文件名称：量化深度卷积网络的有效推理：白皮书
文件大小：820KB
文件格式：PDF
更新时间：2021-09-02 10:54:02
深度学习 量化
谷歌写的关于神经网络量化的白皮书，内容很详细，通俗易懂，值得一读。 1 Introduction 2 Quantizer Design 2.1 Uniform Affine Quantizer 2.2 Uniform symmetric quantizer 2.3 Stochastic quantizer 2.4 Modeling simulated quantization in the backward pass 2.5 Determining Quantizer parameters 2.6 Granularity of quantization 3 Quantized Inference: Performance and Accuracy 3.1 Post Training Quantization 3.1.1 Weight only quantization 3.1.2 Quantizing weights and activations 3.1.3 Experiments 3.2 Quantization Aware Training 3.2.1 Operation Transformations for Quantization 3.2.2 Batch Normalization 3.2.3 Experiments 3.2.4 Lower Precision Networks 4 Training best practices 5 Model Architecture Recommendations 6 Run-time measurements 7 Neural network accelerator recommendations 8 Conclusions and further work 9 Acknowledgements A Impact of Batch Normalization on Quantization