Tensorflow team: we haven't been abandoned

Fish and sheep From the Aofei temple
qubits | official account QbitAI

finally , Google responded “TensorFlow Abandoned ” Hearsay ：

We will continue to be committed to TensorFlow Build a first-class machine learning platform , And JAX Promote machine learning research side by side .

During this period ,“JAX replace TensorFlow” This topic is a hot topic .

Not only Business Insider And other media quoted insiders , call “ Google brain and DeepMind Has generally given up TensorFlow, Switch to JAX”.

even LeCun、fast.ai founder Jeremy Horward Wait for the big guys to come out and watch , Indicates that the competition between deep learning frameworks has entered a new stage .

stay TensorFlow Because of growing bloated 、 High learning costs , More and more criticized by the industry ,“ Google abandons TensorFlow” There seems to be nothing wrong .

however , After a period of public opinion fermentation , Now? ,TensorFlow The official team finally came forward “ rumours ”, Send a message to say ：

We will continue to invest in TensorFlow and JAX Two ML frame , To promote the research and application of millions of users .

Details , See together TensorFlow Blog wrote something .

official ：TensorFlow It is the answer we give to the engineer

TensorFlow The official quoted Stack Overflow The latest statistics of .

data display ,TensorFlow In the field of machine learning as a whole, it is still better than PyTorch Be well received .

in addition ,TensorFlow Currently, the number of downloads per month exceeds 1800 Ten thousand times , And in GitHub Owned on 16.6 Wanbiao star .

by comparison ,PyTorch The number of stars is 5.7 ten thousand ,JAX The number of stars is 1.91 ten thousand .

The official team emphasized , Inside Google ,TensorFlow Almost all AI Support production workflow , Include search 、 advertisement 、YouTube、 mailbox 、 Map 、Play、 Maps, etc .

And apple 、 Net flight 、 tencent 、Uber And other well-known enterprises , Also in use TensorFlow Support their production system .

As for why Google now attaches importance to JAX, The reason is ：

In recent years , We found that , A single common framework is often not applicable to all scenarios —— In particular, the needs of production and cutting-edge research often conflict .

therefore , We created JAX, A simple for distributed numerical computation API.

JAX Has a very excellent performance in the frontier research field ： Reached a new parallel scale , Advanced new algorithms and frameworks , New compilers and systems have also been developed .

Officials also give examples ,AlphaFold and Imagen It has been verified JAX The value of .

by comparison ,TensorFlow It is “ Our answer to the needs of application machine learning developers ”.

in other words ,TensorFlow It will focus more on meeting the requirements of engineers on any scale and any platform , Build and deploy reliable 、 Stable 、 Demand for high-performance machine learning systems .

Besides , Google is officially in JAX and TensorFlow A lot of work has also been done in connectivity .

Such as through jax2tf, Researchers can put JAX The model is placed in TensorFlow Put into production .

Last , The official mentioned ：

Looking forward to the future , We will continue to be committed to TensorFlow Build a first-class machine learning platform , And JAX Promote machine learning research side by side .

We will continue to invest resources in developing these two machine learning frameworks , To promote the research and application of millions of users .

Also hit a wave of job advertisements （ Manual formation ）.

About JAX

2018 year ,JAX Born in the hand of a three person team of Google brain .

At the beginning of the birth of , It aims at high-performance numerical computing and machine learning research .

JAX Can be seen as support GPU、TPU Wait for the accelerator to accelerate 、 Supporting automatic differentiation Numpy variant .

Simply speaking , It can be understood in this way ： When you want to deal with some problems that require high computational power , adopt JAX, You can quickly spread complex problems to multiple TPU On .

at present , Google brain Trax、Flax、Jax-md, as well as DeepMind The neural network library of Haiku And reinforcement learning library RLax etc. , It's all based on JAX Built .

It is worth mentioning that ,JAX When it was born , It is PyTorch Strong impact in academia TensorFlow Time of day .

A research institution of the University of California, Berkeley RISELab Data show that ：

2019 year 1 Month to 6 End of month , stay arXiv Among the papers published in , mention TensorFlow and PyTorch The number is almost the same ,PyTorch Only slightly behind .

But in terms of growth rate , And 2018 year 1-6 Monthly comparison ,PyTorch Of “ share ” increased 194%.TensorFlow Only increased 23%.

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