README.md

Quantization-aware training (QAT)

How to Run

1. train base model with imagenet100 dataset

   • Train Base Model (resnet18)

2. Quantization-aware training (QAT) and export qat onnx

   python onnx_export_qat.py
   

   • fine tuning (We recommend QAT for 10% of the original training epochs)

3. generate tensorrt model

   python onnx2trt.py
   

• int8 qat (Explicit)
  Gpu Mem: 138M
  [TRT_E] Test Top-1 Accuracy: 84.42%
  [TRT_E] Test Top-5 Accuracy: 97.10%
  [TRT_E] 10000 iterations time: 6.3580 [sec]
  [TRT_E] Average FPS: 1572.81 [fps]
  [TRT_E] Average inference time: 0.64 [msec]

Reference

• TensorRT-Model-Optimizer