Deep Learning - Feel the Wild

Recently in the group will explain the framework, in the feeling wild this small knowledge point, we began to produce ambiguity, today I will briefly explain this small knowledge point, but also to bring a beginner to the Receptive Field brand new understanding, if you have an in-depth understanding of just, you can just skip O(∩_∩)O~~!

Now to get down to business!!!

My understanding used to be that feeling wild, that's actually a visual perception area size. For a single layer network, a pixel point in the next layer whose perceptual field size is also the size of the convolutional layer filter is actually quite plainly understood when you think about it, but for multiple layers, that's a little (and just that little bit more complicated)!

Formal definition.

``` In neural networks, the perceptual field is defined as
The size of the region mapped on the original image by the pixel points on the feature map (Feature ap) output from each layer of the convolutional neural network.```

Drew a random example picture myself, mostly depending on the content, O(∩_∩)O thanks!

A reader said a motion picture would be nice, so I've intercepted one for your enjoyment!

Now let's start talking about how to calculate it! (Here comes the watch) It's actually quite simple!

First it can be simply known (and mentioned earlier) that the size of the perceptual field of the output feature image vector of the first convolutional layer is equal to equal to the size of the convolutional layer filter; then its continued forward propagation, in which case the size of the perceptual field of the deeper convolutional layers later is related to the filter size and step size of all previous network layers, and the size of the image Padding is ignored during the calculation. Each layer in the network has a stride which is the product of all previous layer strides, i.e.

According to the information, the perceptual field size is calculated from the deepest layer to the front layer, that is, the deepest layer on the front layer is calculated first, and then gradually feedback to the first layer, the formula is written as follows.

among others

To get the feel of the wild size.

is the perceived wild size of the final layer in the previous layer.

is the convolution layer filter size.

By iterating in this way it is possible to obtain the feeling field for each layer. I see the exact code is also available online, so I'll attach it in passing... the original is available at the link, thanks!

The specific simulation results I got by running Python 3.5 (32-bit) for Windows as follows.

Recommended>>
1、Can a robot understand you
2、IOTA is up again today but that vulnerability that cost users 4 million is still there
3、How to do better permission management for backend applications
4、An application study of a new pharmacy service system based on data center
5、Stephen Hawking died today what precious treasures did he leave behind in the world

已推荐到看一看 和朋友分享想法
最多200字，当前共 发送

已发送

确定
分享你的想法...
取消

确定
最多200字，当前共

发送中