# 用反卷积（Deconvnet）可视化理解卷积神经网络还有使用tensorboard

[En]

Second, using deconvolution to realize feature visualization

1、反池化过程

2、反激活

3、反卷积

[En]

For the deconvolution process, the filter after the convolution process is used (the parameters are the same, but the parameter matrix is flipped horizontally and vertically), which I do not quite understand now, and it is estimated that we should use the relevant mathematical theory to prove it.

[En]

Finally, the visual network structure is as follows:

[En]

Generally speaking, the algorithm mainly has two key points: 1, anti-pooling 2, deconvolution, these two source code implementation methods, need to be well understood.

[En]

Third, understand visualization

[En]

Feature visualization: once we have finished our online training, we can visualize and see what we have learned. But what do you think? How to understand it is another matter. We use the deconvolution network above to view the feature map of each layer.

1、特征可视化结果：

2、特征学习的过程。作者给我们显示了，在网络训练过程中，每一层学习到的特征是怎么变化的，上面每一整张图片是网络的某一层特征图，然后每一行有8个小图片，分别表示网络epochs次数为：1、2、5、10、20、30、40、64的特征图：

3、图像变换。从文献中的图片5可视化结果，我们可以看到对于一张经过缩放、平移等操作的图片来说：对网络的第一层影响比较大，到了后面几层，基本上这些变换提取到的特征没什么比较大的变化。

1、《Visualizing and Understanding Convolutional Networks》

2、《Adaptive deconvolutional networks for mid and high level feature learning》

3、《Stacked What-Where Auto-encoders》

Original: https://www.cnblogs.com/Anita9002/p/9455856.html
Author: Anita-ff
Title: 用反卷积（Deconvnet）可视化理解卷积神经网络还有使用tensorboard

(0)