README.md

ONNX Post-training quantization

How to Run

1. train base model with imagenet100 dataset

   • Train Base Model (resnet18)

2. export input onnx

   • Export input onnx

3. ONNX Post-training quantization and export moq onnx

   python onnx_export_moq.py
   

   • no train
   • calibration

4. generate tensorrt model

   python onnx2trt.py
   

• int8 onnx ptq (Explicit)
  Gpu Mem: 124M
  [TRT_E] Test Top-1 Accuracy: 84.50%
  [TRT_E] Test Top-5 Accuracy: 97.00%
  [TRT_E] 10000 iterations time: 5.2702 [sec]
  [TRT_E] Average FPS: 1897.46 [fps]
  [TRT_E] Average inference time: 0.53 [msec]

Reference

• TensorRT-Model-Optimizer