs test sets, computes BLER, outputs to results/.
• export_onnx.py – converts TF model to ONNX.
• compute_cov_mat.py, export_onnx.py, evaluate.py supporting tasks.
e. Storage
• weights/ – pretrained model checkpoints for various configs.
• results/ – precomputed BLER curves and sweeps.
• Generated at runtime:
• logs/ – training logs, tensorboard.
• onnx_models/ – exported ONNX files.
• data/ – ray‐tracing channel sets.
f. Notebook Interface
• Directory: notebooks/
• Interactive demos: Jumpstart tutorial, model introspection, real‐time latency plots.
g. Deployment / Inference Runtime
• TensorRT engine loading ONNX for real‐time on GPU.
• Latency target: ≤1 ms on NVIDIA A100.
Relationships and data flows
• Config → Training: configs parameterize train_neural_rx.py.
• Training → Weights: trained TF model saved to weights/.
• Weights → ONNX: export_onnx.py loads weights → produces ONNX → stored in onnx_models/.
• ONNX → TensorRT: external step (not in repo) to build TRT engine.
• Data (Sionna + ray tracing) → Training/Evaluation: channel samples fed to model.
• Evaluation → Results: evaluate.py writes metrics to results/.
• Notebooks consume: configs, scripts, weights, results to display plots and latency figures.
Architectural patterns
• Modular ML pipeline: clear separation between data, model logic, training, evaluation, export.
• Two deployment paths:
• Offline: research via notebooks & scripts.
• Real-time: ONNX → TensorRT for low‐latency GPU inference.
• Use of external frameworks as “services”: Sionna (channel simulation), TensorFlow (training), ONNX/TensorRT (inference).