Skip to content

Commit

Permalink
Merge pull request lartpang#19 from LarT2P/master
Browse files Browse the repository at this point in the history
更新内容, 补充可变形卷积网络, 调整之前的内容.
  • Loading branch information
lartpang committed Dec 1, 2018
2 parents 7500b7f + 2bcbb0d commit 498d03b
Show file tree
Hide file tree
Showing 37 changed files with 642 additions and 304 deletions.
8 changes: 4 additions & 4 deletions ML预处理.md
Original file line number Diff line number Diff line change
Expand Up @@ -86,27 +86,27 @@

### 自然灰度图像

==均值消减->PCS/ZCA白化==
均值消减->PCS/ZCA白化

灰度图像具有平稳特性,我们通常在第一步对每个数据样本分别做均值消减(即减去直流分量),然后采用 PCA/ZCA 白化处理,其中的 `epsilon`**足够大以达到低通滤波**的效果。

### 彩色图像

==简单缩放->PCA/ZCA白化==
简单缩放->PCA/ZCA白化

对于彩色图像,<u>色彩通道间并不存在平稳特性</u>。因此我们通常首先对数据进行特征缩放(使像素值位于 [0,1] 区间),然后使用足够大的 `epsilon` 来做 PCA/ZCA。**注意**在进行 PCA 变换前需要对特征进行分量均值归零化。

### 音频 (MFCC/频谱图)

==特征标准化->PCA/ZCA 白化==
特征标准化->PCA/ZCA 白化

对于音频数据 (MFCC 和频谱图),每一维度的取值范围(方差)不同。

例如 MFCC 的第一分量是直流分量,通常其幅度远大于其他分量,尤其当特征中包含时域导数 (temporal derivatives) 时(这是音频处理中的常用方法)更是如此。因此,**对这类数据的预处理通常从简单的数据标准化开始**(即使得数据的每一维度均值为零、方差为 1),然后进行 PCA/ZCA 白化(使用合适的 `epsilon`)。

### MNIST 手写数字

==简单缩放/逐样本均值消减(->PCA/ZCA 白化)==
简单缩放/逐样本均值消减(->PCA/ZCA 白化)

MNIST 数据集的像素值在 [0,255] 区间中。我们首先将其缩放到 [0,1] 区间。实际上,进行逐样本均值消去也有助于特征学习。

Expand Down
Binary file removed Paper/DeeplearningReview-上.png
Binary file not shown.
453 changes: 453 additions & 0 deletions Paper/Deformable Convolutional Networks总结(2017).md

Large diffs are not rendered by default.

2 changes: 2 additions & 0 deletions Paper/GoogLeNet(InceptionV1)总结(2014).md
Original file line number Diff line number Diff line change
Expand Up @@ -46,6 +46,8 @@ Christian在思**考如何才能够减少深度神经网络的计算量,同时

#### 辅助分类器

![1537605876131](../assets/1537605876131.png)

添加到这些中间层. 深度相对较大的网络,有效传播梯度反向通过所有层的能力是一个问题。

在这个任务上,更浅网络的强大性能表明网络中部层产生的特征应该是非常有识别力的。**通过将辅助分类器添加到这些中间层,可以期望较低阶段分类器的判别力**
Expand Down
Empty file added Paper/R-FCN总结(2016).md
Empty file.
19 changes: 19 additions & 0 deletions Paper/STN总结(2015).md
Original file line number Diff line number Diff line change
@@ -0,0 +1,19 @@
# STN(2015)

> Spatial Transformer Networks 论文: https://arxiv.org/abs/1506.02025
>
> 详细解读Spatial Transformer Networks: https://blog.csdn.net/qq_39422642/article/details/78870629
>
> 本文主要做摘录.
## STN作用

普通的CNN能够显示的学习平移不变性,以及隐式的学习旋转不变性,但attention model 告诉我们,与其让网络隐式的学习到某种能力,不如为网络设计一个显式的处理模块,专门处理以上的各种变换。因此,DeepMind就设计了Spatial Transformer Layer,简称STL来完成这样的功能。



## STN架构

## STN流程

## STN 代码
4 changes: 4 additions & 0 deletions Paper/YOLOV2V3总结.md
Original file line number Diff line number Diff line change
Expand Up @@ -2,6 +2,8 @@

## YOLOV2

> https://arxiv.org/abs/1612.08242
YOLO与Fast R-CNN相比有较大的定位误差,与基于region proposal的方法相比具有较低的召回率。因此YOLO v2主要改进是提高召回率和定位能力。

论文主要工作有两部分:
Expand Down Expand Up @@ -81,6 +83,8 @@ YOLO2是依赖于DarkNet-19的结构,这个模型在ImageNet上能达到91%的

## YOLOV3

> https://arxiv.org/abs/1804.02767
### 改进之处

- 多尺度预测 (类FPN)
Expand Down
Binary file added Paper/assets/1543569084925.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added Paper/assets/1543569107259.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added Paper/assets/1543583098259.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added Paper/assets/1543584655182.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added Paper/assets/1543584719028.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added Paper/assets/1543584778874.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added Paper/assets/1543584932404.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added Paper/assets/1543645757847.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added Paper/assets/1543648639596.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added Paper/assets/1543648968239.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added Paper/assets/1543650893502.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added Paper/assets/1543651166636.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added Paper/assets/1543652081019.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added Paper/assets/1543658023608.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Binary file added Paper/assets/1543658566310.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
33 changes: 1 addition & 32 deletions Paper/readme.md
Original file line number Diff line number Diff line change
Expand Up @@ -93,38 +93,7 @@
[中文版](http://noahsnail.com/2018/03/20/2018-03-20-Feature%20Pyramid%20Networks%20for%20Object%20Detection%E8%AE%BA%E6%96%87%E7%BF%BB%E8%AF%91%E2%80%94%E2%80%94%E4%B8%AD%E6%96%87%E7%89%88/)
[中英文对照](http://noahsnail.com/2018/03/20/2018-03-20-Feature%20Pyramid%20Networks%20for%20Object%20Detection%E8%AE%BA%E6%96%87%E7%BF%BB%E8%AF%91%E2%80%94%E2%80%94%E4%B8%AD%E8%8B%B1%E6%96%87%E5%AF%B9%E7%85%A7/)

## OCR

* CRNN
[An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application to Scene Text Recognition](https://arxiv.org/abs/1507.05717)
[中文版](http://noahsnail.com/2017/08/21/2017-8-21-CRNN%E8%AE%BA%E6%96%87%E7%BF%BB%E8%AF%91%E2%80%94%E2%80%94%E4%B8%AD%E6%96%87%E7%89%88/)
[中英文对照](http://noahsnail.com/2017/08/21/2017-8-21-CRNN%E8%AE%BA%E6%96%87%E7%BF%BB%E8%AF%91%E2%80%94%E2%80%94%E4%B8%AD%E8%8B%B1%E6%96%87%E5%AF%B9%E7%85%A7/)

* CTPN
[Detecting Text in Natural Image with Connectionist Text Proposal Network](https://arxiv.org/abs/1609.03605)
[中文版](http://noahsnail.com/2018/02/02/2018-02-02-Detecting%20Text%20in%20Natural%20Image%20with%20Connectionist%20Text%20Proposal%20Network%E8%AE%BA%E6%96%87%E7%BF%BB%E8%AF%91%E2%80%94%E2%80%94%E4%B8%AD%E6%96%87%E7%89%88/)
[中英文对照](http://noahsnail.com/2018/02/02/2018-02-02-Detecting%20Text%20in%20Natural%20Image%20with%20Connectionist%20Text%20Proposal%20Network%E8%AE%BA%E6%96%87%E7%BF%BB%E8%AF%91%E2%80%94%E2%80%94%E4%B8%AD%E8%8B%B1%E6%96%87%E5%AF%B9%E7%85%A7/)

## Mobile

* MobileNetV2
To be added.

## Reward

Will anyone find this?

* Wechat

<img width="275" alt="微信" src="https://user-images.githubusercontent.com/21311442/44447671-7a3dd300-a61b-11e8-88cb-d815acb67c5e.png">



* Alipay

<img width="275" alt="支付宝" src="https://user-images.githubusercontent.com/21311442/44447780-cee14e00-a61b-11e8-8538-70398f84c53a.png">


---

http://noahsnail.com/2017/07/21/2017-7-21-GoogleNet%E8%AE%BA%E6%96%87%E7%BF%BB%E8%AF%91%E2%80%94%E2%80%94%E4%B8%AD%E8%8B%B1%E6%96%87%E5%AF%B9%E7%85%A7/)

Expand Down
39 changes: 0 additions & 39 deletions Paper/各种模型概述.md

This file was deleted.

117 changes: 65 additions & 52 deletions readme.md
Original file line number Diff line number Diff line change
Expand Up @@ -2,69 +2,82 @@

在cs231n笔记的基础上**进一步扩展**, 是我ML&DL学习总结的记录.

---
<details>
<summary> changelog</summary>

> changelog:
>
> 2018年11月15日:
>
> 最近在看目标检测部分内容, 主要会看RCNN->SPP-Net->Fast RCNN->Faster RCNN, 总结的书写方式要改变下, 为了节省时间, 提高效率, 更多增加自己的思考, 决定不再弄论文的翻译复制过来, 在上面的基础上做笔记这样的方式了, 准备按照架构的流程, 理清架构的思路, 以问题推动思考的方式来进行学习.
>
> 希望可以帮助自己更深入的理解.
>
> 2018年11月19日:
>
> 准备开始看Yolo/SSD系列.
>
> 这几天看完了RCNN系列的文章, 略感心累.
>
> 准备过些日子重新在整理下文档结构吧.
2018年11月15日:

---
最近在看目标检测部分内容, 主要会看RCNN->SPP-Net->Fast RCNN->Faster RCNN, 总结的书写方式要改变下, 为了节省时间, 提高效率, 更多增加自己的思考, 决定不再弄论文的翻译复制过来, 在上面的基础上做笔记这样的方式了, 准备按照架构的流程, 理清架构的思路, 以问题推动思考的方式来进行学习.

希望可以帮助自己更深入的理解.

2018年11月19日:

准备开始看Yolo/SSD系列.

## 更新计划(成文-1/未动手-0)

1. 论文
1. 2012
1. AlexNet-1
2. 2013
1. NiN-1
3. 2014
1. OverFeat-1
2. GoogLeNet-1
3. VGG-1
4. R-CNN-1
5. SPP-Net-1
4. 2015
1. BN-GoogLeNet-1
2. InceptionV2/V3-1
3. ResNet-1
4. FCN-1
5. Deconvolution Network(Semantic Segmentation)-1
6. Fast R-CNN-1
7. YOLO-V1-1
5. 2016
1. Faster R-CNN-1
6. 2017
1. SeNet-1
2. DenseNet-1
3. SqueezeNet-1
7. 2018

2.other
这几天看完了RCNN系列的文章, 略感心累.

准备过些日子重新在整理下文档结构吧.

2018年11月30日:

看完了RCNN系列, 看完了SSD&Faster的实现, YOLO简单过了一下, 接下来准备看下其他的关于目标检测算法论文.

</details>

<details>
<summary>已读论文</summary>

1. 2012
1. AlexNet
2. 2013
1. NiN
3. 2014
1. OverFeat
2. GoogLeNet
3. VGG
4. R-CNN
5. SPP-Net
4. 2015
1. BN-GoogLeNet
2. InceptionV2/V3
3. ResNet
4. FCN
5. Deconvolution Network(Semantic Segmentation)
6. Fast R-CNN
7. YOLO-V1
8. SSD
5. 2016
1. Faster R-CNN
2. YOLO-V2
6. 2017
1. SeNet
2. DenseNet
3. SqueezeNet
7. 2018
1. YOLO-V3

</details>

<details>
<summary>未读论文</summary>

- Xception-0
- InceptionV4-0
- Inception ResNet V1/V2-0
- ResNeXt-0
- Mobile系列-0
- SSD-0
- RetinaNet-0

3. 调整文件结构
1. 基础文章收集到一起
2. 论文部分进行一下分类归档
3. 删除文章未使用的图片
</details>

## 后期想法

调整文件结构

1. 基础文章收集到一起
2. 论文部分进行一下分类归档
3. 删除文章未使用的图片

---

Expand Down
Loading

0 comments on commit 498d03b

Please sign in to comment.