#transposed-convolution

Question 1: Single choice

A "stride of 1/2" means the output has twice the height and width of the input. How is such a layer built in practice, and what is it used for?

As a stride-2 convolution followed by max-pooling, to downsample feature maps quickly1
As a $1 \times 1$ convolution, which doubles the resolution by doubling the channels2
As a transposed convolution, to upsample feature maps, for example in segmentation3
As a dilated convolution with dilation 2, which doubles the size of the output4