Section 4.6

Question

Section 4.6

dwf opened this issue 8 years ago · comments

David Warde-Farley commented 8 years ago

i' with a tilde is described but then i' is used in relationship 13.

David Warde-Farley · Answer 1 · Thu Mar 31 2016 11:24:26 GMT+0800 (China Standard Time)

Yeah, the equation for o should use i, not i'. And you probably want to lose the tildes too.

vdumoulin · Answer 2 · Fri Apr 01 2016 00:57:47 GMT+0800 (China Standard Time)

We use the tildes to differentiate the direct convolution view of the transposed convolution from the transposed convolution view itself.

I'm pretty sure the equation in relationship 13 is correct (you can verify that by solving the equation in relationship 6 for i, which corresponds to o' in the transposed convolution case), but we could maybe make that clearer by making use of relationship 2 on the direct convolution view:

\tilde{i}' = i' + (s - 1)(i' - 1),
        o' = (\tilde{i}' - k') + 2p' + 1
           = (i' + (s - 1)(i' - 1) - k) + 2(k - p - 1) + 1
           = s(i' - 1) + k - 2p

David Warde-Farley · Answer 3 · Fri Apr 01 2016 00:59:35 GMT+0800 (China Standard Time)

If i' contains a multiplication by s then you have an s^2 term in there, don't you?

vdumoulin · Answer 4 · Fri Apr 01 2016 01:02:49 GMT+0800 (China Standard Time)

I'm not sure I understand what you mean.

David Warde-Farley · Answer 5 · Fri Apr 01 2016 01:05:52 GMT+0800 (China Standard Time)

What exactly is i' in this situation? I see only tilde-i'.

On Thu, Mar 31, 2016 at 1:02 PM, vdumoulin notifications@github.com wrote:

I'm not sure I understand what you mean.

—
You are receiving this because you authored the thread.
Reply to this email directly or view it on GitHub
#2 (comment)

vdumoulin · Answer 6 · Fri Apr 01 2016 01:34:32 GMT+0800 (China Standard Time)

i' corresponds to the input size of the transposed convolution or, alternatively, the output size of the converse convolution.