Hand-coding YOLO and ResNet in TensorRT, layer by obsessive layer

What it does

TensorRTx is a collection of reference implementations for running well-known computer vision models—ResNet, YOLO variants, MobileNet, ViT, and others—using NVIDIA’s TensorRT C++ API. Instead of converting PyTorch or TensorFlow graphs through an ONNX parser, you export weights to a plain-text .wts file and define every layer manually in code. The result is a compiled TensorRT engine that runs inference without the usual parser intermediaries.

The interesting bit

The project treats parser-free construction as a feature, not a burden: the author argues that hand-rolling the network makes it easier to merge layers, swap operations, or wire preprocessing directly into the graph, and to inspect intermediate tensors when things go wrong. It is essentially a hands-on textbook for how these networks map to TensorRT primitives.

Key highlights

• Covers dozens of models, from LeNet and AlexNet to YOLOv11, YOLOv9, ViT, and DETR, with weights often sourced from PyTorch reference repos.
• Uses a custom .wts text format to ferry weights from PyTorch/TensorFlow into the TensorRT builder.
• Explicitly targets TensorRT 8.x as the stable compatibility sweet spot; TensorRT 7.x and 10.x are marked as not recommended, and some models are pinned to specific SDK versions.
• Includes newer experimental ports such as a YOLO11 example written in Tripy, NVIDIA’s newer Python programming model for TensorRT.

Caveats

• The top-level CMake build that compiles multiple subprojects at once is still under development, and the README warns that most subprojects are not adapted to it yet.
• Model support is fragmented across TensorRT versions; you need to check individual subfolder READMEs to see whether your SDK version matches.

Verdict

Grab this if you are optimizing inference pipelines and need to surgically alter a network graph beyond what an ONNX parser allows, or if you are trying to understand how modern CV architectures decompose into TensorRT layers. Skip it if you just want a one-click converter and do not care about the internals.