Caffe: How can I understand a deconvolution layer

Created on 24 Mar 2016 · 4Comments · Source: BVLC/caffe

Hi,

I have read codes in caffe about deconvolution layer. But I'm confused about the codes. In convolutional layer, it is easy to understand, while it is difficult for me to understand the deconvolution layer, I cannot get a formula how to do deconvolution from the codes. Can someone help me?

Thanks.

Source

ginobilinie

All 4 comments

In the easiest case (stride = 1), deconvolution is the same as a full convolution with the weights flipped on the axis.
See:
http://www.johnloomis.org/ece563/notes/filter/conv/convolution.html

The standard convolution would be a "valid" convolution in the illustration above. As the weights are machine learned, they don't really need to be flipped on the axis / inverted.

naibaf7 on 24 Mar 2016

👎2

@naibaf7

Thank you. if the stride is larger than 1, then for deconvolution, what's the situation? According the the FCN paper, "Upsampling is backwards strided convolution", I do not understand it well.

ginobilinie on 24 Mar 2016

@naibaf7

I got it. Thank you.

ginobilinie on 24 Mar 2016

😕10 🎉1

From https://github.com/BVLC/caffe/blob/master/CONTRIBUTING.md:

_Please do not post usage, installation, or modeling questions, or other requests for help to Issues._
Use the caffe-users list instead. This helps developers maintain a clear, uncluttered, and efficient view of the state of Caffe.