yafeng

Results 15 comments of yafeng

用这个 https://github.com/alibaba/MNN/tree/master/tools/MNNPythonOfflineQuant

1. 导出到onnx的时候,torch.onnx.export中设置 training=1即可保存BN 2. 不保存BN也可以用这个工具进行量化,只是精度可能没有保留BN的好

已知bug,如果等不了,可以自己修复,然后源码编译python包。具体的,请在 express/NeuralNetWorkOp.cpp中,第1492,1528,1578行,这三处之后添加下面4行代码: ` if (pads.size() == 2) { conv2D->common->padX = pads[0]; conv2D->common->padY = pads[1]; } else { conv2D->common->pads = std::move(pads); } ` 然后到pymnn/pip_package源码编译MNN python,命令: python build_deps.py pymnn/pip_package/MNN中新建version.py文件,里面写上__version__ = "test-version" python...

cannot reproduce with Python 3.8.13, change python version may solve this problem

https://github.com/alibaba/MNN/tree/master/tools/MNNPythonOfflineQuant

不太可能出现你说的的情况,离线量化之后没有这个scale?

这个scale如果能合并,在原始模型到MNN的转换过程中就会优化掉这个scale