Algorithm 1 The forward-propagation algorithm of quantization sub-layer for input features
Input:	Values of input features over a mini-batch: $\mathrm{B}=\left\{{\mathrm{x}}_1,{\mathrm{x}}_2,\dots ,{\mathrm{x}}_m\right\}$;
	Boundary values to be estimated: $max\_moving$,$min\_moving$
Output:	$\left\{q_{idx}=\mathrm{Quantization}\left(x_{idx}\right)\right\}$
Step1.	$max\_x=\frac{1}{b}{\displaystyle \sum _{idx=0}^{idx=b-1}\max \left(x_{idx}\right)}$,$min\_x=\frac{1}{b}{\displaystyle \sum _{idx=0}^{idx=b-1}\min \left(x_{idx}\right)}$ //Real Boundary values
Step2.	$max\_moving=\left(1-\lambda \right)\cdot max\_x+\lambda \cdot max\_moving$,
	$max\_moving=\left(1-\lambda \right)\cdot max\_x+\lambda \cdot max\_moving$ //Moving Boundary values
Step3.	$S_x=\frac{\max \left(\left|max\_x\right|,\left|min\_x\right|\right)}{2^{N-1}-1}$ //Scaling factor
Step4.	$q_{idx}=\frac{x_{idx}}{\max \left(S_x,\epsilon \right)}$ //Quantization