auto qat is a tool, which automatically insert qdq nodes in torch nn.module, calibrate models, and finetune.
- lib
impl of auto quantize - __models
models for unit testing, include structure definition of networks
cd pytorch-quantization_v2.1.2
python3 setup.py build
# you can try install follow !!!
pip3 install loguru
pip3 install
pip3 install onnx-simplifier==0.4.10
cd auto_qat
python3 ./qat_resnet50.py
- auto qdq insert
- calib
- load pretrained weights
- finetune
- export quant onnx
- qat vs ptq benchmark
[1] https://github.com/NVIDIA/TensorRT/tree/main/tools/pytorch-quantization
[2] https://docs.nvidia.com/deeplearning/tensorrt/pytorch-quantization-toolkit/docs/userguide.html
[3] https://docs.nvidia.com/deeplearning/tensorrt/developer-guide/index.html#working-with-int8
[4] https://google.github.io/styleguide/pyguide.html
[5] https://github.com/zh-google-styleguide/zh-google-styleguide/tree/master/google-python-styleguide
[6] https://elinux.org/index.php?title=TensorRT/CommonErrorFix
[7] https://developer.nvidia.com/zh-cn/blog/tensorrt-measuring-performance-cn/
[8] https://developer.nvidia.com/blog/exploring-tensorrt-engines-with-trex/
auto qat is provided under the [Apache-2.0 license](LICENSE).