def svm$\_$loss$($X$,$y$,$Weight,reg$\_$para,batch$\_$size,num$\_$classes$)$:

# Here you are required to calculate the hinge loss with L$2$ regularization, and the gradient dW for updating the weight

margin$\_$delta $=1$ # define the delta to determine the margin softness

loss $=0\mathrm{.}0$ #initialize the loss with $0\mathrm{.}0$ value

dW $=$ np$.$zeros$($Weight$.$shape$).$astype$(\text{'}$float$\text{'})$ # initialize the gradient as zero

# $\%\%\%\%\%\%\%\%\%\%\%\%\%\%$ implement your code below $(10$~$15$ lines$)\%\%\%\%\%\%\%\%\%\%\%\%\%\%$

scores $=$ np$.$dot$($X$,$Weight$)$ # predict the class score to generate a $($number of sample, number of class$)$ tensor

correct$\_$class$\_$scores $=$ scores$[$np$.$arange$($num$\_$train$),$ y$]$ # find the predict score of the correct class return a $($ number of sample, $)$ tensor

correct$\_$class$\_$scores $=$ np$.$reshape$($correct$\_$class$\_$scores,$($batch$\_$size,$1))$ # reshape the predct score into $($number of sample, $1)$ tensor for matrix propagation

margin $=$ scores $-$ correct$\_$class$\_$scores $+$ margin$\_$delta # calculate the margin

margins$[$np$.$arange$($num$\_$train$),$ y$]=$

margins$[$margins $<=0]=$

margins$[$margins $>0]=$

loss $+=$ #calculate loss with regularization part $1-2$ lines

row$\_$sum $=$ # sum the margin per data point $1$ by N

margins$[$np$.$arange$($num$\_$train$),$ y$]=$

dW $+=$

# $\%\%\%\%\%\%\%\%\%\%\%\%\%\%$ your code ends here $\%\%\%\%\%\%\%\%\%\%\%\%\%\%$

return loss, dW # return the loss and gradient