12

Chapter12/Figures/figure-point-product-attention-model.tex

@@ -59,7 +59,7 @@
 {
 \node [rectangle,inner sep=0.2em,rounded corners=1pt,fill=green!10,drop shadow,draw=ugreen,minimum width=10em] [fit = (line1) (line2) (line3) (line4)] (box1) {};
 \node [rectangle,inner sep=0.1em,rounded corners=1pt,very thick,dotted,draw=ugreen] [fit = (Q1) (K1) (V1)] (box0) {};
-\draw [->,dotted,very thick,ugreen] ([yshift=-1.5em,xshift=1.2em]box1.east) -- ([yshift=-1.5em,xshift=0.1em]box1.east);
+\draw [->,dotted,very thick,ugreen] ([yshift=-1.5em,xshift=1.8em]box1.east) -- ([yshift=-1.5em,xshift=0.1em]box1.east);
 }
 {
 \node [rectangle,inner sep=0.2em,rounded corners=1pt,fill=blue!20!white,drop shadow,draw=blue] [fit = (line11) (line12) (line13)] (box2) {};

Chapter12/Figures/figure-process-of-5.tex

@@ -116,7 +116,11 @@
 % )
 \node(bra2) at ([xshift=0.2em,yshift=0]mid.east){)};
 %红色框
-\node[rectangle,minimum width=4.0em,minimum height=1.5em,draw=red](p222) at([xshift=0em,yshift=-1.0em]mid.north) {};
+\node[rectangle,minimum width=4.0em,minimum height=1.5em,draw=red,line width=1pt](p222) at([xshift=0em,yshift=-1.0em]mid.north) {};
+
+\node[rectangle,minimum width=4.0em,minimum height=1.5em,draw=ugreen,ultra thick,dotted,thick,font=\footnotesize](sub) at([xshift=-12em,yshift=1.0em]p222.west) {按行进行Softmax};
+\draw[->,dotted,very thick,draw=ugreen] (p222.west) .. controls +(north:0.5) and +(east:1) .. (sub.east);
+
 
 %%%% v
 \node(tbv3) at ([xshift=0.5em,yshift=0]bra2.east){

Chapter12/chapter12.tex

@@ -415,7 +415,7 @@
 
 \parinterval 在Transformer的训练过程中，由于引入了残差操作，将前面所有层的输出加到一起，如公式：
 \begin{eqnarray}
-x_{l+1} = x_l + \mathcal{F} (x_l)
+x_{l+1} = x_l + F (x_l)
 \label{eq:12-50}
 \end{eqnarray}