Pytoch Learning Series: Introduction

Reading guide

In the new year, we still need to get better , This official account has been silent for nearly two months before and after the festival , I haven't updated the original tweet for a long time , I feel I can't degenerate like this anymore . So I've been thinking about what to tweet ： It should meet the needs of the current work , Forced growth ; It should also be beneficial to the readers , Instead of simply marketing operations .

therefore , Finally decided “ Meddle in ” In depth learning direction , And I plan to update it first Pytorch Study the tutorial . Of course , This will be a series .

torch, Original intention “ torch ”

mention Pytorch We have to start with deep learning . All those engaged in data related positions know , Deep learning is a sub direction of machine learning , It is mainly based on neural network , Through flexible combination of a certain number of layers of network to achieve specific model functions , Especially good at computer vision （CV） And natural language processing （NLP） Direction . Its development history , After two highs and two lows , At present, it is in the third booming period of rapid development .

Net diagram , Invasion and deletion

here , The depth of in-depth learning is mainly reflected in the number of model layers , So called “ deep ”; But there is an important assumption hidden here , That is, the models constructed take neurons as the smallest unit of the network , So strictly speaking, it should be called deep learning based on neural network . natural , It can also not be based on neural network , For example, zhouzhihua team explored and proposed a deep random forest model a few years ago , It can be said to be a new research idea of deep learning .

From theoretical research to industrial application , Among them, there must be mature industrial realization . With python Language is the foundation , For the classical machine learning model , Then everyone must know scikit-learn; And when it comes to deep learning , The corresponding tool kit is less “ Centralization and unification ”, It can even be called the place of big factory disputes . among , The most representative and widely used is TensorFlow and Pytorch, The former comes from google, The latter comes from Facebook; The former is mainly used in industry , The latter is popular in academia . Of course , In terms of academia or industry , There is no clear boundary between the two .

At first , In learning TensorFlow Widely used in industry , And I have long been far away from colleges and universities, so I went directly into the pit TensorFlow, For a while TF boy, Especially knowing that TensorFlow2.0 Overcome the early 1.0 The version has been criticized for its static graph problem , So I don't think TensorFlow What are the disadvantages . But later , With the deepening of learning , In addition, through the understanding of surrounding colleagues , Find out Pytorch It has more excellent characteristics ： For example, with Numpy The design of is closer to , The grammar style is more Pythonic wait . therefore , The individual is also decisive to turn to Pytorch camp .

This paper is the first one , Just to introduce Pytorch Can do , And personal understanding of why it is so designed .

Torch It is an old-fashioned deep learning framework , It was based on lua Language development , Because of the minority of its development language , So its development and application are also subject to many restrictions . since Facebook Open source Python Ecosphere Torch tool kit ——Pytroch after , It has always been a match TensorFlow A heavyweight tool for . at present Pytroch stay GitHub Get on 54k star（TensorFlow Currently in GitGHub Get on 163k star, The gap is still large , There are about 3 Twice as many ）.

It is precisely because the most extensive stage of in-depth learning lies in the application direction of image, voice and text , So with Pytorch Three supporting toolkits and a model service toolkit ：

• torchvision
• torchtext
• torchaudio

Of course ,Pytorch It is still the foundation and core

As a deep learning toolkit ,Pytorch What can I do with it ？ The official documents are quoted here to describe its location , Broadly speaking, it has two functions ：

namely ：

• Support GPU Accelerated Tensor Ability to calculate
• Deep neural network construction supporting automatic derivation

So here comes the question ： All say Pytorch It is a deep learning tool , Why is its core function designed as the above two points ？ Regarding this , Personal understanding is as follows ：

firstly ：Tensor It is the foundation of deep learning model construction and training , Its status is like array To Numpy、DataFrame To Pandas, It is itself a data structure , But it constitutes Pytorch The soul of . here ,Tensor The original meaning of English is “ tensor ”, In fact, it corresponds to a multidimensional array , Essentially follow numpy Of ndarray It's consistent .

From this point of view ,Pytorch It can be seen as numpy Upgraded version , The upgrade here is mainly reflected in the availability of GPU Powerful parallel computing power . If there is Numpy Basics , Study Pytorch It can be very simple ; On the other hand , learn Pytorch It can also be regarded as right Numpy A supplement to , Regardless of the purpose of building a deep learning model .

second ：Pytorch Positioned as a deep learning tool , Its main function is to support the construction and training of deep learning model . meanwhile , Different from the mature models in classical machine learning , Most deep learning networks have no fixed model or paradigm , Generally, it is composed of multiple basic modules flexibly matched by users （ Of course , In fact, there are some mature models , for example LeNet-5、AlexNet and VGGNet etc. , But more generally, users still need to customize themselves ）, therefore Pytorch The support for deep learning does not lie in how many mature models are integrated , Instead, it provides a basic deep learning module , These are like scaffolding , It can be combined and matched arbitrarily , So as to realize more free and customized functions .

Pytorch The functions are rich and complicated , The best learning platform is to consult its official documents ,https://pytorch.org/. From the Pytroch The universality of the group , At present, its documents support multiple languages , Including Chinese documents , This also provides more channels for self scholars to get started quickly . I personally benefited from it , Subsequent tweets will also use this as an important reference framework .

So much for this tweet , Benchmarking goes from tool introduction to model modeling , The following articles will be updated every week Pytorch Learn a series of tweets .