This Pytorch Library ‘SpeedTorch’ Might Speedup Your Pytorch Deep Learning Projects

While browsing on reddit, I found out about ‘SpeedTorch’. SpeedTorch is a library developed for Pytorch, for fast transfer rates between pinned CPU tensors and GPU Pytorch variables.

This Pytorch library rotates around Cupy memory maps pinned to CPU which can achieve upto 110 times faster CPU to GPU transfers in comparison to general Pytorch pinned CPU tensors.

Github Link: https://github.com/Santosh-Gupta/SpeedTorch

Reddit Discussion: https://www.reddit.com/r/MachineLearning/comments/d4recl/p_speedtorch_4x_faster_pinned_cpu_gpu_data/

https://github.com/Santosh-Gupta/SpeedTorch/tree/master/SpeedTorch