I disagree with the answer to the last question of the week 4 quiz. May I ask teaching staff to double check it?
According to the lecture, the number of channels in the input and the kernel should match. Here we have an input of 32x32x32x16 (where 16 is the number of channels), whereas the filter has the dimension of 3x3x3. In order to be able to perform this convolution, we need a filter of size 3x3x3x16, right?
Thanks for your notice.
You are right; this is implicitly assumed in the question. For clarity, it may be better to make it explicit.
the problem is there is an option for invalid operation
Yes, I’ll see if I can get this to be changed by people working on the backend. Thanks!
No Chinese language video, it’s not so good.
hi, I have another question:why they should match ?
Just like the kernel multiply the input image array, why we cannot implement through the channel?
I am not sure I understand your question correctly, but I’ll try to give an answer.
The 3x3x3 filter is in fact run over the 16 channels. But the 3x3x3 filter has to be compatible with the other dimensions of the input volume. (if e.g. the input volume had been 32x32x2x16 the dimensions would not have matched up).