problem

[watchmexiang]

Hello, may I ask if this Autofocus Layer can directly replace Dilated convolution? How is the effect, thank you

[perslev]

Hi, bear in mind I am not among the authors of this paper (please refer to https://arxiv.org/pdf/1805.08403.pdf). However, this layer implements exactly dilated convolutions, although with multiple dilation rates applied in parallel using the same kernel. An attention mechanism then selects from the convolved signals produced under variable dilations to produce an output of appropriate scale given the context - hence the "autofocus". The idea is to let the network 'decide' on a suitable dilation rate for the given problem in each layer (while being able to dynamically change the receptive field given the context in a given layer).

That being said, I have had limited success with using this layer so far, although this may be due to an implementation error or lack of experimentation.

[watchmexiang]

Thank you for your reply. I would like to ask if you have experimented with replacing the Autofocus Layer with the convolution layer. How much improvement?

…

------------------ 原始邮件 ------------------ 发件人: "Mathias Perslev"<notifications@github.com>; 发送时间: 2019年4月30日(星期二) 下午3:36 收件人: "perslev/Autofocus-Layer-TF"<Autofocus-Layer-TF@noreply.github.com>; 抄送: "z"<1264754202@qq.com>; "Author"<author@noreply.github.com>; 主题: Re: [perslev/Autofocus-Layer-TF] problem (#1) Hi, bear in mind I am not the authors of this paper (please refer to https://arxiv.org/pdf/1805.08403.pdf). However, this layer implements exactly dilated convolutions, although with multiple dilation rates applied in parallel using the same kernel. An attention mechanism then selects from the convolved signals produced under variable dilations to produce an output of appropriate scale given the context - hence the "autofocus". The idea is to let the network 'decide' on a suitable dilation rate for the given problem in each layer (while being able to dynamically change the receptive field given the context in a given layer). That being said, I have had limited success with using this layer so far, although this may be due to an implementation error or lack of experimentation. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.