Supporting assets for model deployment benchmarking — TensorRT export, timing methodology, and configuration used for the real latency numbers in this article.
Exports a PyTorch checkpoint to a TensorRT engine at a given precision, reusing an existing engine if one is already present.
DeploymentWarmup-then-measure latency benchmarking with median, P95, and FPS computed from per-iteration timings.
EvaluationThe exact resolution, warmup/iteration counts, and confidence/IoU thresholds used to produce the latency numbers in this article.
ReferenceSelected resources are designed to support experimentation and practical understanding through focused implementations, utilities, and companion materials accompanying each article.