Timeline for What are deconvolutional layers?
Current License: CC BY-SA 4.0
21 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| May 18, 2020 at 17:19 | history | edited | David Dao | CC BY-SA 4.0 |
Add a clearer way to attribute the great visualisations
|
| Aug 8, 2019 at 6:59 | comment | added | Ben | And how did you make the animation? | |
| Jul 7, 2019 at 20:37 | comment | added | meduz | Is the animation yours or could you give the right credit? | |
| Jul 7, 2019 at 20:36 | comment | added | meduz | Could you explain more than visually ? | |
| Apr 30, 2018 at 13:37 | comment | added | Tahlil | In the second image there is 1 zero padded pixel between each input pixel vertically and horizontally. So performing convolution with stride 1 here actually works like stride 2 as now a zero padded pixel sits between each input pixel. @Curious | |
| Dec 23, 2017 at 22:46 | comment | added | IssamLaradji | this is still very confusing. I don't see how the stride, padding, and transpose are affecting the operation ... The first gif has zero padding but you said there is no padding. The second gif uses a stride of 1, and you said the stride is two. | |
| Mar 30, 2017 at 6:54 | comment | added | stillanoob | @Demonedge Shown above is a transposed convolution. 'stride two' means stride in the corresponding original convolution is two. This is precisely why you have 1 (=2-1, 2 being the original stride) layer of zeros in between rows and columns. Transposed convolution is generally used in backward pass. It is called transposed because of the analogy with fully connected layer where you multiply with the transpose of the weight matrix during a backward pass. | |
| Mar 1, 2017 at 5:51 | comment | added | Bill Yan | Why do you call it "transpose"? What does transpose mean? | |
| Jan 15, 2017 at 22:05 | comment | added | Andrei | I'm sorry, but I think your answer is wrong regarding the first animation. As seen here (deeplearning.net/software/theano_versions/dev/tutorial/…) you are showing a convolution and not a transposed convolution. It will however have the same result as a transposed convolution with 0 padding going from a 2x2 space to a 4x4 space. That's why transposed convolutions are more efficient (no padding required). | |
| Dec 22, 2016 at 0:26 | comment | added | RockTheStar | @DavidDao I think to be more precise, I think transpose convolution = add padding and then do convolution with flipped (vertically & horizontally) filters (from original conv layer) | |
| Dec 20, 2016 at 23:36 | comment | added | RockTheStar | @MartinThoma, did the filter that we used in tranpose convolution flipped or transpose? Or in the tranpose convolution we just add padding and do convolution again? | |
| Dec 4, 2016 at 6:21 | review | Suggested edits | |||
| Dec 5, 2016 at 19:57 | |||||
| Nov 22, 2016 at 21:29 | comment | added | Martin Thoma | Could you link the arxiv paper A guide to convolution arithmetic for deep learning, please? | |
| Aug 17, 2016 at 14:54 | comment | added | Mikael Rousson | Personally, I prefer to see it as a convolution with fractional stride. In the cited paper, there is no real deconvolution in the sense of "reversing the effect of a convolution". The only objective is to learn the up-sampling kernels. In that case, both "deconvolution" and "transposed convolution" are not correct. | |
| Aug 10, 2016 at 14:01 | comment | added | Demonedge | Thanks for this very intuitive answer, but I'm confused about why the second one is the 'stride two' case, it behaves exactly like the first one when kernel moves. | |
| Aug 8, 2016 at 14:08 | comment | added | Martin Thoma | By the way: It is called transposed convolution now in TensorFlow: tensorflow.org/versions/r0.10/api_docs/python/… | |
| Jul 30, 2016 at 13:06 | comment | added | Stas S | Why do you say "no padding" in Figure 1, if actually input is zero-padded? | |
| Jun 30, 2016 at 20:47 | comment | added | David Dao | Yes, a deconvolution layer performs also convolution! That is why transposed convolution fits so much better as name and the term deconvolution is actually misleading. | |
| Jun 8, 2016 at 5:00 | comment | added | Martin Thoma | Just to make sure I understood it: "Deconvolution" is pretty much the same as convolution, but you add some padding? (Around the image / when s > 1 also around each pixel)? | |
| Jun 8, 2016 at 4:58 | vote | accept | Martin Thoma | ||
| Jun 7, 2016 at 20:15 | history | answered | David Dao | CC BY-SA 3.0 |