#bilinear-interpolation

Question 1: Select all that apply

d2l builds a fully convolutional network (FCN) for semantic segmentation of Pascal VOC (21 classes) from a pretrained ResNet-18. Which statements about this network are true? Select all that apply.

Its transposed convolution starts as bilinear upsampling, not with random weights1
Its input must be exactly $320 \times 480$ , the image size used in training2
It drops ResNet-18's global average pooling and fully connected layers3
A $1 \times 1$ convolution turns the 512 feature channels into 21, one per class4
Its output has a single channel holding each pixel's class index, so no argmax is needed5

#bilinear-interpolation

Timed test