I believe this is a duplicate of #889, which contains a workaround. I'll close this one, but we still have to fix it.

It's not a duplicate, atrous convolution is adding holes in the input patch, whereas #889 is about dense patches but large strides.

Hello Martin,

Thanks for the super quick reply.
I attached a short PPT to illustrate the a-trou Algorithm.

Let me know if you need additional details.
Your tutorial on deep dream was great !!

On Thu, Apr 7, 2016 at 5:12 PM, Martin Wicke notifications@github.com
wrote:

I believe this is a duplicate of #889
#889, which contains a
workaround. I'll close this one, but we still have to fix it.

—
You are receiving this because you authored the thread.
Reply to this email directly or view it on GitHub
#1815 (comment)

For future reference, this is supported in Caffe master branch now and it is one of the features in Caffe rc3 https://github.com/BVLC/caffe/releases/tag/rc3. It is called dilation in convolution parameter and added by pull request BVLC/caffe#3487.

There was detailed discussion about whether this feature should be named dilation, hole or atrous in this pull request BVLC/caffe#3452. Relevant discussion started on Dec 18, 2015 in the pull request. This feature is eventually called dilation in convolution or dilated convolution following the term in Holschneider et al. (1987). The Caffe developers concluded that "dilated convolution" is a historically more accurate name than "à trous convolution".

@gpapan I assigned you based on your email to @yaroslavvb. Thanks!

Thanks. This will be ready in the next few days.

The change is in the process of being committed to github's Tensorflow.

Thank @gpapan for adding this to TensorFlow!

You are very welcome @fyu! Hope that people find this feature useful.

Hi,
Thank you for having implemented the new feature. I'm doing some test and so far it looks great. What is the reason the regular stride parameter has been removed from the API ?
Does it make sense to have dilated input map processed with different overlapped receptive field (stride parameter) ?

all the best
Laurent
Sent from my iPhone

On Apr 27, 2016, at 8:16 PM, George Papandreou notifications@github.com wrote:

You are very welcome @fyu! Hope that people find this feature useful.

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

@laurentk67
I have not yet come across a use case that needs both stride > 1 and rate > 1, so I decided to omit the stride parameter from the API:
You typically use atrous_conv2d() in order to preserve the spatial resolution of your input feature map, whereas conv2d() with stride greater than one lowers the resolution. Not exposing conv2d's stride parameter thus prevents the user from using atrous convolution incorrectly. If some legitimate use case of atrous convolution comes up where both stride and rate may be greater than one, we can easily add a stride parameter and pass it along to conv2d().

Having atrous convolution in TensorFlow is great - thank you @gpapan. I'd like access to strides for learning multi-scale raw audio filters, where using a stride of 1 is computationally infeasible.

You should file a new issue with that feature request, I doubt it will be seen hidden in this thread.