Twister2 Deep Learning is currently working on supporting Pytorch, Tensorflow and MXNet. With our 0.5.0 release we have introduced an experimental version of a deep learning with distributed data parallel Pytorch implementation. We use an MPI backend to support distributed training and currently we are working on providing better performance with GPU.

Twister2:DeepNet

Twister2:DeepNet is the distributed runtime for deep learning algorithm training and inference. Currently we support an experimental version of a distributed training for Pytorch. Our initial implementation uses Apache Arrow and Parquet to move data from our data pre-processing pipeline to the data training framework. We are actively working on transforming this process into a seamless in-memory computation model.