TPU-NNTC Development Reference Manual

Statement

• Statement

Basic Concept of NNToolChain

• Basic concept of NNToolChain

Overall Architecture of NNToolChain

• Overall architecture of NNToolChain

Installation Instructions of NNToolChain

• Installation instructions of NNToolChain

Model Compilation

• Model compilation
  • Use of BMNETC
  • Use of BMNETD
    • Common Problems
  • Use of BMNETM
  • Use of BMNETO
    • Common Problems
  • Use of BMNETP
    • Common Problems
  • Use of BMNETT
  • Use of BMPADDLE

Model Quantization

• Model quantization
  • Introduction to quantization-tools
    • Example program
  • Quantization steps
    • auto-cali quantization
    • Step-by-step quantization
    • Use of BMNETU
    • Example–YOLOv5s quantization
  • Quantization skills
    • Threshold calculation
    • Mixed execution
    • Network graph optimization
    • asym_threshold
    • fp16_opt
    • auto_cali quantitative optimization strategy
  • Common Exceptions
    • The path of lmdb dataset is not set correctly
    • There are unsupported layers
    • The whole network input is 0
  • Quantitative demo
    • Example 1: classify_demo
    • Example 2: create_lmdb_demo
    • Example 3: face_demo
    • Example 4: object_detection_python_demo
    • Example 5: caffemodel_to_fp32umodel_demo
    • Example 6: tf_to_fp32umodel_demo
    • Example 7: pt_to_fp32umodel_demo
    • Example 8: mx_to_fp32umodel_demo
    • Example 9: dn_to_fp32umodel_demo
    • Example 10: on_to_fp32umodel_demo
    • Example 11: pp_to_fp32umodel_demo
    • Example 12: auto_cali_demo
    • Example 13: yolov5s_demo
    • Example 14: view_demo
  • U-FrameWork python interface

Appendix

• Appendix 1: Support of Front-end Operators
  • Caffe support
  • TensorFlow support
  • Pytorch support
  • MxNet support
  • Darknet support
  • ONNX support
  • PADDLE support
• Appendix 2: Support of Quantization Operator
  • Supported layer