Heart of the Machine reports

Participation: Zhang Qian, Du Wei

A few days ago, OpenAI announced a “full shift to PyTorch” via their blog, with plans to unify all frameworks for their platform into PyPyTorch. This news has reignited community debate about the merits of the two frameworks. As an upstart, has PyTorch really caught up with TensorFlow? To investigate this question, data scientist Jeff Hale examined the status of the two frameworks in terms of number of online jobs, number of mentions in top papers, online search results, and developer usage.

By comparing these four areas, the authors conclude that TensorFlow is still leading in most areas, but PyTorch is making progress and closing the gap.

This is not the first time the author has investigated deep learning frameworks. From 2018 to 2020, he conducted three surveys and published reports. From these reports, we can see the hot history of deep learning frameworks in multiple dimensions.

2018: TensorFlow crushes PyTorch

Jeff Hale’s first study was published in September 2018. In that survey, he found that TensorFlow was the absolute champion. GitHub activity, Google searches, Medium articles, Amazon books, and arXiv papers are among the largest. TensorFlow also has the largest number of developer users and the largest number of online job descriptions.

Deep learning framework power score ranking from 2018 survey. \

By contrast, PyTorch came in third, scoring well below Keras, who was in second place at the time.

Weighted scores for each framework in the 2018 survey results. As you can see, TensorFlow is the absolute champion in most metrics, while PyTorch is mostly in third place.

2019: PyTorch is firing on all cylinders, TensorFlow is experiencing sluggish growth

In April 2019, Jeff Hale published a second survey. This time, he looked at the growth of several frameworks over the past six months (the time elapsed between this survey and the last one). TensorFlow was still the most in demand and the fastest growing framework at the time, but PyTorch has also grown faster than Keras in the past six months.

Deep learning framework growth rankings from 2019 research.

It is worth noting that PyTorch was particularly impressive in terms of job growth at the time, and was very close to TensorFlow. Moreover, PyTorch added more jobs than TensorFlow on all job sites except linkedin.

TensorFlow was already experiencing negative growth in terms of the relative number of Google searches. This represents a decrease in the number of searches relative to TensorFlow and an increase in the number of searches relative to PyTorch over the past six months.

2020: PyTorch Top will take the lead, catching up to TensorFlow in the workplace

Fast forward to 2020, and the framework battle is down to PyTorch and TensorFlow. Therefore, this time, the author put all his energy on these two frames.

At the time of this survey, the two frameworks had become more and more similar, that is, there was a “convergence” trend. Both can now run in either dynamic eager Execution mode or static graph mode.

As of now, PyTorch has been updated to 1.4, adding new features to cater to the industry and making it easier to run on Google Cloud TPU. PyTorch’s community is also growing. In addition to recent OpenAI, Preferred Networks (PFN), the keeper of the deep learning open source framework Chainer, announced late last year that the team would no longer be making major updates to Chainer, Future research will focus on PyTorch.

TensorFlow 2.0 also introduces a number of new improvements that make the API more streamlined and brain-friendly. In addition, TensorFlow tightly integrates Keras as its front-end and advanced APIS.

TensorFlow still has more functionality in product and edge device deep learning than PyTorch does, but PyTorch’s functionality is getting better.

In this context, the survey looked at four metrics: number of online jobs, number of mentions in top papers, online search results, and developer usage.

Number of online jobs

On January 26, 2020, the author searched for TensorFlow and PyTorch on Indeed, Monster, SimplyHired, and LinkedIn.

As the search results below show, TensorFlow appears about twice as often on each job site as PyTorch.

The results are shown in percentage terms as follows:

Ten months ago, TensorFlow was listed three times as often as PyTorch in the 2019 survey, and now the gap has narrowed to twice that.

Top will appear in the paper

Currently, PyTorch appears most frequently in top conference papers. As shown below, Horace He, a former PyTorch intern, made a chart showing the proportion of PyTorch occurrences in TensorFlow/PyTorch occurrences in papers from the top conferences (2017-2019) : A ratio of more than 50% means that PyTorch appears more often than TensorFlow in this summit paper. As you can see from the chart, PyTorch appears more frequently at conferences than TensorFlow.

Photo source: chillee. Making. IO/pytorch – v -…

The diagram below shows a visual comparison and trend development between PyTorch and TensorFlow in NeurIPS conference papers (2016-2019), with solid lines representing PyTorch and dotted lines representing TensorFlow: In the 2019 NeurIPS conference papers, PyTorch appeared 166 times and TensorFlow 74 times, twice as many. In the 2018 NeurIPS conference paper, PyTorch appeared less frequently than TensorFlow. These can be seen the development trend of the two frameworks in recent years.

Photo source:Chillee. Making. IO/pytorch – v -…

Online Search results

The authors used Google Trends to find the relative volume of searches for PyTorch (software) and TensorFlow (computer application) between January 26, 2017 and January 26, 2020.

Below is a linear trend line for Google Search Results, with TensorFlow in blue and PyTorch in red. As you can see from the figure, the sex pair search volume of TensorFlow decreases while that of PyTorch increases, and the gap between the two is getting smaller and smaller.

Photo source:Public.tableau.com/profile/jef…

Developer Usage

In a Stack Overflow developer survey in early 2019, 10% of respondents used TensorFlow and 3.3% used Torch/PyTorch. The same is true for professional developers (9.4%vs 2.9%). But given that the data are from early 2019, the true picture may now be different.

Overall, TensorFlow still has more online jobs than PyTorch, but the gap is closing; PyTorch continues to lead in terms of the number of occurrences in top papers and has further narrowed the gap with TensorFlow in Google search results; In a recent Stack Overflow developer survey, TensorFlow was still three times as popular as PyTorch.

Finally, the author says that his research on both TensorFlow and PyTorch deep learning frameworks will continue, but it is still unclear which framework will be the best choice for the next two years. But he believes TensorFlow is a safer option.

Finally, the author provides the want to learn the depth study of readers some useful resources: such as course. Fast. Ai website: course. Fast. Ai/index. HTML.

In addition, The Heart of Machine has recently compiled a list of excellent tutorials from the past few years for you to collect and learn from (see “Collect and Quit in one Go, The Heart of Machine 2019 is here”).

Related reading:

The most popular Deep Learning Frameworks of 2018?

2019 Findings: Is TensorFlow being unseated in 2019?

Reference links:

Towardsdatascience.com/is-pytorch-…

Note: the menu of the official account includes an AI cheat sheet, which is very suitable for learning on the commute.

Machine Learning Online Manual Deep Learning online Manual AI Basics download (PDF updated to25Note: to join this site's wechat group or QQ group, please reply to "add group" to get a discount station knowledge planet coupon, please reply to "knowledge planet"Copy the code

Like articles, click Looking at the