Google refutes rumors and gives up tensorflow. It's still alive!

author | Su mi

Produce | CSDN（ID：CSDNnews）

I don't know if it was before “TensorFlow Will die ” The rumor of has spread too far ,Google Recently, I released an urgent article entitled 《Bringing Machine Learning to every developer’s toolbox》（ Bring machine learning into every developer's toolbox ） Announcement of , Spread the word ,TensorFlow No, “ die ”, And all kinds of data show that , It is now developing very well , It's also global 300 The most commonly used by software developers ML Tools .

meanwhile ,Google  And I didn't give up on continuing development TensorFlow, In the future, it will work with JAX amen .

Caught between the left and the right TensorFlow

TensorFlow, yes Google stay 2015 Open source end-to-end machine learning platform in . Once released , The industry is like apple 、Netflix、Stripe、 tencent 、Uber、LinkedIn、Twitter、 Baidu 、Orange、LVMH And other companies often use it to optimize their operations , It is also applied to the training and reasoning of deep neural networks .

Iterating over time ,TensorFlow The strongest competitor PyTorch stay 2016 In the , It brings more than TensorFlow Faster prototyping . Besides ,PyTorch And Python Ecosystem integration ratio TensorFlow More closely , The debugging experience is also much simpler .

source ：The Gradient

For many years between ,TensorFlow And PyTorch It can be said that the competition has reached a more anxious state .

However , When time comes 2020 year ,Google A decision of , Give Ben and PyTorch The competition is slightly passive TensorFlow Poured a wave of cold water , It belongs to the same parent company Alphabet Flag DeepMind To express to the outside , They are using JAX To speed up AI/ML Research .

according to JAX stay GitHub The official introduction on shows ：

JAX By Google Brain A new framework for high-performance machine learning research launched by the team , Its predecessor is Autograd and XLA The combination of .

adopt Autograd Updated version of ,JAX It can be done to Python Procedure and NumPy The operation performs automatic differentiation , And it supports circulation 、 Branch 、 recursive 、 Derivation of closure function . in addition ,JAX You can also use XLA Realize in GPU and TPU Compile and run on NumPy Program . By default , The compilation is done according to the system call JIT Compile and execute . however JAX It also allows developers to use a single function API jit To his own Python The function compiles to XLA Optimized kernel .

Today, JAX The momentum of development seems to be stronger than before TensorFlow When it was first launched, it was even faster . Up to now , Its presence GitHub You have obtained 19k individual Star.

therefore , It's hard not to get people to TensorFlow Worry about the future .

Even Turing Prize Winners 、CNN The father of Yann LeCun Also commented that ,“ The fierce competition between deep learning frameworks has entered a new stage . Now? Google Of TensorFlow Has lost to Meta Of PyTorch,Google The interior is also turning JAX.”

that ,Google Is it really a rising star JAX Will be known to all TensorFlow Instead, ？

TensorFlow It is still the most commonly used framework for developers

For now ,Google Gave a sure and affirmative answer ： Can't .

In the latest announcement ,Google Use a lot of data to illustrate TensorFlow The current status of .

Just like before Stack Overflow Developer research The data results show that ,TensorFlow Is the most commonly used by developers ML Tools , By the world 300 Million software developers use , To enhance their products and solutions . meanwhile ,TensorFlow It is also the framework that many developers want to use most in the future , It is expected in the near future ,TensorFlow The user base will reach 400 ten thousand .

The framework and library you want to try most

TensorFlow Now download more than every month 1800 Ten thousand times , And in GitHub On the accumulation of 166k Pieces Star, This data is also in the leading position in all current deep learning frameworks .

Google call , Within the company ,TensorFlow Almost all of it AI R & D workflow , Include search 、 advertisement 、YouTube、GMail、 Map 、 Play 、 Photos and so on . meanwhile , Every month ,Google Scholar Will be more than 3,000 The article mentions TensorFlow or Keras Index new scientific publications , This includes important applied sciences , Such as CANDLE（ Cancer distributed learning environment framework ） Research on understanding cancer .

One framework is not enough to go around the world

Why should we develop a new framework ,Google Also explain ,「 A single common framework cannot be applied to all situations —— especially ,“ Realistic production environment ” and “ Top scientific research ” The needs of often conflict .」

so , It introduces a minimalism for distributed numerical computation API——JAX, To provide power for scientific computing research in the next era .

Simple view ,TensorFlow and JAX The audience has different tendencies , The former tends to ML Developer community , The latter is mainly for researchers .

Google Express ,“ In this new multi frame world ,TensorFlow It's our application ML Response to developer needs —— On any scale and platform , Engineers need to build and deploy reliable 、 Stable 、 High performance ML System . Our vision is to create a cohesive ecosystem , Researchers and engineers can take advantage of components that work together , No matter which framework they originated from . We've been JAX and TensorFlow Significant progress has been made in interoperability , Especially through jax2tf. Development JAX The researchers of the model will be able to pass TensorFlow Platform tools put them into production .” 

expectation TensorFlow The future of ,Google Further shows the attitude ,「 We intend to continue to develop TensorFlow, As an application ML First class platform , And JAX side by side , Push ML The scope of the study . We will continue to invest in these two ML frame , To promote the research and application of millions of users .」

Domestic in-depth learning framework breaks through “ monopoly ” The opportunity

TensorFlow Never give up ,PyTorch Still lead ,JAX do all one can to catch up , The field of deep learning framework has become more and more lively . And look at home , There are also many in-depth learning frameworks emerging rapidly .

By 2021 year 12 month , Baidu “ Flying propeller ” Deep learning platform （ namely  PaddlePaddle）, Has broken through the past in the Chinese market Google、Facebook（Meta） The monopoly situation of , Become the first comprehensive market share of China's deep learning platform .

The global AI Open source framework Star Count 4 month 、5 Monthly list TOP2, source ：OSS Insight data

meanwhile , domestic MindSpore、OneFlow、MegEngine、Jittor And other frameworks are also being applied in various fields .

For this trend , First class technology  OneFlow Founder yuanjinhui also commented in the circle of friends not long ago ,“ I thought Google  To give up Tensorflow, Because I haven't seen any significant updates for a long time , But recently it seems that dtensor, And refactoring runtime, This explanation TF Not given up , It should be a two legged strategy . Recently I saw , The domestic framework has also made rapid progress , Focus on innovation and localization strategies , There is still a good chance .”

however , According to the data survey ,TensorFlow And PaddlePaddle Of Star The number difference is close to 10：1. meanwhile ,TensorFlow And PaddlePaddle Of Commits Number , The gap between China and the United States is close 3 times . Although from AI In the course of development , Foreign countries have certain first mover advantages , But open source frameworks Star Sum of numbers  Commits The lack of numbers can still reflect some problems in China's open source ecosystem .

The report of prospective industry research institute points out that , China AI The development of is more inclined to the application layer . Regarding this ,CSDN founder & Chairman of the board of directors 、 Jiang Tao, founding partner of geek group venture capital, said ,“ Re apply , Light ecology ” Of AI The open source model is not a long-term solution .「 There are still problems in the current open source ecosystem , Even formed a kind of “ act of one 's own free will ” The situation of , This leads to internal consumption , Increase user selection costs , And the difficulty of technology reuse , Hinder the large-scale development of the whole industry . therefore , Open source ecological construction is very important for China's development .」 And this is also “ Overtaking in curve ” The best breakthrough .

Reference link ：

https://blog.tensorflow.org/2022/06/%20bringing-machine-learning-to-every-developers-toolbox.html?m=1

https://github.com/google/jax

— Recommended reading  —

Twenty years ago ,《 The programmer 》 When it was first published , We should pay full attention to the growth of software people . today , Our original intention remains the same ： Behind the lines of code , It's a living developer's ambition to change the world . therefore ,《 New programmers 004》 from C++ The father of Bjarne Stroustrup、C# The father of Anders Hejlsberg、MySQL The father of Michael "Monty" Widenius、PostgreSQL Co founder of global development group Bruce Momjian Wait for the programmer's grandfather , Go to jiayangqing, vice president of Alibaba 、 Pan Aimin, founder and chairman of instruction set 、Vue.js By you Yuxi ……48 It's a big tech guy , Create our programming life together 、 Our technological age .《 New programmers 004》 Fully listed , Welcome to subscribe to ！