Chainer is an open-source deep learning framework used for training and inference. It supports CUDA/cuDNN object-oriented APIs to build and train networks. It was developed in Python and runs on NumPy and CuPy. One of the key features is “define-by-run) which provider automatic differentiation APIs. 

Japanese-based Preferred Networks developed the product in collaboration with IBM in 2015. However, in 2019, the company transitioned its developed efforts to PyTorch. 

Project Background

  • Framework: Chainer
  • Author: Seiya Tokui
  • Initial Release: June 9, 2015
  • Type: Python-based open-source deep-learning library
  • License: MIT
  • Contains: Chainer has four extension libraries – ChainerMN, ChainerRL, ChainerCV, and ChainerUI.
  • Language: Python
  • GitHub: chainer
  • Runs On: Windows, macOS, and Linux
  • Twitter: Chainer


  • Building computational graph: It uses a unique technique and builds computation graph “on-the-fly” during training. With this, users can change the graph at each iteration depending on conditions.
  • Academic papers and research: Chainer is ideal for research projects focused on computer vision, speech processing, robotics, and natural language processing. Moreover, it is a great deep learning library for the research and development of new services and products.
  • Imperative API in plain Python and NumPy: Chainer uses imperative API that offers great flexibility for the implementation of complex neural networks.
  • Big firms like Toyota Motors, Panasonic, and FANUC use Chainer extensively. It supports computation on CPUs and GPUs. With CUDA computation, it only needs a few lines of code to leverage a GPU. In addition, users can use it on multiple GPUs with no effort. IBM, Intel, Microsoft, NVIDIA, and AWS all these companies support Chainer.

Scroll to Top