singa

Why we included this project

Apache SINGA is a deep learning library aimed at teams that want to train models across several GPUs without assembling a cluster stack on their own. It handles data-parallel training on one or more nodes, keeps a computation graph so backward propagation runs automatically, and ships with standard optimizers like Adam, RMSProp, and SGD with momentum. Two details set it apart: it exchanges models through ONNX for interoperability with other tools, and it can call trained models as stored procedures inside a relational database, so predictions live where the application data already sits. That combination suits research groups and teams that want training results queryable right next to their own data.