4. Floating-point Model Convert to Fixed-point Model Guide

D-Robotics provides you with two methods to convert floating-point model to fixed-point model, which are PTQ model quantization scheme and QAT model quantization scheme. This section brings you the description of Post-training Quantization (PTQ) and Quantized Awareness Training (QAT) .

• 4.1. Post-training Quantization (PTQ)
  • 4.1.1. PTQ Principle and Steps
    • 4.1.1.1. PTQ Conversion Process and Steps
    • 4.1.1.2. Operator Constraints
    • 4.1.1.3. Floating-point Model Preparation
    • 4.1.1.4. Check the Model
      • 4.1.1.4.1. Use the hb_mapper checker Command to Check Your Model
      • 4.1.1.4.2. Exception Handling
      • 4.1.1.4.3. Interpret Model Check Results
      • 4.1.1.4.4. A Guide to Optimize the Check Results
    • 4.1.1.5. Prepare Calibration Data
    • 4.1.1.6. Model Quantization and Compilation
      • 4.1.1.6.1. Convert the Model Using the hb_mapper makertbin Tool
        • 4.1.1.6.1.1. Configuration File Template
        • 4.1.1.6.1.2. Specific Parameter Information
        • 4.1.1.6.1.3. param_value Configuration
      • 4.1.1.6.2. Model Conversion Interpretation
      • 4.1.1.6.3. Interpret Conversion Results
      • 4.1.1.6.4. Interpret Conversion Output
    • 4.1.1.7. Model Performance Analysis and Optimization
      • 4.1.1.7.1. Use the hb_perf Tool to Evaluate Model Performance
      • 4.1.1.7.2. Evaluate Actual Model Performance on Dev Board
      • 4.1.1.7.3. Model Performance Optimization
    • 4.1.1.8. Model Accuracy Analysis and Optimization
      • 4.1.1.8.1. Model Accuracy Analysis
      • 4.1.1.8.2. Accuracy Optimization
      • 4.1.1.8.3. Further Improve Model Accuracy Using the QAT Solution
        • 4.1.1.8.3.1. About Quantization
        • 4.1.1.8.3.2. About Model Conversion
        • 4.1.1.8.3.3. About Model Quantization & Compilation Workflow
        • 4.1.1.8.3.4. QAT Model Quantization & Compilation Introduction
    • 4.1.1.9. Other Dev Tools (Optional)
      • 4.1.1.9.1. Pack Multiple Bin Models into One File
      • 4.1.1.9.2. View Model Information
      • 4.1.1.9.3. Modify the Nodes of Bin Models
    • 4.1.1.10. Post-training Quantization (PTQ) FAQ
      • 4.1.1.10.1. How to understand the two forms of BPU acceleration and CPU computation mentioned in the operator constraint?
      • 4.1.1.10.2. How does model segmentation affect the performance?
      • 4.1.1.10.3. Why some OPs supported by BPU at the tail part of the model running on CPU?
      • 4.1.1.10.4. Does it support asymmetric quantization?
      • 4.1.1.10.5. How to understand D-Robotics’ default calibration method?
      • 4.1.1.10.6. How to understand D-Robotics’ mix calibration method?
      • 4.1.1.10.7. How to understand the compiler optimization level parameters in yaml files?
      • 4.1.1.10.8. Why does the input size derived from the nv12 model hb_perf not match the prediction library?
      • 4.1.1.10.9. Is the data layout of the inputs to the quantized model and the on-board bin model necessarily the same?
      • 4.1.1.10.10. How to compile to get a multi-batch model?
      • 4.1.1.10.11. Is it normal for the order of model inputs to change during the conversion of a multi-input model?
    • 4.1.1.11. Custom OP Development
      • 4.1.1.11.1. General Descriptions
      • 4.1.1.11.2. Custom OP Included Model Conversion
        • 4.1.1.11.2.1. Modify Model File
          • 4.1.1.11.2.1.1. Caffe Model
          • 4.1.1.11.2.1.2. ONNX Model
        • 4.1.1.11.2.2. OP Implementation
      • 4.1.1.11.3. Run Custom OP on Dev Board
        • 4.1.1.11.3.1. A C++ Template of Custom OP
        • 4.1.1.11.3.2. Custom OP Registration
  • 4.1.2. PTQ Tools Guide
    • 4.1.2.1. Model Conversion Process
      • 4.1.2.1.1. Model Checking (hb_mapper checker)
      • 4.1.2.1.2. Calibration Image Preparation
      • 4.1.2.1.3. Model Conversion (hb_mapper makertbin)
      • 4.1.2.1.4. Single Image Inference
      • 4.1.2.1.5. Model Accuracy Evaluations
      • 4.1.2.1.6. Model Performance Evaluations
      • 4.1.2.1.7. [Reference] Supported Calibration Methods
      • 4.1.2.1.8. [Reference] OP List
    • 4.1.2.2. The hb_mapper Tools
      • 4.1.2.2.1. Model Checking Tool (hb_mapper checker)
      • 4.1.2.2.2. Model Compiling Tool (hb_mapper makertbin)
      • 4.1.2.2.3. Inference Tool (hb_mapper infer)
      • 4.1.2.2.4. Key Configuration Parameter (calibration_parameters.preprocess_on)
      • 4.1.2.2.5. Configuration File Reference Example
        • 4.1.2.2.5.1. Reference and Descriptions
        • 4.1.2.2.5.2. About input_type_rt/ input_type_train
    • 4.1.2.3. The hb_perf Tool
      • 4.1.2.3.1. How to Use
      • 4.1.2.3.2. Parameters
      • 4.1.2.3.3. Output Contents
    • 4.1.2.4. The vec_diff Tool
      • 4.1.2.4.1. How to Use
      • 4.1.2.4.2. Parameters
      • 4.1.2.4.3. Output Contents
    • 4.1.2.5. The hb_model_info Tool
      • 4.1.2.5.1. How to Use
      • 4.1.2.5.2. Parameters
      • 4.1.2.5.3. Output Contents
    • 4.1.2.6. The hb_pack Tool
      • 4.1.2.6.1. How to Use
      • 4.1.2.6.2. Parameters
      • 4.1.2.6.3. Output Contents
    • 4.1.2.7. The hb_verifier Tool
      • 4.1.2.7.1. How to Use
      • 4.1.2.7.2. Parameters
      • 4.1.2.7.3. Examples of Reference Usage
      • 4.1.2.7.4. Description of the Output Content
    • 4.1.2.8. The hb_model_modifier Tool
      • 4.1.2.8.1. How to Use
      • 4.1.2.8.2. Parameters
      • 4.1.2.8.3. Output Contents
    • 4.1.2.9. The hb_custom_op Tool
      • 4.1.2.9.1. How to Use
      • 4.1.2.9.2. Parameters
      • 4.1.2.9.3. Output Contents
    • 4.1.2.10. The hb_eval_preprocess Tool
      • 4.1.2.10.1. How to Use
      • 4.1.2.10.2. Parameters
      • 4.1.2.10.3. Output Contents
    • 4.1.2.11. The HB_ONNXRuntime Inference Library
      • 4.1.2.11.1. How to Use
      • 4.1.2.11.2. Parameters
    • 4.1.2.12. Accuracy Debug Tool
      • 4.1.2.12.1. Quickstart
        • 4.1.2.12.1.1. Calibration Models and Data Storage
        • 4.1.2.12.1.2. Accuracy Debug Module Import and Usage
        • 4.1.2.12.1.3. Analysis Result Display
      • 4.1.2.12.2. Function Description
        • 4.1.2.12.2.1. get_sensitivity_of_nodes
        • 4.1.2.12.2.2. plot_acc_error
        • 4.1.2.12.2.3. plot_distribution
        • 4.1.2.12.2.4. get_channelwise_data_distribution
        • 4.1.2.12.2.5. sensitivity_analysis
        • 4.1.2.12.2.6. runall
  • 4.1.3. PTQ Model Conversion Samples Guide
    • 4.1.3.1. Release Introduction
      • 4.1.3.1.1. Model Quantification Example Package: horizon_model_convert_sample
      • 4.1.3.1.2. Model Release: model_zoo
    • 4.1.3.2. Common Algorithm Model Samples
      • 4.1.3.2.1. Where to Find
      • 4.1.3.2.2. Prepare Dataset
        • 4.1.3.2.2.1. Dataset Download Address
        • 4.1.3.2.2.2. Dataset Reference Structure
          • 4.1.3.2.2.2.1. ImageNet
          • 4.1.3.2.2.2.2. COCO
          • 4.1.3.2.2.2.3. VOC
          • 4.1.3.2.2.2.4. Cityscapes
          • 4.1.3.2.2.2.5. CIFAR-10
      • 4.1.3.2.3. Prepare Models
        • 4.1.3.2.3.1. Fcos_efficientnetb0
        • 4.1.3.2.3.2. Fcos_efficientnetb1
        • 4.1.3.2.3.3. Fcos_efficientnetb2
        • 4.1.3.2.3.4. MobileNetv1/v2
        • 4.1.3.2.3.5. GoogleNet
        • 4.1.3.2.3.6. ResNet18
        • 4.1.3.2.3.7. Mobilenet_onnx
        • 4.1.3.2.3.8. EfficientNet_Lite0/1/2/3/4
        • 4.1.3.2.3.9. YOLOv2_Darknet19
        • 4.1.3.2.3.10. YOLOv3_Darknet53
        • 4.1.3.2.3.11. YOLOv5s
        • 4.1.3.2.3.12. SSD_MobileNetv1
        • 4.1.3.2.3.13. EfficientDetd0
        • 4.1.3.2.3.14. CenterNet_Resnet50
        • 4.1.3.2.3.15. Fcos_efficientnetb0
        • 4.1.3.2.3.16. UNet_mobilenet
        • 4.1.3.2.3.17. DeeplabV3plus_efficientnetb0
        • 4.1.3.2.3.18. Fastscnn_efficientnetb0
      • 4.1.3.2.4. A Demonstration of the Algorithm Model Samples
        • 4.1.3.2.4.1. Docker Container Prep
        • 4.1.3.2.4.2. Obtain the Original Model and Calibration Dataset
        • 4.1.3.2.4.3. Check If the Model is Executable
        • 4.1.3.2.4.4. Prepare Calibration Dataset
        • 4.1.3.2.4.5. Building Heterogeneous Models
        • 4.1.3.2.4.6. Single Image Inference
        • 4.1.3.2.4.7. Accuracy Evaluation
      • 4.1.3.2.5. FAQ
        • 4.1.3.2.5.1. How to match the inference results of the ONNX original floating-point model and the hb_mapper makertbin tool generated *_original_float_model.onnx model?
        • 4.1.3.2.5.2. Why is the reproduced accuracy slightly different from the indicators in the documentation?
        • 4.1.3.2.5.3. Why is the fixed-point model accuracy not aligned with the on-board accuracy of the bin file in the ai_benchmark example?
    • 4.1.3.3. Other Algorithm Model Samples
      • 4.1.3.3.1. How to Use
      • 4.1.3.3.2. Miscellaneous Samples
        • 4.1.3.3.2.1. lenet_gray
          • 4.1.3.3.2.1.1. Directory
          • 4.1.3.3.2.1.2. Description
        • 4.1.3.3.2.2. resnet50_feature
          • 4.1.3.3.2.2.1. Directory
          • 4.1.3.3.2.2.2. Description
        • 4.1.3.3.2.3. vector-diff
          • 4.1.3.3.2.3.1. Directory
          • 4.1.3.3.2.3.2. Output List
        • 4.1.3.3.2.4. multi_input_example
          • 4.1.3.3.2.4.1. Directory
          • 4.1.3.3.2.4.2. Description
        • 4.1.3.3.2.5. model_verifier
          • 4.1.3.3.2.5.1. Directory
          • 4.1.3.3.2.5.2. Description
        • 4.1.3.3.2.6. model_info
          • 4.1.3.3.2.6.1. Directory
          • 4.1.3.3.2.6.2. Description
        • 4.1.3.3.2.7. mobilenet_bgr
          • 4.1.3.3.2.7.1. Directory
          • 4.1.3.3.2.7.2. Description
        • 4.1.3.3.2.8. mobilenet_yuv444
          • 4.1.3.3.2.8.1. Directory
          • 4.1.3.3.2.8.2. Description
      • 4.1.3.3.3. User Custom OP Sample
        • 4.1.3.3.3.1. Directory
        • 4.1.3.3.3.2. Description
  • 4.1.4. Appendix
    • 4.1.4.1. norm_type Configuration Description
      • 4.1.4.1.1. Explanation of Parameters
      • 4.1.4.1.2. Calculation Formula and Example Description
        • 4.1.4.1.2.1. Calculation Formula for Normalization of Data During Model Training
        • 4.1.4.1.2.2. Calculation Formula for Model Inference
    • 4.1.4.2. Image Processing Transformer Description
      • 4.1.4.2.1. AddTransformer
      • 4.1.4.2.2. MeanTransformer
      • 4.1.4.2.3. ScaleTransformer
      • 4.1.4.2.4. NormalizeTransformer
      • 4.1.4.2.5. TransposeTransformer
      • 4.1.4.2.6. HWC2CHWTransformer
      • 4.1.4.2.7. CHW2HWCTransformer
      • 4.1.4.2.8. CenterCropTransformer
      • 4.1.4.2.9. PILCenterCropTransformer
      • 4.1.4.2.10. LongSideCropTransformer
      • 4.1.4.2.11. PadResizeTransformer
      • 4.1.4.2.12. ResizeTransformer
      • 4.1.4.2.13. PILResizeTransformer
      • 4.1.4.2.14. ShortLongResizeTransformer
      • 4.1.4.2.15. PadTransformer
      • 4.1.4.2.16. ShortSideResizeTransformer
      • 4.1.4.2.17. PaddedCenterCropTransformer
      • 4.1.4.2.18. BGR2RGBTransformer
      • 4.1.4.2.19. RGB2BGRTransformer
      • 4.1.4.2.20. RGB2GRAYTransformer
      • 4.1.4.2.21. BGR2GRAYTransformer
      • 4.1.4.2.22. RGB2GRAY_128Transformer
      • 4.1.4.2.23. RGB2YUV444Transformer
      • 4.1.4.2.24. BGR2YUV444Transformer
      • 4.1.4.2.25. BGR2YUV444_128Transformer
      • 4.1.4.2.26. RGB2YUV444_128Transformer
      • 4.1.4.2.27. BGR2YUVBT601VIDEOTransformer
      • 4.1.4.2.28. RGB2YUVBT601VIDEOTransformer
      • 4.1.4.2.29. YUVTransformer
      • 4.1.4.2.30. ReduceChannelTransformer
      • 4.1.4.2.31. BGR2NV12Transformer
      • 4.1.4.2.32. RGB2NV12Transformer
      • 4.1.4.2.33. NV12ToYUV444Transformer
      • 4.1.4.2.34. WarpAffineTransformer
      • 4.1.4.2.35. F32ToS8Transformer
      • 4.1.4.2.36. F32ToU8Transformer
    • 4.1.4.3. Common Failure Resolutions
      • 4.1.4.3.1. Common Failure - hb_mapper checker
      • 4.1.4.3.2. Common Failure - hb_mapper makertbin
      • 4.1.4.3.3. Common Failure - hb_model_modifier
      • 4.1.4.3.4. Common Failure - hb_model_verifier
      • 4.1.4.3.5. Common Failure - hb_onnxruntime
      • 4.1.4.3.6. Common Failure - libDNN
• 4.2. Quantized Awareness Training (QAT)
  • 4.2.1. Environmental Dependence
  • 4.2.2. Quick Start
    • 4.2.2.1. Building A Floating Point Model
    • 4.2.2.2. Pre-train A Floating Point Model
    • 4.2.2.3. Set BPU Architectures
    • 4.2.2.4. Operator Fusion
    • 4.2.2.5. Convert A Floating Point Model To A Quantitative Model
    • 4.2.2.6. Quantization Aware Training (QAT)
    • 4.2.2.7. Converting a Quantized Model to a Fixed-point Model
    • 4.2.2.8. Check and Compile a Fixed-point Prediction Model
  • 4.2.3. TUTORIAL
    • 4.2.3.1. Floating Point Model Preparation
      • 4.2.3.1.1. Operator Substitution
      • 4.2.3.1.2. Insert Quantization and Dequantization Nodes
      • 4.2.3.1.3. Set up Quantization Parameters
      • 4.2.3.1.4. Example of Custom Floating Point Model
    • 4.2.3.2. Operator Fusion
      • 4.2.3.2.1. Why Do Operator Fusion
      • 4.2.3.2.2. Speed up Computation
      • 4.2.3.2.3. Improve Quantization Accuracy
      • 4.2.3.2.4. Fusion Operator
      • 4.2.3.2.5. Purpose of BN (Batch Normalization) Absorption
      • 4.2.3.2.6. BN Absorption Method
      • 4.2.3.2.7. Operator Fusion Example
    • 4.2.3.3. Set Up Different BPU Architectures
    • 4.2.3.4. Heterogeneous Model QAT
      • 4.2.3.4.1. Differences Between Heterogeneous Models and Non-heterogeneous Models
      • 4.2.3.4.2. Description of Main Interface Parameters
        • 4.2.3.4.2.1. horizon_plugin_pytorch.quantization.prepare_qat_fx
        • 4.2.3.4.2.2. horizon_plugin_pytorch.quantization.prepare_calibraiton_fx
        • 4.2.3.4.2.3. horizon_plugin_pytorch.quantization.convert_fx
        • 4.2.3.4.2.4. horizon_plugin_pytorch.utils.onnx_helper.export_to_onnx
      • 4.2.3.4.3. Example
    • 4.2.3.5. FX Based Quantization
      • 4.2.3.5.1. Restriction
    • 4.2.3.6. Building Quantization-friendly Floating-point Models
      • 4.2.3.6.1. Background
      • 4.2.3.6.2. Phenomena Not Suitable for Quantization
        • 4.2.3.6.2.1. Extensive Use of OPs Not Suitable for Quantization
        • 4.2.3.6.2.2. Asymmetric Input or High Resolution Requirements
          • 4.2.3.6.2.2.1. Solution
        • 4.2.3.6.2.3. Large Feature Map in Model
          • 4.2.3.6.2.3.1. Solution
        • 4.2.3.6.2.4. Large Range of Model Weight
          • 4.2.3.6.2.4.1. Solution
      • 4.2.3.6.3. Tips
        • 4.2.3.6.3.1. Not Suitable for Quantification Does Not Mean That It Cannot Be Quantified
        • 4.2.3.6.3.2. int16 Quantization
    • 4.2.3.7. QAT Experience Summary
      • 4.2.3.7.1. Hyperparameter
      • 4.2.3.7.2. Quantization Parameters
    • 4.2.3.8. Model Precision Debug Tool
      • 4.2.3.8.1. Similarity
        • 4.2.3.8.1.1. Example
      • 4.2.3.8.2. Visualization
        • 4.2.3.8.2.1. Model Visualization
        • 4.2.3.8.2.2. PT File Visualization
      • 4.2.3.8.3. Statistics
        • 4.2.3.8.3.1. Example
      • 4.2.3.8.4. Step Quantization
        • 4.2.3.8.4.1. Example
      • 4.2.3.8.5. Shared OP Check
        • 4.2.3.8.5.1. Example
      • 4.2.3.8.6. Fuse Check
        • 4.2.3.8.6.1. Example
      • 4.2.3.8.7. Single-operator Conversion Precision Debugging
        • 4.2.3.8.7.1. Example
      • 4.2.3.8.8. Quantitative Configuration Check
        • 4.2.3.8.8.1. Example
      • 4.2.3.8.9. Model Weight Comparison
        • 4.2.3.8.9.1. Example
      • 4.2.3.8.10. Deploy Device Check of Hybrid Models
        • 4.2.3.8.10.1. Example
      • 4.2.3.8.11. Integrated Interface
        • 4.2.3.8.11.1. Example
      • 4.2.3.8.12. GPU Memory Profiler
        • 4.2.3.8.12.1. Example
    • 4.2.3.9. Ideas of Quantization Precision Debug
      • 4.2.3.9.1. BaseLine
        • 4.2.3.9.1.1. Quantization Precision Standard
      • 4.2.3.9.2. Flow Chart
      • 4.2.3.9.3. NAN
      • 4.2.3.9.4. Low Calibration Precision
      • 4.2.3.9.5. Low QAT Precision
      • 4.2.3.9.6. Low Quantized Precision
  • 4.2.4. API REFERRENCE
    • 4.2.4.1. Quantization API
    • 4.2.4.2. D-Robotics Operator API
    • 4.2.4.3. Supported Operators
      • 4.2.4.3.1. Supported Torch Operator
      • 4.2.4.3.2. Supported Torchvision Operator
  • 4.2.5. NOTE
    • 4.2.5.1. Similarities and Differences in Using the QAT Interface Compared to the Community
      • 4.2.5.1.1. Operator Fusion Function
      • 4.2.5.1.2. QConfig Parameter Settings
      • 4.2.5.1.3. Model Conversion between Different Stages
  • 4.2.6. Calibration
    • 4.2.6.1. Calibration (Experimental Support)
      • 4.2.6.1.1. Calibration Method
      • 4.2.6.1.2. How to Use Calibration in Plugin
      • 4.2.6.1.3. Limitations of Calibration in Plugin
    • 4.2.6.2. Calibration v2 (Experimental Support)
      • 4.2.6.2.1. Usage
      • 4.2.6.2.2. Limitations of Calibration in Plugin