TensorFlow, Chainer, and PyTorch
Era 4 · Deep learning and foundation models (2012–2021) › Infrastructure and institutions
- Lineage
- Machines and computation
- Kind
- Implementation
Who
Google Brain, Preferred Networks, Facebook AI Research
What happened
Google releases TensorFlow under an open-source license on November 9, 2015. A few months earlier the Japanese company Preferred Networks had launched Chainer, and in January 2017 Facebook releases PyTorch, which adopts Chainer's way of programming networks.
Why it mattered
The tools for building and training deep networks become open and shared: the research code of the big labs can be used outside them.
The full story
Since 2011 Google Brain had trained its networks with DistBelief, an internal system. According to Jeff Dean and Rajat Monga, it was built only for neural networks, it was hard to configure, and it was so tightly tied to Google's infrastructure that sharing research code was nearly impossible. TensorFlow, its second generation, was released on November 9, 2015, under the open-source Apache 2.0 license. The same program could run on anything from a phone to hundreds of machines with thousands of GPUs.
In June 2015 Tokyo's Preferred Networks had launched Chainer with a different idea. In TensorFlow the whole network had to be defined before it was run; in Chainer the network is built as the computation runs, like an ordinary Python program. It was the first deep learning framework with this approach, called define-by-run, designed for networks that did not fit the frameworks built for convolutional ones, such as recurrent networks.
PyTorch, developed at Facebook AI Research from the Torch library, was released in January 2017. Its authors later wrote that its imperative design had been pioneered by Chainer and DyNet, and that PyTorch extended it to the whole workflow.
PyTorch and TensorFlow became two of the most widely used deep learning libraries, and in December 2019 Preferred Networks itself moved its development from Chainer to PyTorch.
Sources
Original source
- TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed Systems (arXiv:1603.04467) (opens in a new tab). Martín Abadi and 39 others (Google), 2016
Scholarship and context
- TensorFlow: Google's latest machine learning system, open sourced for everyone (Google Research, November 9, 2015) (opens in a new tab). Jeff Dean and Rajat Monga, 2015
- Chainer: a Next-Generation Open Source Framework for Deep Learning. LearningSys workshop at NIPS 2015 (opens in a new tab). Seiya Tokui, Kenta Oono, Shohei Hido, and Justin Clayton, 2015
- PyTorch: An Imperative Style, High-Performance Deep Learning Library (arXiv:1912.01703, NeurIPS 2019) (opens in a new tab). Adam Paszke, Sam Gross, Francisco Massa, and 18 others, 2019
- p. 2 (opens in a new tab) The imperative design, "pioneered by Chainer and DyNet."
How to cite
“TensorFlow, Chainer, and PyTorch.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/tensorflow-chainer-pytorch
This entry was last reviewed on November 13, 2026.