ResNet output dimensions of initial convolution don’t yield in an integer

Question

I am trying to understand the ResNet dimensions, but got stuck at the first layer. We are passing a [224x224x3] image into 64 filters with kernel size 7x7 and stride=2. According to the ResNet source code from pytorch we are also using zero padding of size 3. The output size should be 112, but I get a output size of 112.5. To get an output size of 112 we need padding of 2.5.. See:

I do not understand how the output of 112 is created. Is the padding adjusted by pytorch automatically to match floor (output)?

Answer 1

As you can see in the pytorch documentation for torch.nn.Conv2d, the output size of a 2d convolutional layer can be calculated that is very similar to the formula you show:

$$ H_{out} / W_{out} = \lfloor \frac{H_{in} + 2 * padding - dilation * (kernel\_size - 1) - 1}{stride} + 1 \rfloor $$

So after performing the calculations pytorch indeed floors the output to get the largest integer smaller than or equal to the calculated output since half pixels do not exist.