6. Post-training Quantization (PTQ)

D-Robotics provides you with the PTQ scheme. The following sections will guide you on how to use Post Training Quantization (PTQ) .

PTQ Conversion Principle And Process: This section introduces you to the overall process of PTQ model conversion and details of the whole process.

PTQ Conversion Tools Guide: This section provides you with a detailed introduction to the PTQ toolkit provided by algorithm toolchain.

PTQ Conversion Steps: This section provides instructions on how to use it from model preparation, model checking, prepare calibration data, model quantization and compilation, performance analysis, accuracy analysis, and so on.

PTQ Conversion Samples Guide: This section introduces conversion sample package of the horizon_model_convert_sample model and its usage instructions. Provide a quick sample of converting a floating-point model to a fixed-point model using the floating-point model conversion toolchain, including a single inference and accuracy verification sample.

Custom OP Development: If you are using a model with operators that are not supported by the D-Robotics toolchain, refer to this section for custom operator development.

FAQ And Common Failure Resolutions: This section provides you with answers to some common questions about the PTQ conversion process as well as generalized suggestions for solving common trouble-shooting phenomena.

Appendix: This section introduces the descriptions and analysis of norm_type related parameters and related calculation formulas, as well as the concept of each transformer used in image scaling and cropping, parameter descriptions and examples, and general suggestions for solving common abnormalities and failures.

• 6.1. PTQ Conversion Principle and Process
  • 6.1.1. Overview
  • 6.1.2. PTQ Conversion Process
  • 6.1.3. Model Conversion Process Detailed Explanation
• 6.2. PTQ Conversion Tools Guide
  • 6.2.1. PTQ Conversion Tool Reading Overview
  • 6.2.2. The hb_mapper Tool
  • 6.2.3. The hb_perf Tool
  • 6.2.4. The vec_diff Tool
  • 6.2.5. The hb_model_info Tool
  • 6.2.6. The hb_pack Tool
  • 6.2.7. The hb_verifier Tool
  • 6.2.8. The hb_model_modifier Tool
  • 6.2.9. The hb_custom_op Tool
  • 6.2.10. The hb_eval_preprocess Tool
  • 6.2.11. The HB_ONNXRuntime Inference Library
  • 6.2.12. Accuracy Debug Tool
• 6.3. PTQ Conversion Steps
  • 6.3.1. Floating-point Model Preparation
  • 6.3.2. Model Checking
  • 6.3.3. Prepare Calibration Data
  • 6.3.4. Model Quantization and Compilation
  • 6.3.5. Model Performance Analysis
  • 6.3.6. Model Accuracy Analysis
• 6.4. PTQ Conversion Samples Guide
  • 6.4.1. Release Introduction
  • 6.4.2. Common Algorithm Model Samples
  • 6.4.3. Other Algorithm Model Samples
• 6.5. Custom OP Development
  • 6.5.1. General Descriptions
  • 6.5.2. Custom OP Included Model Conversion
  • 6.5.3. Run Custom OP on Dev Board
• 6.6. FAQ and Common Failure Resolutions
  • 6.6.1. Post-training Quantization (PTQ) FAQ
  • 6.6.2. Common Failure Resolutions
• 6.7. Appendix
  • 6.7.1. Configuration of norm_type
  • 6.7.2. Image Processing Transformer