(W-F+2P)/S+1 works out to be a decimal

If you have data with a height of 10, a width of 10, a number of channels of 3, and a lot size of 1, after nn.Conv2d(in_channels=3, out_channels=64, kernel_size=5, stride=2, padding=1), If the theoretical output has a height and width of 4.5, but the actual output has a height and width of 4, is it possible to show that nn.Conv2d is rounded down?

I’m unsure how you want to “show” the rounding, but the docs are giving you the output sizes using the floor function ⌊x⌋, which means you will round down to the next Integer.