Artificial Intelligence Now

Chapter 10. Complex Neural Networks Made Easy by Chainer

Shohei Hido

Chainer is an open source framework designed for efficient research into and development of deep learning algorithms. In this post, we briefly introduce Chainer with a few examples and compare with other frameworks such as Caffe, Theano, Torch, and TensorFlow.

Most existing frameworks construct a computational graph in advance of training. This approach is fairly straightforward, especially for implementing fixed and layer-wise neural networks like convolutional neural networks.

However, state-of-the-art performance and new applications are now coming from more complex networks, such as recurrent or stochastic neural networks. Though existing frameworks can be used for these kinds of complex networks, it sometimes requires (dirty) hacks that can reduce development efficiency and maintainability of the code.

Chainer’s approach is unique: building the computational graph on-the-fly during training.

This allows users to change the graph at each iteration or for each sample, depending on conditions. It is also easy to debug and refactor Chainer-based code with a standard debugger and profiler, since Chainer provides an imperative API in plain Python and NumPy. This gives much greater flexibility in the implementation of complex neural networks, which leads in turn to faster iteration, and greater ability to quickly realize cutting-edge deep learning algorithms.

In the following sections, I describe how Chainer actually ...