Google refutes rumors of abandoning TensorFlow, it’s still alive!

Author | Su Mi

Produced | CSDN (ID: CSDNnews)

I don’t know if the rumors of “TensorFlow will die” have spread too much. Google has urgently released an article titled “Bringing Machine Learning to every developer’s toolbox” (bringing machine learning into every developer’s toolbox) The announcement, widely advertised, TensorFlow is not “dead”, and various data show that it is now developing very well, and it is also the most commonly used ML tool for 3 million software developers around the world.

At the same time, Google has not given up on continuing to develop TensorFlow, which will be with JAX in the future.

TensorFlow being attacked left and right

TensorFlow is an end-to-end machine learning platform open sourced by Google in 2015. Once released, companies in the industry such as Apple, Netflix, Stripe, Tencent, Uber, LinkedIn, Twitter, Baidu, Orange, LVMH, etc. often use it to optimize operations, and it is also used in deep neural network training and inference.

Over time, TensorFlow’s strongest competitor, PyTorch, emerged in 2016, bringing faster prototyping than TensorFlow. Additionally, PyTorch is more tightly integrated with the Python ecosystem than TensorFlow, and the debugging experience is much simpler.

Source: The Gradient

Over the years, the competition between TensorFlow and PyTorch can be said to have reached a more anxious state.

However, when the time came to 2020, a decision by Google poured cold water on TensorFlow, which was already in a somewhat passive situation of competing with PyTorch. DeepMind, which also belongs to the same parent company Alphabet, said to the public that they were using JAX. to accelerate AI/ML research.

According to the official introduction of JAX on GitHub:

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

With the updated version of Autograd, JAX can perform automatic differentiation of Python programs and NumPy operations, and supports loop, branch, recursion, and closure function derivation. Additionally, JAX can use the XLA implementation to compile and run NumPy programs on GPUs and TPUs. By default, compilation is JIT-compiled and performed as a system call. But JAX also allows developers to compile their own Python functions into XLA-optimized kernels using a single-function API jit.

JAX seems to be gaining momentum today more rapidly than when TensorFlow was first introduced. As of now, it has obtained 19k Stars on GitHub.

So it’s hard not to worry about the future of TensorFlow.

Even the Turing Award winner and father of CNN, Yann LeCun, once commented, “The fierce competition between deep learning frameworks has entered a new stage. Now Google’s TensorFlow has lost to Meta’s PyTorch, and Google is also turning to JAX internally. .”

So, is Google really replacing the well-known TensorFlow with the rising star JAX?

TensorFlow is still the most commonly used framework for developers

For now, Google has given an affirmative and affirmative answer: no .

In the latest announcement, Google used a lot of data to illustrate the current state of TensorFlow.

As data from previous Stack Overflow developer surveys show, TensorFlow is the most commonly used ML tool by developers, used by 3 million software developers worldwide to enhance their products and solutions. At the same time, TensorFlow is also the framework that many developers want to use in the future. It is expected that in the near future, the user base of TensorFlow will reach 4 million.

Most Wanted Frameworks and Libraries

TensorFlow is now downloaded more than 18 million times a month and has accumulated 166k Stars on GitHub, which is also the leader in all current deep learning frameworks.

Google says that within its company, TensorFlow supports almost all AI R&D workflows, including search, advertising, YouTube, GMail, maps, playback, photos, and more. Meanwhile, every month, Google Scholar indexes more than 3,000 new scientific publications that mention TensorFlow or Keras, including important applied sciences such as CANDLE (Cancer Distributed Learning Environment Framework) research on understanding cancer .

One framework is not enough

As for why it developed a new framework, Google also explained that ” a single general framework cannot be suitable for all situations – in particular, the needs of “real-world production environment” and “top-of-the-line scientific research” often conflict . “

Therefore, it has launched a minimalist API for distributed numerical computing, JAX, to power scientific computing research in the next era.

In a nutshell, TensorFlow and JAX have different audiences, with the former oriented towards the ML developer community and the latter mainly towards researchers.

Google said, “In this new multi-framework world, TensorFlow is our answer to the needs of applied ML developers – engineers who need to build and deploy reliable, stable, high-performance ML systems at any scale and on any platform. We Our vision is to create a cohesive ecosystem where researchers and engineers can leverage components that work together, no matter which framework they originate from. We’ve made great strides in JAX and TensorFlow interoperability, especially through jax2tf. Researchers developing JAX models will be able to put them into production through the tools of the TensorFlow platform.”

Looking forward to the future of TensorFlow, Google further stated its attitude, ” We intend to continue to develop TensorFlow as a first-class platform for applying ML, side by side with JAX, to promote the scope of ML research. We will continue to invest in both ML frameworks to drive research and applications for millions of users. “

The domestic deep learning framework breaks through the “monopoly” opportunity

TensorFlow has not given up, PyTorch is still leading, JAX is catching up, and the field of deep learning frameworks has become more and more lively. Looking at the country, there are also many deep learning frameworks rising rapidly.

As of December 2021, Baidu’s “Flying Paddle” deep learning platform (ie PaddlePaddle) has broken through the monopoly of Google and Facebook (Meta) in the Chinese market in the past, and has become the largest deep learning platform in China in terms of comprehensive market share.

Global AI open source framework star number TOP2 in April and May, source: OSS Insight data

At the same time, domestic frameworks such as MindSpore, OneFlow, MegEngine, and Jittor are also infiltrating applications in various fields.

Regarding this trend, Yuan Jinhui, the founder of the first-class technology OneFlow, also commented in the circle of friends not long ago, “I originally thought that Google gave up Tensorflow, because I haven’t seen a significant update for a long time, but recently I see that dtensor is being introduced and the runtime is being refactored. , which shows that TF has not been abandoned, it should be a strategy of walking on two legs. Recently, we have seen that the domestic framework has also made rapid progress, and there is still a lot of opportunities to focus on innovation and localization strategies.”

However, according to the data survey, the star number gap between TensorFlow and PaddlePaddle is close to 10:1. At the same time, the number of Commits between TensorFlow and PaddlePaddle is nearly three times the gap between China and the United States. Although foreign countries have a certain first-mover advantage in terms of AI development, the lack of open-source framework Stars and Commits can still reflect some of the problems existing in China’s open-source ecosystem.

The report of the Prospective Industry Research Institute pointed out that the development of AI in China is more inclined to the application layer. In this regard, Jiang Tao, founder and chairman of CSDN and founding partner of Geekbang Venture Capital, said that the AI ​​open source model of “emphasizing application and ignoring ecology” is not a long-term solution . “The current open source ecology still has problems, and even forms a situation of “separate management”, which will lead to internal consumption, increase the cost of user selection, and the difficulty of technology reuse, hindering the large-scale development of the entire industry. Therefore, the construction of open source ecology It is crucial for China’s development.” And this is also

The text and pictures in this article are from CSDN