【原创 深度学习与TensorFlow 动手实践系列 - 4】第四课:卷积神经网络 - 高级篇
2017-06-19 15:41
706 查看
【原创 深度学习与TensorFlow 动手实践系列 - 4】第四课:卷积神经网络 - 高级篇
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/9541e4ff8e412cadb0ac0f2d0c3ddeac.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/a2faa3de42f95c203373579364c388b5.png)
提纲:
1. AlexNet:现代神经网络起源
2. VGG:AlexNet增强版
3. GoogleNet:多维度识别
4. ResNet:机器超越人类识别
5. DeepFace:结构化图片的特殊处理
6. U-Net:图片生成网络
7. 实例:剖析VGG,用模型进行模型参数可视化,特征提取,目标预测
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/e44b76361b038c500f063ddd90336704.png)
期待目标:
1. 掌握AlexNet结构特点,神经网络各层之间特征传导关系,模型参数总数计算
2. 了解VGG,GoogLeNet,ResNet等复杂ImageNet模型的结构特点,简单设计思想
3. 针对特殊数据,特殊任务设计的神经网络结构
4. 深度剖析VGG TF代码,学会对已有模型进行参数读取,目标预测,特征提取。
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/a33ebb6cd79702a30fac66d9a790e75a.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/05cf26dfd03b602186e125e282b3046e.png)
AlexNet:现代神经网络起源
背景介绍:
ImageNet Challenge:1000类物体,每类1000张图片
传统方法思路:
1. 图片特征提取
2. 机器学习分类
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/9faf79d58d0b05ce8adcdd18aad9c01c.png)
背景介绍:
2010年冠军
System Overview
Dense Grid descriptor:HOG,LBP
Coding:Local coordinate super-vector
Pooling, SPM
Linear SVM
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/4dcd85418cf2da9b4e9728e0943f750d.png)
2011年冠军:Xerox Lab
1. 特征提取
2. Fisher 压缩
3. SVM分类
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/67096d4bcfb5b10bb912a295685e725e.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/ec8de20c29c92e8707f3c6d83863a182.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/4be89b4e8d9048d0b55f0e0cad49e9a3.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/fb6deb238e8d588c6ff8fc437e30b5f3.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/3c28a6edc69ed662da4741ee0f31b008.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/8a0765e967feeeb7cf165602fa22c2b8.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/6dee36dba6fa965e1381cd0894b1722a.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/92c8ffd0539d71b4bef648b5aef15fdb.png)
VGG:AlexNet增强版
1. VGG-AlexNet 对比卷积层 - 卷积群参数个数:138m - 60m
2. 识别率(top5)7.3% - 15.3%
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/011fea889f90df36f07d6768c97da8a9.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/264fc14b867174aa5686c322f86c1f84.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/61ea3cd44744f019fee7938cee3a2d16.png)
VGG作用:
1. 结构简单:同AlexNet结构类似,均为卷积层,池化层,全连接层的组合。
2. 性能优异:同AlexNet提升明显,同GoogleNet,ResNet相比,表现接近。
3. 选择最多的基本模型:方便进行结构的优化,设计,SSD,RCNN,等其他任务的基本模型(base model)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/dd415ea3da32891e7e401d1dccc4f7ab.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/5a6ccc9289e1eb3f78428b73984dc9a2.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/aa5c863d1092115a1ea9bb71a47b69fe.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/3db512762a4fc49b2db6c5c31c7f877a.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/f69a9c821f33ae6538809f58b05f5533.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/9d09efacdc5f5ba35e3ae13f9924bf49.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/0b2c0f89be4604b5dd499053f0613829.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/8b12c2af52f0797c91935701777dfa1f.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/74a08a91532590deeec3f3b6163ad64b.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/8f8358e8cc6eacd3eeae3f3e48e58d5a.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/21d57057149ad5e888366bd0e96c8b99.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/fb4ae348645d2b80160d8951b572e747.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/11817e61b28231d90c8283f24035337a.png)
1. 为什么ResNet有效?
1. 前向计算:低层卷积网络高层卷经济网络信息融合;层数越深,模型的表现力越强。
2. 反向计算:导数传递更直接,越过模型,直达各层。
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/908129451eb49a51479c1a2fc61dadd5.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/ebbd6b620518c0dd192983b984cd1656.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/045ac7ba083aad66a72d453f48361214.png)
人脸识别数据特点:
结构化:所有人脸,组成相似,理论上能够实现对齐
差异化:相同位置,形貌不同
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/d3c054efc34e7f513dfe7386003318ba.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/1472bdb492a21c5cbef2ea29f93d4787.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/5cee452d42c1d34ba7c7025d419c2979.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/6af5841b61279a61c2243696ca38b546.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/c9340937c828de2eab30c21d2de55e18.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/7cb9dafd4600f46f437882cb25765412.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/73f9a1badf3bee04d68a72aecdc7cd30.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/7085dfecc205a69efcc59bf2d701e073.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/adeb0215f931f9174ea2bc48d336bbfb.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/6ff30faa0945fd9713ee22637c7becf8.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/5b036d24629242b432f633d950ac0bc8.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/fd8304e881ee021e585d074ffa984f67.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/6a45604b9f1f9d6fa18b635dcac63527.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/7f7575c8b9e27a36ce587040540089e8.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/8e6b857ff3b4bd990237d1d89aa4a610.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/751fae93df610820e038463a2d65787e.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/9541e4ff8e412cadb0ac0f2d0c3ddeac.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/a2faa3de42f95c203373579364c388b5.png)
提纲:
1. AlexNet:现代神经网络起源
2. VGG:AlexNet增强版
3. GoogleNet:多维度识别
4. ResNet:机器超越人类识别
5. DeepFace:结构化图片的特殊处理
6. U-Net:图片生成网络
7. 实例:剖析VGG,用模型进行模型参数可视化,特征提取,目标预测
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/e44b76361b038c500f063ddd90336704.png)
期待目标:
1. 掌握AlexNet结构特点,神经网络各层之间特征传导关系,模型参数总数计算
2. 了解VGG,GoogLeNet,ResNet等复杂ImageNet模型的结构特点,简单设计思想
3. 针对特殊数据,特殊任务设计的神经网络结构
4. 深度剖析VGG TF代码,学会对已有模型进行参数读取,目标预测,特征提取。
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/a33ebb6cd79702a30fac66d9a790e75a.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/05cf26dfd03b602186e125e282b3046e.png)
AlexNet:现代神经网络起源
背景介绍:
ImageNet Challenge:1000类物体,每类1000张图片
传统方法思路:
1. 图片特征提取
2. 机器学习分类
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/9faf79d58d0b05ce8adcdd18aad9c01c.png)
背景介绍:
2010年冠军
System Overview
Dense Grid descriptor:HOG,LBP
Coding:Local coordinate super-vector
Pooling, SPM
Linear SVM
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/4dcd85418cf2da9b4e9728e0943f750d.png)
2011年冠军:Xerox Lab
1. 特征提取
2. Fisher 压缩
3. SVM分类
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/67096d4bcfb5b10bb912a295685e725e.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/ec8de20c29c92e8707f3c6d83863a182.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/4be89b4e8d9048d0b55f0e0cad49e9a3.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/fb6deb238e8d588c6ff8fc437e30b5f3.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/3c28a6edc69ed662da4741ee0f31b008.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/8a0765e967feeeb7cf165602fa22c2b8.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/6dee36dba6fa965e1381cd0894b1722a.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/92c8ffd0539d71b4bef648b5aef15fdb.png)
VGG:AlexNet增强版
1. VGG-AlexNet 对比卷积层 - 卷积群参数个数:138m - 60m
2. 识别率(top5)7.3% - 15.3%
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/011fea889f90df36f07d6768c97da8a9.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/264fc14b867174aa5686c322f86c1f84.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/61ea3cd44744f019fee7938cee3a2d16.png)
VGG作用:
1. 结构简单:同AlexNet结构类似,均为卷积层,池化层,全连接层的组合。
2. 性能优异:同AlexNet提升明显,同GoogleNet,ResNet相比,表现接近。
3. 选择最多的基本模型:方便进行结构的优化,设计,SSD,RCNN,等其他任务的基本模型(base model)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/dd415ea3da32891e7e401d1dccc4f7ab.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/5a6ccc9289e1eb3f78428b73984dc9a2.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/aa5c863d1092115a1ea9bb71a47b69fe.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/3db512762a4fc49b2db6c5c31c7f877a.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/f69a9c821f33ae6538809f58b05f5533.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/9d09efacdc5f5ba35e3ae13f9924bf49.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/0b2c0f89be4604b5dd499053f0613829.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/8b12c2af52f0797c91935701777dfa1f.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/74a08a91532590deeec3f3b6163ad64b.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/8f8358e8cc6eacd3eeae3f3e48e58d5a.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/21d57057149ad5e888366bd0e96c8b99.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/fb4ae348645d2b80160d8951b572e747.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/11817e61b28231d90c8283f24035337a.png)
1. 为什么ResNet有效?
1. 前向计算:低层卷积网络高层卷经济网络信息融合;层数越深,模型的表现力越强。
2. 反向计算:导数传递更直接,越过模型,直达各层。
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/908129451eb49a51479c1a2fc61dadd5.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/ebbd6b620518c0dd192983b984cd1656.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/045ac7ba083aad66a72d453f48361214.png)
人脸识别数据特点:
结构化:所有人脸,组成相似,理论上能够实现对齐
差异化:相同位置,形貌不同
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/d3c054efc34e7f513dfe7386003318ba.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/1472bdb492a21c5cbef2ea29f93d4787.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/5cee452d42c1d34ba7c7025d419c2979.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/6af5841b61279a61c2243696ca38b546.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/c9340937c828de2eab30c21d2de55e18.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/7cb9dafd4600f46f437882cb25765412.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/73f9a1badf3bee04d68a72aecdc7cd30.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/7085dfecc205a69efcc59bf2d701e073.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/adeb0215f931f9174ea2bc48d336bbfb.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/6ff30faa0945fd9713ee22637c7becf8.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/5b036d24629242b432f633d950ac0bc8.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/fd8304e881ee021e585d074ffa984f67.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/6a45604b9f1f9d6fa18b635dcac63527.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/7f7575c8b9e27a36ce587040540089e8.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/8e6b857ff3b4bd990237d1d89aa4a610.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/751fae93df610820e038463a2d65787e.png)
![](https://oscdn.geek-share.com/Uploads/Images/Content/201706/07666e62cd92368652cfed2ef5bae31b.png)
相关文章推荐
- 【原创 深度学习与TensorFlow 动手实践系列 - 3】第三课:卷积神经网络 - 基础篇
- 【原创 深度学习与TensorFlow 动手实践系列 - 1】第一课:深度学习总体介绍
- 【原创 深度学习与TensorFlow 动手实践系列 - 2】第二课:传统神经网络
- 深度学习动手实践:用 TensorFlow 打造“会看”的机器人
- 深度学习动手实践:用 TensorFlow 打造“会看”的机器人
- 两个月入门深度学习,全靠动手实践!一位前端小哥的经验分享
- 深度学习与计算机视觉系列(9)_串一串神经网络之动手实现小例子
- 手把手教你搭建谷歌TensorFlow深度学习开发环境和更高级的keras开发安装!
- 深度学习笔记 (二) 在TensorFlow上训练一个多层卷积神经网络
- TensorFlow 深度学习笔记 卷积神经网络
- 【深度学习_4.1_2】在TensorFlow中应用卷积神经网络
- 深度学习算法实践10---卷积神经网络(CNN)原理
- 深度学习之卷积神经网络CNN及tensorflow代码实现示例
- 深度学习与计算机视觉系列(9)_串一串神经网络之动手实现小例子
- 深度学习与计算机视觉系列(9)_串一串神经网络之动手实现小例子
- 深度学习算法实践10---卷积神经网络(CNN)原理
- 深度学习之卷积神经网络CNN及tensorflow代码实现
- ArXiv最受欢迎开源深度学习框架榜单:TensorFlow第一,PyTorch第四
- TensorFlow深度学习笔记 循环神经网络实践
- 深度学习与计算机视觉系列,细说卷积神经网络