Realtek AmebaPro2 NN Quantization Discussion

By quantizing the model, it can decrease the size of model and speed up the inference. But it will decrease the accuracy about 1~3% as well.

Before starting quantization, please prepare a dataset for calibration by choosing images from your dataset randomly and copy them to a folder. Then create a text file, called dataset.txt to list down the path of the images. Kindly place the dataset.txt in the same directory as the model.

• Inputmeta File Modification
  • Crucial step before quantization
    • YOLOv4-tiny
    • YOLOv7-tiny for Pytorch version
    • MobileFaceNet
    • SCRFD
    • YAMNet
    • CNN Model for Image classification

Examples

• NN Example (Yolov7_tiny)
  • Using customized NN model
    • Setup Acuity toolkit on PC
    • Step for customized model conversion
• NN Example (Yolov9_tiny)
  • Using customized NN model
    • Setup Acuity toolkit on PC
    • Step for customized model conversion
    • Supported model quantization type on NN accelerator
  • SDK configuration for customized NN model
    • NN related file in SDK
    • Add customized model network binary to SDK
    • Create a model object can be used by VIPNN module
  • Post-process PC Development Tool
  • Hybrid Quantize
  • Troubleshooting