User:Packet0/sandbox

Chainer en
Chainer is a flexible, open source library for training and deploying neural networks. It adopts a "Define-by-run" scheme, where it dynamically generates data structure necessary for back propagation and automatic differentiation during the actual forward calculation, on the fly. This is different from symbol based APIs found in many neural network libraries, although there are other libraries such as MXNet that support similar programing scheme. Chainer itself is written purely in Python on top of NumPy and CuPy. The development is led by Japanese venture company Preferred Networks with support from IBM, Intel, Microsoft, and Nvidia.