【网课笔记】Fast.ai-2019 || Week3 - Multi-label Classification & Segmentation

注： 此博客仅为博主复习之用，省去了大量细节和博主已在其他地方做记录的知识点；有兴趣学习fast.ai课程的同学建议扎实学习Howard的视频以及Forum中的资源！共勉！

---

2019/6/28

课前热身

介绍了Andrew Ng的ML课程：内容部分虽然稍微过时，但大部分还是非常excellent，其采用自底向上模式，结合fast.ai自顶向下模式，可以更好地学习！
combine bottom-up style and top-down style and meet somewhere in the middle.
当然Howard也介绍了自己的ML课程，约为DL课程的两倍时长；（涉及很多基础知识）

学习建议：All these courses together if you want really dig deeply into the material,DO ALL OF THEM. A lot of people who have had end up saying: oh I got more out of each one by doing the whole lot。

Production

Howard提供了部署web app的简单方案！前端javascript，后端数据JSON；链接在课程导航页左侧的production菜单中；

show works!

值得去Forum一一回看，会给你很多启发！！！ （不一定视频，有GitHub详细笔记！）

• What car!
  
• 识别视频中的人物表情 or 手势
  查找Forum和资料，弄懂如何实现图片到视频的跨越
• yourcityfrom
  通过卫星图像识别出是哪个城市！
• 识别面部表情的类别–结合2013年一篇论文

| Multi-label prediction

dataset: Planet Amazon dataset (satellite images), from Kaggle
很多人已经在深度学习中使用satellite images, but only scratching the surface!

Get dataset from Kaggle

按照notebook教程做即可

Create DataBunch - use data block API

DataBunch(train_dl: DataLoader(datasets(), batchsize = 8), valid_dl: 同) ： 即一层一层作为参数传入！

注： fastai 所有文档 都是notebook，可以clone下来在Jupyter上练习！具体怎么打开可以回看lesson3-29`

----------------up to 56` min

---

| Image Segmentation with CamVid

注意点：调整图片时，mask也要相应调整，否则不匹配了！

训练方法： progressive resizing

从size较少的数据集先训练（更快更容易），然后transfer learning，逐步增大图片的size； 但是目前还没有准确的理论指出每个level的大小应该设为多少，Howard经验：
低于64x64往往没有帮助

fit_one_cycle()原理

简单来说，就是让lr_rate先变大变小！

Loss的形态类似于下图，
变大的目的：helping the model to explore the whole function surface and try and find areas where both the loss is low and also it`s not bumpy! 防止困在某个局部最优处！（所以传入fit_one_cycle()的其实是max learning rate）

所以如果用fit_one_cycle()，Loss的变化会如下，也就是说，如果你发现:
if you find that it’s just getting a little bit worse and then it gets a lot better, you’ve found a really good maximum learning rate.

那如果你发现：Loss一直下降，可以稍微上调一下学习率
so if you find that Loss is kind of always going down, particularly after you unfreeze, that suggests you can probably bump your learning rate up a little bit.
Because you really want to see this kind of shape! It’s going to train faster and generalize better.

Howard： 知道这个理论和真正运用的差距，你需要看大量这种情况的变化图！
所以每次训练完，看Loss的变化情况，good results怎样，bad results怎样，不但调整learning rate 和epochs，观察图像变化！

Mixed precision training

如果训练时遇到内存超限的问题，可以尝试使用16bit精度的浮点数进行训练！
means: Instead of using single-precision floating-point numbers, you can do most of the calculations in your model with half-precision floating-point numbers.
So, 16 bits instead of 32 bits. 因为比较新，所以你可能需要最新的硬件(CUDA drivers, etc)才行，fastai提供了接口，只需在create_learner的末尾加上to_fp16()即可

----------------up to 94` min

---

| Regression with BIWI head pose dataset

【Image Regression Model: find the center of the face，two float numbers】
Regression: any kind of model where your output is some continuous number or set of numbers.

----------------up to 101` min

---

| NLP quick version - IMDB

第一个学生提问： 没太听懂，回看！

----------------up to 110` min

最后提到了Michael Nielsen的书 《Neural Network and Deep Learning》 中的动画：
If you have enough little matrix multiplications followed by sigmoid(ReLU), you can create arbitrary shapes. Combinations of linear functions and nonlinearities can create arbitrary shapes

Universal approximation theory: If you have stacks of linear functions and nonlinearities, you can approximate any function arbitrarily, closely.

之前这本书看到一半扔下来，看来要捡起来才行！里面有很多细节值得一看！

---

Question： 如果图片数据是2 channel或者4channel的，如何使用3channel的 pretrained model?

Howard表示之后会在fastai中加入此功能；
2 channel的数据: there is few things you can do, but basically you can create a third channel as either being all zeros or being the average of the other two other channels.
4 channel的数据: 当然不能丢弃 4th channel的数据， 所以只能修改你的model，简单来说就是给weight tensor多加一维，赋值为0或随机值；后面课程会细讲！