Horizon Journey5 Algorithm Toolchain

Horizon Journey 5 (J5) algorithm toolchain is a algorithm solution developed based on the Horizon Journey 5 processor, which includes two major components: Model Algorithm Processing and Embedded Model Prediction Library.

• The model algorithm processing provides PTQ (Post Training Quantization) and QAT (Quantization Aware Training) schemes. The converted fixed-point models can be executed on Horizon computing platforms after processing with the model compilation tools.

• The embedded model prediction library provides a series of supporting interfaces for completing inference using fixed-point models.

To allow developers to quickly get started with the key steps of model training/conversion, deployment, validation, and inference, the components of the algorithm toolchain also provides Sample packages as well as Docker images and model distributions to be used with the sample packages.

• The sample packages encapsulate the core business logic and key configuration parameters of each tool into scripts.

• The Docker image comes with preset development environments required by the sample packages.

• The model distribution object contains rich built-in algorithmic models.

This document thoroughly introduces the J5 algorithm toolchain and its related contents. The content structure of the manual is listed below.

Catalog:

• 1. OE Document Introduction
• 2. Product Introduction
  • 2.1. OpenExplorer Introduction
  • 2.2. Toolchain Overview
• 3. Key Concepts
• 4. Environmental Installation
  • 4.1. Environment Deployment
  • 4.2. Version Management of OE Package
• 5. Quick Start
  • 5.1. PTQ and QAT Introduction
  • 5.2. Algorithm Model PTQ + On-board Deployment Quick Start
  • 5.3. Algorithm Model QAT + Deployment Quick Start
  • 5.4. J5 Computing Platform Evaluation Guide
• 6. Post-training Quantization (PTQ)
  • 6.1. PTQ Conversion Principle And Process
  • 6.2. PTQ Conversion Tools Guide
  • 6.3. PTQ Conversion Steps
  • 6.4. PTQ Conversion Samples Guide
  • 6.5. Custom OP Development
  • 6.6. FAQ And Common Failure Resolutions
  • 6.7. Appendix
• 7. Quantized Awareness Training (QAT)
  • 7.1. Environmental Dependence
  • 7.2. Quick Start
  • 7.3. TUTORIAL
  • 7.4. API REFERRENCE
  • 7.5. NOTE
  • 7.6. Calibration
• 8. Model Performance/Accuracy Tuning Guide
  • 8.1. Model Performance Optimization
  • 8.2. Efficient Model Design Advice
  • 8.3. PTQ Model Accuracy Optimization
  • 8.4. QAT Model Accuracy Optimization
• 9. Embedded Application Development (Runtime) Manual
  • 9.1. Embedded Application Development
  • 9.2. BPU SDK API DOC
  • 9.3. Basic Sample User Guide
  • 9.4. AI Benchmark User Guide
  • 9.5. Tool Introduction
• 10. Advanced Contents
  • 10.1. Horizon Torch Samples
  • 10.2. DSP Development Documentation
  • 10.3. DSP Runtime API Manual
  • 10.4. DSP Sample Package Instruction
  • 10.5. Target Detection Full-process Example Document
• 11. Benchmark of Model Performance
  • 11.1. Descriptions
  • 11.2. Important Model Performance Data
  • 11.3. Model Full Performance Data
• 12. Appendix
  • 12.1. Toolchain Operator Support Constraint List
  • 12.2. Dataset Download
  • 12.3. Common Abbreviations
  • 12.4. Overview of Journey 5
  • 12.5. Development Board Kit and System Image Instructions