Preferred Networks is migrating its deep learning research platform from its own open source framework Chainer to PyTorch. The Japanese artificial intelligence startup unveiled the plan last week, assigning its new Chainer V7 to a “maintenance phase” in advance of the move. Preferred Networks will provide documentation and a library for Chainer users to facilitate the transition to PyTorch.
According to a Nikkei survey, Preferred Networks ranks No.1 on estimated corporate value among 181 Japanese startups, with an estimated valuation of JP￥351.5 billion (US$3.24 billion). Japanese auto maker Toyota has been working closely with Preferred Networks since its founding in 2014 and has pumped more than JP ¥11 billion (US$101 million) into the company’s deep learning, robotics and self-driving R&D.
Preferred Networks initially released and open sourced Chainer in June 2015. The framework has been widely adopted as a standard method by current mainstream deep learning researchers for its intuitiveness and flexibility.
The maturation of deep learning frameworks over the past years is the reason why Preferred Networks made this decision. Instead of making small adjustments to differentiate itself from competitors, the company is more interested in collaborating with PyTorch community contributors such as Facebook to ensure sustainable growth and a healthy ecosystem for PyTorch.
Preferred Networks President and CEO Toru Nishikawa says migrating to PyTorch was an important decision for his company: “We firmly believe that by participating in the development of one of the most actively developed frameworks, PFN can further accelerate the implementation of deep learning technologies, while leveraging the technologies developed in Chainer and searching for new areas that can become a source of competitive advantage.”
The move is the latest gain for Facebook-developed PyTorch in its ongoing war with Google’s TensorFlow. PyTorch has shown overwhelming growth among machine learning researchers this year, with 69% of CVPR papers, 75+% of NAACL and ACL papers, and 50+% of ICLR and ICML papers using it. The consensus is that researchers prefer PyTorch’s simplicity, great API and performance. Tensorflow however still dominates the industry market and has a loyal base within Google/DeepMind.
PyTorch is clearly a framework on the rise, but it’s too early to speculate whether or when it might emerge as a virtual monopoly.
Author: Yuqing Li | Editor: Michael Sarazen