Column directory:
This article +
pytorch quick introduction and actual combat - 1. Knowledge preparation (introduction to elements)
pytorch quick introduction and actual combat - 2. Deep learning classic network development
pytorch quick introduction and actual combat - 3. Unet realizes
pytorch quick introduction and actual combat ——4. Network training and testing
Notice：
The tutorial modules are highly independent, and can be read anywhere, regardless of whether it is between different articles or different modules of the article.
How to be happy. Anyway, it's all from " what's it all " to " hehe, this "
part is not listed in detail because it can be skipped, even if you don't understand it.
This article targets the me who was two months ago, almost pure and new. UNet does image processing
All the examples in this article are the simplest examples, don't be limited to this, its divergence is almost infinite.
My humble servant: I am new to deep learning, majoring in computer science, second year of graduate school, not good at mathematics, good at coding, with tasks in hand, eager to apply deep learning, playful in class (if I am not playful, I will not learn much).
Problem-Solving Ideas: What is it? Why? what to do? Learning concept: almost crazy pursuit of efficiency, so choose emergency learning.
Emergency learning: Search and solve problems after encountering them, that is, purpose-driven learning.
The introductory concept carding line is based on:
Don't bother with python (understand the concept)
Xiaoqiang DL (read papers to see network development)
Deep Learning Quick Start
There are too many ways to get started with deep learning, starting with basic concepts and historical development.
In short, most of them are systematic learning, but systematic learning is wide enough for those who do not need to apply deep learning in the short term, but it is too time-consuming for those who are on task, and the sense of accomplishment is low and easy. bored, tired. Especially in the face of too many jerky concepts, it is easy to create a sense of fear.
But this does not mean that I do not pay attention to a solid foundation, the importance of the cornerstone is self-evident, but in the face of reality, I choose to be lazy and take shortcuts.
After two weeks of fighting, I finally let the code run and understand some basic debugging (pure manual code is not a big problem, but of course it does not include improving the network and loss functions.) I faced Deep Learning two weeks ago
. When I was with pytorch, I was also confused, as if there was a big mountain in front of me, and now it is also a corner of his mysterious veil.
1 What deep learning can do
First of all, it is necessary to understand what deep learning can do:Classification、detection、Split
ClassificationNot much to say, I have read all the watermelon books, that is:
judging whether a melon is good or bad according to its color, roots and other characteristics.
Target Detection
Circle the car, person, etc. in the picture.
semantic segmentation
Note that it is not limited to NLP (Natural Language Understanding). For example, in a picture, whether a certain pixel belongs to the background is also semantics.
But don’t be limited to this. I don’t understand many problems, but they can be used if they diverge. For example, I use deep learning instead of CT section reconstruction.
2 Three elements of neural network
datainput tonetworkoutputresult, compare the label, and calculate the loss backpropagation correction parameters. Go to the next round of training.
2.1 Data
You can download minist handwriting training, and use your own if you have it.
Or if you try it out, you can randomly use some two-dimensional matrices as grayscale images.
2.2 Network
Just find a baseline, you can change it or not, you can find the source code on github, or you can find a paper and reproduce it yourself through the network structure diagram. In fact, it is not that difficult.
2.3 Result return
Torch has its own loss function, L1 regular? MSE mean square loss function? Anyway, to get started, just define one and it’s OK.
Anyway, this is the case, so let’s understand some concepts from these three elements:
pytorch quick introduction and actual combat - 1. Knowledge preparation (introduction to elements)