FeiYull/TensorRT-Alpha

TensorRT‑Alpha – Fast TensorRT Deployments for Vision Models

What it is – A collection of C++/CUDA examples that show how to take popular computer‑vision networks (YOLO series, EfficientDet, U‑2‑Net, face‑mesh, etc.), convert them from PyTorch → ONNX → TensorRT, and run them with custom multi‑batch preprocessing, decoding and NMS on NVIDIA GPUs.

Key points

• Supported platforms – Ubuntu 18.04 and Windows 10 with CUDA 11.3, cuDNN 8.2, TensorRT 8.4.2.4, OpenCV 3/4, and a recent GCC or Visual Studio.
• Models – More than 30 pre‑built pipelines are provided (YOLOv3‑v8, YOLOX, YOLO‑R, EfficientDet, PP‑HumanSeg, U‑2‑Net, libfacedetection, facemesh, etc.). The repo also hosts ready‑made ONNX files via Weiyun or Google Drive links.
• Performance focus – The code implements its own CUDA kernels for image pre‑processing, inference, and post‑processing (decode + NMS) to minimise latency. Example numbers are shown for a RTX 2070 Mobile (yolov8‑n, 1080p video, 8‑batch, ~14 % GPU utilisation, ~1 GB VRAM).
• Accuracy – The authors claim “strict precision alignment” with the original PyTorch/ONNX results; side‑by‑side screenshots compare official outputs with TensorRT‑Alpha outputs for several models.
• How to get started –
  1. Create a Python 3.8 conda environment (optional) and install the Python requirements.
  2. Clone the repo, edit cmake/common.cmake to point TensorRT_ROOT at your TensorRT installation.
  3. Build with CMake and run the example README for a model (e.g., yolov8/README.md).
• Installation docs – Detailed step‑by‑step guides are provided for Ubuntu 18.04 and a Dockerfile for Linux users.
• Citation – A ready‑to‑copy BibTeX entry is included for academic use.

Who might use it – Engineers who need low‑latency inference of vision models on NVIDIA hardware and prefer a C++/CUDA codebase that already handles batch preprocessing and NMS, without writing their own TensorRT plugins.

---

All information above is taken directly from the repository’s README.