deconvolution interface #3872
deconvolution interface #3872
Conversation
|
@dwf, since you opened the original issue, your feedback would be welcome. |
| input. Generates output of shape: input shape - filter shape + 1 | ||
| * ``'full'``: apply filter wherever it partly overlaps with the input. | ||
| Generates output of shape: input shape + filter shape - 1 | ||
| * ``'half'``: pad input with a symmetric border of ``filter rows // 2`` |
There was a problem hiding this comment.
The "traditional" border modes are usually full, valid and same (see e.g. matlab's conv).
I suggest to implement same, maybe in place of half which seems almost the same.
There was a problem hiding this comment.
I would just add the 'same' mode, but maybe as part of another PR.
This is just exposing an interface and "half" is already there.
There was a problem hiding this comment.
same is tricky, because it may mean uneven padding on both sides, which is not implemented in some back-ends (if not all), or taking a subtensor (which makes the tensor non-contiguous).
This is why we have half, which is equivalent to same when same is achievable with symmetric padding.
There was a problem hiding this comment.
I didn't know that, let's stick with half then.
There was a problem hiding this comment.
I still think we should add 'same', but not in this PR.
|
Pascal and I checked Zeiler's deconvolutional networks papers (https://www.cs.nyu.edu/~gwtaylor/publications/zeilertaylorfergus_iccv2011.pdf, http://www.matthewzeiler.com/pubs/cvpr2010/cvpr2010.pdf). Apparently, what they do is using alternating layers of convolutional sparse coding instead of taking the gradient of conv outputs w.r.t its inputs. |
|
So, the discussion at TensorFlow for this functionality is at tensorflow/tensorflow#256. Interesting points:
TensorFlow settled on For the moment, I would be OK with merging this PR, and open a new ticket to add a |
| used by the convolution, such that the output_grad is upsampled | ||
| to the input shape. | ||
|
|
||
| :type output_grad: symbolic 4D tensor. |
There was a problem hiding this comment.
IMO it's quite confusing to call the input of the function output_grad. I understand why it was called like this, but probably something like input_tensor or 'to_be_upsampled_tensor' would be better.
There was a problem hiding this comment.
Let's keep output_grad here for now.
|
OK, so we agreed with @vdumoulin, @fvisin, and @nouiz to have a |
|
So in the end, there is only one small change, and then we can merge. Thanks! |
|
I applied the change. |
|
Thanks! |
PR for issue #3681