https://www.zhihu.com/question/362455124?sort=created
微軟的模型壓縮工具：distiller（重點(diǎn)）
https://github.com/NervanaSystems/distiller?(支持PyTorch厉斟，官方是在PyTorch1.3上測試的，在GitHub上搜PyTorch Pruning最多星)
微軟的AutoML工具：nni中也有模型壓縮的模塊擦秽，https://github.com/microsoft/nni/blob/master/examples/model_compress/QAT_torch_quantizer.py
https://nni.readthedocs.io/zh/latest/Compressor/Quantizer.html
https://github.com/microsoft/nni/blob/master/examples/model_compress/DoReFaQuantizer_torch_mnist.py
Pytorch自帶的量化工具（PyTorch>1.3）
https://zhuanlan.zhihu.com/p/81026071
https://github.com/pytorch/glow/blob/master/docs/Quantization.md
https://github.com/pytorch/QNNPACK
nni中的量化從例子看起來還蠻好用的（Pytorch的官方量化文檔看暈了，不適合剛?cè)胧至炕男“祝?/p>
https://github.com/microsoft/nni/blob/master/examples/model_compress/DoReFaQuantizer_torch_mnist.py
nni中有4種量化方式：
1.?Naive Quantizer缩搅，2.?QAT Quantizer触幼，3.?DoReFa Quantizer硼瓣，4.BNN Quantizer
其中置谦，1貌似是最low的，推理的時(shí)候32位變成8位霉祸，不想用，4慢宗，二進(jìn)制神經(jīng)網(wǎng)絡(luò)是啥？按位運(yùn)算挺好的镜沽，不知道部署時(shí)候是否有坑
2，Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference（谷歌2018）nni中說不支持批量歸一化折疊缅茉，不知道有沒有影響
3，DoReFa-Net: Training Low Bitwidth Convolutional Neural Networks with Low Bitwidth Gradients（Face++ 2018）例子看起來簡單https://arxiv.org/abs/1606.06160译打，https://blog.csdn.net/langzi453/article/details/88172080
nni中的量化只是模擬不是加速？https://github.com/microsoft/nni/issues/2332
https://github.com/microsoft/nni/blob/master/examples/model_compress/auto_pruners_torch.py
paper with code 的量化github排名：
https://paperswithcode.com/search?q_meta=&q=Quantizer
第一名：Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference（谷歌2018）
https://paperswithcode.com/paper/quantization-and-training-of-neural-networks? ? Tensorflow勸退
第二名：Training with Quantization Noise for Extreme Model Compression（2020）
https://paperswithcode.com/paper/training-with-quantization-noise-for-extreme?PyTorch實(shí)現(xiàn)
模型壓縮的benchmark:?https://paperswithcode.com/task/model-compression
模型壓縮的benchmark:https://paperswithcode.com/task/quantization
DoReFa還有這個(gè)人的：https://github.com/666DZY666/model-compression\
QAT還有這個(gè)人的：https://github.com/Xilinx/brevitas
