Tools OV INT8Model

[junxnone]

OpenVINO INT8 Model

• OpenVINO INT8 Model - 增加性能/减少模型大小
• POT - Post-training Optimization Tool
• API/CMD/Workbench 三种方式量化
• Quantization Methods
  • Default Quantization
    • 仅需要未注释的数据集
    • Accuracy Drop 不可控
  • Accuracy-aware Quantization
    • 需要标注的数据集
    • Accuracy Drop 可控

Default Quantization

• POT Python API/POT cmd-line

Reference

• Optimizing models post-training
• calibration_tool
• NNCF
• Intel® Neural Compressor
• Post-Training Quantization with TensorFlow Classification Model