In the lecture notes, concerning the formula for the output format when applying valid padding, shouldn't the output format be
$$N_1-K_1+1$$
rather than
$$N_1-2(K_1-1)$$
? (we only consider the first dimension here). The provided example in Fig 2 has N_1 = 4 and K_1=2. The resulting output format has indeed dimension 4 - 2 + 1 = 3 rather than 4 - 2*(2-1) = 2.
The formula is correct but I think you are mixing up the size of the kernel and the definition of \(K_1\). The kernel value is zero for \( |l|\geq K_1 \). Therefore the size of the kernel is \(2 K_1 -1\) (this notation does not allow to have kernel with even size and I agree that the illustration is slightly misleading).
You want to find the last positions so that the whole filter lies inside the original data so you have to stop \(K_1-1\) pixels before the last one both at right and at left. Therefore the correct formula is \( N_1-2(K_1-1) \). It is easier to convince yourself doing a drawing.
Valid Padding output format
In the lecture notes, concerning the formula for the output format when applying valid padding, shouldn't the output format be
$$N_1-K_1+1$$
rather than
$$N_1-2(K_1-1)$$
? (we only consider the first dimension here). The provided example in Fig 2 has N_1 = 4 and K_1=2. The resulting output format has indeed dimension 4 - 2 + 1 = 3 rather than 4 - 2*(2-1) = 2.
Hi,
The formula is correct but I think you are mixing up the size of the kernel and the definition of \(K_1\). The kernel value is zero for \( |l|\geq K_1 \). Therefore the size of the kernel is \(2 K_1 -1\) (this notation does not allow to have kernel with even size and I agree that the illustration is slightly misleading).
You want to find the last positions so that the whole filter lies inside the original data so you have to stop \(K_1-1\) pixels before the last one both at right and at left. Therefore the correct formula is \( N_1-2(K_1-1) \). It is easier to convince yourself doing a drawing.
If it is still not clear please come to the Q&A.
Best,
Nicolas
Ok I understand now, I was indeed confusing K with the kernel size. Thanks !
Add comment