Comparison of the influence of 11 deep learning frameworks

Editor's note: Jeff Hale assessed the impact of 11 in-depth learning frameworks based on online recruitment, research reports, online search, papers, tutorials, GitHub and other data.

At present, the most popular direction in the field of data science is still deep learning. Accordingly, the framework of in-depth learning is undergoing rapid change. The most popular framework today, except Theano, did not exist five years ago.

I wanted to collect some more interesting evidence about which framework, so I did this impact score. I used 11 data sources of seven different categories to assess the usage, interest and popularity of the framework. Then I weighted the data and got the results of this paper. Specific code can be seen on Kaggle Kernel: discdiver/deep-learning-framework-power-scores-updated on September 20, 2018: At the request of many readers, I expanded the evaluation scope of the framework, adding Caffe, Deeplearning 4J, Caffe2, Chainer. Current evaluations include a framework where KDNuggets use more than 1% of all reports in the survey.

Update 21 September 2018: I have improved some measurement methods.

No more verbosity. Here is the impact score of the Deep Learning Framework.

There is no doubt that TensorFlow ranks first. But I have some unexpected findings. See below for details.

frame

All frameworks for evaluation are open source, with Python interfaces provided in addition to one, and some frameworks provide interfaces in R or other languages.

TensorFlow is the undisputed winner. GitHub is the hottest, Google searches the most, Medium articles, Amazon books, ArXiv papers the most. Most developers use TensorFlow, which is also mentioned in most job descriptions of online recruitment. Google stands behind TensorFlow.

Keras has an API designed for humans, not machines. It is the second most popular framework in almost all assessments. Keras uses TensorFlow, Theano or CNTK as the underlying engine. If you are new to intensive learning, start with Keras.

Overall, PyTorch is the third most popular framework and the second most popular independent framework. It is younger than TensorFlow, and its popularity is growing rapidly. It also supports some customizations that TensorFlow does not yet support. Behind PyTorch stands Facebook.

Caffe is the fourth most popular framework. It's almost five years old. Relatively speaking, some employers also require familiarity with Caffe, and some academic papers also use Caffe, but few people use Caffe recently.

Theano, developed by the University of Montreal in 2007, is the oldest and most influential Python in-depth learning framework. Its popularity has dropped considerably, and major developers have announced that they will no longer release large versions of new features. However, there are still some updates. Theano is still ranked fifth.

Amazon uses Apache-hatched MXNET. It is the sixth most popular deep learning library.

CNTK is Microsoft's cognitive toolkit. It reminds me of many other Microsoft products that tried to compete with the tools offered by Google and Facebook, but didn't win much use.

Deeplearning4J, also known as DL4J, is used with the Java language. It is the only quasi-popular framework that does not provide Python interfaces. However, you can import models written in Keras. This is also the only framework where two different search terms (Deeplearning4J and DL4J) occasionally return different results. I used higher numbers in each measure. Since the score of this framework is rather low, there has been no substantial change.

Caffe2 is another Facebook open source product. It is built on Cafe and is now part of the PyTorch project (the same GitHub repository). Since Caffe2 does not have its own code repository, I use GitHub data from the old repository.

Chainer is a framework developed by Japanese company Preferred Networks. It has a small number of users.

FastAI is built on PyTorch. Its API borrows from Keras and requires very little code (less than Keras) to produce powerful results. FastAI is currently a cutting-edge framework and is in the process of rewriting code for version 1.0, which is expected to be released in October 2018. Jeremy Howard, author of FastAI, is chairman of Kaggle. He once wrote an Introducing Pytorch for fast.ai, discussing why FastAI moved from using Keras to creating its own framework.

At present, there are no jobs that need to use this framework, and it is not widely used. However, because FastAI's free online courses are popular, there are naturally many users. This framework is powerful and easy to use, so it may soon become ubiquitous.

standard

I chose the following seven categories to assess the popularity and attention of in-depth learning frameworks.

Online Recruitment Description

Research on the Use of KDnuggets

Google Search Volume

Medium article

Amazon Data

ArXiv Papers

GitHub heat

Data acquisition from September 16, 2018 to September 21, 2018, source data can be accessed through Google Trial Form: https://docs.google.com/spreadsheets/d/1mYfHMZfuXGpZ0ggBVDot3SJMU-VsCsEGceEL8xd1QBo/edit?Usp=sharing?

I used Python's pandas library to explore popularity and plotly library for visualization. If you want to view interactive plotly charts, visit Kaggle Kernel, which I mentioned at the beginning of this article.

Online Recruitment Description

What kind of deep learning libraries are in higher demand in the current job market? I collected data from LinkedIn, Indeed, Simply Hired, Monster, Angel List.

There is no doubt that TensorFlow is the winner. If you want to find a job with in-depth study, you can learn one.

Please read the Chinese version for details.