论文标题
训练后的批评重新校准
Post-Training BatchNorm Recalibration
论文作者
论文摘要
我们重新访问Shomron and Weiser（2020）先前介绍的同时多线程（NB-SMT）。 NB-SMT通过偶尔将多个线程“挤压”到共享的乘数和积累（MAC）单元中来交易精度以换取性能。但是，在共享MAC单元中容纳多个线程的方法可能会为计算造成噪音，从而更改模型的内部统计数据。我们表明，考虑到NB-SMT的存在，可以通过对批准层的运行均值和运行方差统计的训练后重新校准来恢复实质性模型性能。
We revisit non-blocking simultaneous multithreading (NB-SMT) introduced previously by Shomron and Weiser (2020). NB-SMT trades accuracy for performance by occasionally "squeezing" more than one thread into a shared multiply-and-accumulate (MAC) unit. However, the method of accommodating more than one thread in a shared MAC unit may contribute noise to the computations, thereby changing the internal statistics of the model. We show that substantial model performance can be recouped by post-training recalibration of the batch normalization layers' running mean and running variance statistics, given the presence of NB-SMT.
