#fcn

An FCN upsamples its backbone's feature map with a transposed convolution. For a $480 \times 640$ input image, the ResNet-18 backbone and the $1 \times 1$ convolution give a $15 \times 20$ map with 21 class channels.

import torch
from torch import nn

conv_trans = nn.ConvTranspose2d(21, 21, kernel_size=64, padding=16, stride=32)
X = torch.rand(1, 21, 15, 20)
print(conv_trans(X).shape)

What is the last number printed, the output width?