Skip to content

Commit

Permalink
fixed
Browse files Browse the repository at this point in the history
  • Loading branch information
FelixFu520 committed Nov 2, 2021
1 parent 84663a0 commit bc3a1e8
Show file tree
Hide file tree
Showing 5 changed files with 129 additions and 7 deletions.
91 changes: 86 additions & 5 deletions README.md
Original file line number Diff line number Diff line change
@@ -1,9 +1,6 @@
# 🍭理论篇

- [200+本计算机书籍](https://github.com/imarvinle/awesome-cs-books)
- [CUDA](https://docs.nvidia.com/cuda/index.html#)
- [TRT](https://docs.nvidia.com/deeplearning/tensorrt/index.html)
- [CVPR](https://github.com/extreme-assistant/CVPR2021-Paper-Code-Interpretation)
-

## 1⃣️ 数学

Expand Down Expand Up @@ -58,6 +55,9 @@
- [extern “C”的作用详解](https://www.cnblogs.com/xiangtingshen/p/10980055.html)
- [C++程序利用std::cout打印到日志文件方法](https://blog.csdn.net/weixin_42096202/article/details/109364061)
- [C/C++ 获取文件夹下所有文件名 windows和linux通用](https://blog.csdn.net/u012005313/article/details/50687297)
- [std::string 与 char* 之间的转换](https://blog.csdn.net/u010023457/article/details/41855007)
- [文件复制和移动](https://www.cnblogs.com/mfryf/archive/2013/03/17/2964207.html)| [2](https://blog.csdn.net/u012750702/article/details/52738859)|[3](https://blog.csdn.net/yc461515457/article/details/57432680)|[4](https://www.jianshu.com/p/a8cc836d5dff)
- [字符串操作](https://blog.csdn.net/xjw532881071/article/details/49154911)|[2](https://blog.csdn.net/yuejisuo1948/article/details/88643491)|[3](https://blog.csdn.net/dcrmg/article/details/99650783)|[4](https://blog.csdn.net/man_sion/article/details/54847573)|[5](https://blog.csdn.net/guowenyan001/article/details/11231927)

#### 🍃 C++工具

Expand Down Expand Up @@ -150,6 +150,36 @@

- 🐾 [docker容器python中文乱码](theory/python/zh_error.md)
- [list 排列组合](https://blog.csdn.net/vitaminc4/article/details/78922612)
- [Python中下划线的5种含义](https://zhuanlan.zhihu.com/p/36173202)
- [numpy: np.logical_and/or/not (逻辑与/或/非)](https://blog.csdn.net/JNingWei/article/details/78651535)
- [np.where](https://www.cnblogs.com/massquantity/p/8908859.html)
- [np.vstack, np.hstack](https://blog.csdn.net/u012609509/article/details/70319293)
- [Opencv实现多幅图像显示在同一窗口(基于Python)](https://www.codenong.com/cs105613611/)
- [opencv读中文图片](https://blog.csdn.net/PanYHHH/article/details/111561477)
- [pytorch计算余弦相似度](https://blog.csdn.net/tszupup/article/details/100711874)
- [(torch.full()、torch.full_like())创造全value的矩阵](https://blog.csdn.net/Fluid_ray/article/details/109855155)
- [torch.cat 和 torch.stack的区别](https://blog.csdn.net/wukong999/article/details/89284055)
- [ToTensor 到底做了什么?](http://spytensor.com/index.php/archives/42/)
- [named_children()和named_modules()的区别](https://blog.csdn.net/watermelon1123/article/details/98036360)
- [transforms的二十二个方法](https://zhuanlan.zhihu.com/p/53367135)
- [pytorch手动实现滑动窗口操作,论fold和unfold函数的使用](https://blog.csdn.net/LoseInVain/article/details/88139435) | [2](https://blog.csdn.net/qq_34914551/article/details/102940368)
- [Pytorch中index_select() 函数的理解](https://blog.csdn.net/kdongyi/article/details/103099589)
- [pytorch:torch.clamp()](https://blog.csdn.net/u013230189/article/details/82627375)
- [torch.nonzero的理解](https://blog.csdn.net/monchin/article/details/79750216)
- [Pytorch中index_select() 函数的理解](https://blog.csdn.net/kdongyi/article/details/103099589)
- [PIL读入图片转为BGR](https://blog.csdn.net/zicai_jiayou/article/details/115541307)
- [np.random.randint、np.random.choice、random.sample三种随机函数的用法案例](https://blog.csdn.net/zsc201825/article/details/80918450)
- [Python——自定义排序 sort() / list.sort()](https://blog.csdn.net/Eternity666/article/details/108301177)
- [如何直观地理解「协方差矩阵」?](https://zhuanlan.zhihu.com/p/37609917)
- [pprint的用法](https://blog.csdn.net/u013061183/article/details/79094460)
- [python dir()和vars()的区别](https://blog.csdn.net/sxingming/article/details/52116157)
- [Python装饰器abstractmethod、property、classmethod、staticmethod及自定义装饰器](https://www.jianshu.com/p/a340ffb231a0)
- [python 获取当前日期的前n天或后n天](https://blog.csdn.net/qq_15058425/article/details/90268190)
- [Python计数器collections.Counter用法详解](https://blog.csdn.net/sinat_28576553/article/details/99131954)
- [python读写json文件](https://www.cnblogs.com/bigberg/p/6430095.html)
- [python限定方法参数类型、返回值类型、变量类型等](https://www.cnblogs.com/xxpythonxx/p/12198876.html)
- [【python】错误解决经历](https://www.cnblogs.com/ytxwzqin/p/12012025.html) |[2](https://www.cnblogs.com/henuliulei/p/13258297.html)
-

### 🍬Linux

Expand Down Expand Up @@ -242,6 +272,8 @@
- 🐾 [实验管理工具sacred](envs/tools/scared.md)
- 🐾 [bypy 百度云下载python](envs/tools/bypy.md)
- 🐾 [Python日志库loguru](envs/tools/loguru.md)
- [字典2dot--dotmap](https://github.com/drgrib/dotmap)
-

### 🍬 IDE&Tools安装

Expand Down Expand Up @@ -321,7 +353,7 @@

#### 🍃CNN

- 🐾 [CNN介绍](train/cnn/introduce_cnn.md) --[🐾CNN网络代码讲解](https://github.com/WZMIAOMIAO/deep-learning-for-image-processing)
- 🐾 [CNN介绍](train/cnn/introduce_cnn.md) --[🐾CNN网络代码讲解加部署 ⭐](https://github.com/WZMIAOMIAO/deep-learning-for-image-processing)
- 🐾 [LeNet, AlexNet, VGG](train/cnn/lenet_alexnet_vgg.md)
- 🐾 [DSC理解](train/cnn/dsc.md)
- 🐾 [Inception](train/cnn/inception.md)
Expand Down Expand Up @@ -409,6 +441,9 @@

- 🐾 [交叉墒](train/loss/entropy.md)
- 🐾 [Pytorch的损失函数](train/loss/losses_pytorch.md)
- [三个相关系数](https://www.cnblogs.com/yjd_hycf_space/p/11537153.html)
- [机器学习-距离度量](https://blog.csdn.net/xjp_xujiping/article/details/108576496)|[2](https://zhuanlan.zhihu.com/p/46626607)
- [方差](https://blog.csdn.net/lijinxiu123/article/details/52450858)|[2](https://geek.digiasset.org/pages/mathbasic/correlation-co-variances-variance-coeffi_21Mar07115144053588/#%E4%B8%89%E5%8D%8F%E6%96%B9%E5%B7%AE)|[3](https://zhuanlan.zhihu.com/p/86181679)|[4](https://zhuanlan.zhihu.com/p/68967565)|[5](https://blog.csdn.net/lilong117194/article/details/78399568)
- Lovasz-Softmax Loss
- Exponential Logarithmic loss
- Focal Loss + Dice Loss
Expand Down Expand Up @@ -613,3 +648,49 @@

- 🐾 [onnxruntime, tensorrt, tvm, ...](https://github.com/FelixFu520/deploy_demo)

# 🍭书签

- [200+本计算机书籍](https://github.com/imarvinle/awesome-cs-books)
- [CUDA](https://docs.nvidia.com/cuda/index.html#)
- [TRT](https://docs.nvidia.com/deeplearning/tensorrt/index.html)
- [CVPR](https://github.com/extreme-assistant/CVPR2021-Paper-Code-Interpretation)

### Language

#### C++

- **[C++中文网](http://c.biancheng.net/cpp/)**
- [C++ reference - cppreference.com](https://en.cppreference.com/w/cpp)

- [Boost C++](https://wizardforcel.gitbooks.io/the-boost-cpp-libraries/content/0.html)
- [cplusplus](http://www.cplusplus.com/)
- [C++官方](https://isocpp.org/)
- [GCC参数](https://www.runoob.com/w3cnote/gcc-parameter-detail.html)

#### Python

- [Python教程](http://www.coolpython.net/python_senior/index.html)|[2](https://www.liaoxuefeng.com/wiki/1016959663602400)

- [Python官方](https://docs.python.org/zh-cn/3/index.html)
- [PYPI](https://pypi.org/manage/projects/)
- [PyGObject](https://pygobject.readthedocs.io/en/latest/getting_started.html)
- [GTK+](https://python-gtk-3-tutorial.readthedocs.io/en/latest/index.html#)
- [pycuda](https://documen.tician.de/pycuda/driver.html)
- [pandas中文教程](https://www.w3cschool.cn/hyspo/)
- [numpy](https://www.numpy.org.cn/reference/) | [2](https://www.numpy.org.cn/about/)
- [matplotlib](https://www.yiibai.com/matplotlib/pyplot-tutorial.html)|[2](https://matplotlib.org/)
- [scikit-image](https://scikit-image.org/docs/stable/)

#### 其他

- [git](https://www.liaoxuefeng.com/wiki/896043488029600) | [2](https://backlog.com/git-tutorial/cn/)

#### 读图

- [数字图像处理](http://www.vue5.com/dip/histogram_stretching.html)
- [Opencv-python教程](https://www.osgeo.cn/opencv-python/index.html)
- [Opencv-c++](https://docs.opencv.org/master/d2/d96/tutorial_py_table_of_contents_imgproc.html) |[2](https://docs.opencv.org/3.4.3/d9/df8/tutorial_root.html) |[3](https://www.yiibai.com/opencv/)
- [Pillow]()
-

- https://pycharm.iswbm.com/
2 changes: 1 addition & 1 deletion train/cnn/dsc.md
Original file line number Diff line number Diff line change
Expand Up @@ -52,7 +52,7 @@



Pointwise Convolution实际为1×1卷积,在DSC中它起两方面的作用。第一个作用是让DSC能够自由改变输出通道的数量;第二个作用是对Depthwise Convolution输出的feature map进行通道融合。第一个作用比较容易理解,这因为单独的Depthwise Convolution无法改变输出通道数量,因而采用1×1卷积来改变输出通道数量是比较直观和简单的做法。为理解Pointwise Convolution的第二点作用,可以考虑在只使用Depthwise Convolution来堆叠网络时会发生什么情况。假设输入为IN,它的第i个通道记为IN_i;第一层Depthwise Convolution的输出记为DC1,它的第i个通道记为DC1_i;第二层Depthwise Convolution的输出记为DC2,它的第i个通道记为DC2_i。
Pointwise Convolution实际为1×1卷积,在DSC中它起两方面的作用。第一个作用是让**DSC能够自由改变输出通道的数量**;第二个作用是**对Depthwise Convolution输出的feature map进行通道融合**。第一个作用比较容易理解,这因为单独的Depthwise Convolution无法改变输出通道数量,因而采用1×1卷积来改变输出通道数量是比较直观和简单的做法。为理解Pointwise Convolution的第二点作用,可以考虑在只使用Depthwise Convolution来堆叠网络时会发生什么情况。假设输入为IN,它的第i个通道记为IN_i;第一层Depthwise Convolution的输出记为DC1,它的第i个通道记为DC1_i;第二层Depthwise Convolution的输出记为DC2,它的第i个通道记为DC2_i。

由Depthwise Convolution的工作机理可知,DC1_i只与IN_i有关,DC2_i只与DC1_i有关,进而DC2_i也只与IN_i有关。简单来说,输入、输出各不同通道间没有任何计算将它们联系起来。而1×1卷积本身就具有通道融合的能力,因此在Depthwise Convolution之后接Pointwise Convolution能够有效解决上述问题。

Expand Down
2 changes: 1 addition & 1 deletion train/cnn/mobilenet.md
Original file line number Diff line number Diff line change
Expand Up @@ -79,7 +79,7 @@ MobileNet 需要尽可能维持其中发展较快的计算机视觉和深度学

![img](imgs/1226410-20191107145306585-415671776.png)

 与标准卷积网络不一样的是,我们将卷积核拆分成单通道形式,在不改变输入特征图像的深度的情况下,对每一通道进行卷积操作,这样就得到了和输入特征图通道数一致的输出特征图。如上图,输入`12*12*3 `的特征图,经过`5*5*1*3`的深度卷积之后,得到了8*8*3的输出特征图。输入和输出的维度是不变的3,这样就会有一个问题,通道数太少,特征图的维度太少,能获得足够的有效信息吗?
 与标准卷积网络不一样的是,我们将卷积核拆分成单通道形式,在不改变输入特征图像的深度的情况下,对每一通道进行卷积操作,这样就得到了和输入特征图通道数一致的输出特征图。如上图,输入`12*12*3 `的特征图,经过`5*5*1*3`的深度卷积之后,得到了`8*8*3`的输出特征图。输入和输出的维度是不变的3,这样就会有一个问题,通道数太少,特征图的维度太少,能获得足够的有效信息吗?

#### 1.2.2 逐点卷积

Expand Down
41 changes: 41 additions & 0 deletions train/detection/yolox.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,41 @@
# YoloX

⌚️: 2021年11月02日

📚参考

- [YoloX网络结构](https://blog.csdn.net/nan355655600/article/details/119329727)
-

---

Yolox共有**七种**网络结构,包含**2种**轻量级网络,和**5种**标准网络。

**① 轻量级网络**
(1) Yolox-Nano可视化网络结构图:点击查看

(2) Yolox-Tiniy可视化网络结构图:点击查看



**② 标准网络**
(1)Yolox-s可视化网络结构图:点击查看

(2) Yolox-m可视化网络结构图:点击查看

(3) Yolox-l可视化网络结构图:点击查看

(4) Yolox-x可视化网络结构图:点击查看

(5) Yolox-Darknet53可视化网络结构图:点击查看



此外也放上Yolov3、Yolov4、Yolov5的网络结构,可以对比查看:
(1) Yolov3可视化网络结构图:点击查看
(2)Yolov4可视化网络结构图:点击查看
(3) Yolov5可视化网络结构图:点击查看

但最好的方式,还是使用netron工具打开cfg文件或者Yolox的onnx文件,可以看到网络的各个细节,因此也放上下载链接,和大白画得网络结构图一起查看,思路更加清晰:
(1) 大白的另一篇文章《网络可视化工具netron详细安装流程》:点击查看

Empty file added train/detection/yolox_arch.md
Empty file.

0 comments on commit bc3a1e8

Please sign in to comment.