chainer

Why we included this project

Chainer is worth a look if you want to see how a deep learning framework actually works instead of just calling into one. Its define-by-run design builds the computational graph as your code runs, so debugging, writing custom layers, and testing unusual architectures feel more direct than in static-graph frameworks. The project went into maintenance in 2019 and now only receives bug fixes, but the codebase still reads as a clear reference for automatic differentiation, CUDA support through CuPy, and multi-GPU training. That makes it genuinely useful for researchers digging into how backpropagation is implemented, and for teams stuck maintaining older Chainer models. Just do not pick it for greenfield projects today.