deepstream-import-vision-model

What problem does it solve?

This Skill solves the time-consuming, error-prone process of converting a HuggingFace or NVIDIA NGC vision model into a working, benchmarked NVIDIA DeepStream pipeline.

Core Features & Use Cases

• Model acquisition (HF/NGC): Downloads the right model artifacts and extracts class labels from config.json.
• ONNX + TensorRT engine automation: Exports SafeTensors to ONNX when needed, builds a dynamic TRT engine, benchmarks with fixed log filenames, and derives throughput-based capacity.
• DeepStream deployment + parsing: Generates and compiles a custom nvinfer bbox parser and runs validated single-stream and multi-stream performance benchmarks, producing a complete HTML+PDF report.

Quick Start

Use the deepstream-import-vision-model skill to import a HuggingFace object-detection model, build a TensorRT engine, run DeepStream benchmarks on sample_720p.mp4, and generate a styled HTML+PDF benchmark report in models/<model_name>/reports/.