Problem $1[4$ points$]$

Consider the following code snippet to define a neural network in PyTorch:

$```$

claan MyNetwork$($nn$.$Module$)$:

def $\_\_$init$\_\_($self$,$ input$\_$size, hidden$\_$size, output$\_$size$)$:

A Bollerplate not shown

self.ll $=$ nn$.$Linear$($input$\_$size, hidden$\_$size$)$

self.nonlin $-$ nn$.$Tanh$()$

self$\mathrm{.}12=$ nn$.$Linear$($hidden$\_$size, output$\_$size$)$

self.log$\_$sm $=$ nn$.$LogSoftmax$($dim$=0)$

# Initialization, etc. not shown

def forward$($self$,$ x$)$:

return self.log$\_$sm$($self$,$nonlin$($self$\mathrm{.}12($self$,$nonlin$($self$\mathrm{.}11($x$)))))$

$```$

$1\mathrm{.}1[2$ points$]$ Let b be a batch size $>1.$ Which of the following tensor shapes can be input to this network without causing a code crash?

I. $[$inp$]$

II$.[$b$,$ inp$]$

III. $[$ing$,$ b$]$

A$.1$

B$.$ I and II

C$.$ I and III

D$.$ II and III

E$.$ I, II$,$ and III

$1\mathrm{.}2[2$ points$]$ What is "wrong" with the above network compared to standard neural network models?

A$.$ Tanh is not differentiable so backpropagation will not work

B$.$ There is a dimension mismatch which will lead to a crash

C$.$ Reusing the self.nonlin layer will cause a crash

D$.$ The presence of the nonlinearity before the softmax results in a non$-$standard model