One simple graphic: Researchers love PyTorch and TensorFlow

Interest in PyTorch among researchers is growing rapidly.

In a recent survey—AI Adoption in the Enterprise, which drew more than 1,300 respondents—we found significant usage of several machine learning (ML) libraries and frameworks. About half indicated they used TensorFlow or scikit-learn, and a third reported they were using PyTorch or Keras.

I recently attended an interesting RISELab presentation delivered by Caroline Lemieux describing recent work on AutoPandas and automation tools that rely on program synthesis. In the course of her presentation, Lemieux reviewed usage statistics they had gathered on different deep learning frameworks and data science libraries. She kindly shared some of that data with me, which I used to draw this chart:

Figure 1. Number of papers posted on that mention each framework. Source: Data from RISELab and graphic by Ben Lorica.

The numbers are based on simple full-text searches of papers posted on the popular e-print service Specifically, they reflect the number of papers which mention (in a full-text search) each of the frameworks. Using this metric, the two most popular deep learning frameworks among researchers are TensorFlow and PyTorch. From January to the end of June 2019, about 1,800 papers mentioned TensorFlow and a comparable number mentioned PyTorch. Most notably, interest in PyTorch among researchers is growing rapidly: it grew 194% year-over-year (Jan-Jun 2018 compared to Jan-Jun 2019).

To the extent that researchers and teachers are harbingers and strongly influence what future professionals might use, look for PyTorch to also gain users among data scientists, developers, ML engineers, and companies. In a recent post, we outlined the suite of tools for managing machine learning in the enterprise. Many of the tools and companies we highlighted in that post support the popular ML libraries and deep learning frameworks (particularly TensorFlow and PyTorch), thus we predict both of these frameworks will be equally viable options for enterprise users.

[A version of this post appears on the O’Reilly Radar.]

Update: This post is the subject of an interesting thread on reddit.

Related content:

%d bloggers like this: