OpenExplorer
Guide
Horizon Smart Car Developer Community
Horizon Official Website
Guide
Horizon Smart Car Developer Community
Horizon Official Website
English

PTQ Conversion Tool Overview

ToolsTools Introduction
The hb_compile ToolThe hb_compile is a tool provided by Horizon, which maps floating-point models to quantized models and provides the model checking function as well as the model modification function. It supports following usage scenarios: model checking, model quantized compilation and model modification.
The hb_model_info ToolThe hb_model_info is a tool provided by Horizon, which is used to parse the dependencies and parameters of the *.hbm and the *.bc in compilation, and support the checking of *.bc deletable nodes.
The hb_verifier ToolThe hb_verifier tool is a tool used to support performing cosine similarity comparisons of model-by-layer operator outputs. The tool supports performing comparisons between onnx models and between onnx models and hbir models.
The hb_eval_preprocess ToolThe hb_eval_preprocess is a tool provided by Horizon for preprocessing image data on x86 system (certain image data operations before being fed into the model, e.g., image resizing, image cropping, image padding, etc.) when you evaluate model accuracy.
The hb_config_generator ToolThe hb_config_generator is a tool provided by Horizon for obtaining simplest yaml configuration files and yaml configuration files with all parameter defaults.
The HBRuntime InferenceThe HBRuntime is a x86-side model inference library provided by Horizon, which supports inference on the original ONNX models directly exported by commonly used training frameworks, the ONNX models at various stages generated during the PTQ conversion process of the Horizon toolchain, and the HBIR(*.bc) models generated during the Horizon toolchain conversion process.
Accuracy Debug ToolThe Accuracy Debug Tool is a tool provided by Horizon for helping you to locate accuracy problems in the model quantization process on your own. It can help you to analyze the node-granularity quantization error of the calibration model, and finally quickly locate the nodes with accuracy anomalies.