PruneTrain: Fast Neural Network Training by Dynamic Sparse Model Reconfiguration

# 🚀 Feature request

PruneTrain. {...} By using a structured-pruning approach and additional reconfiguration techniques we introduce, the pruned model can still be efficiently processed on a GPU accelerator. Overall, **PruneTrain achieves a reduction of 39% in the end-to-end training time of ResNet50 for ImageNet by reducing computation cost by 40% in FLOPs, memory accesses by 37% for memory bandwidth bound layers, and the inter-accelerator communication by 55%.**

## Motivation

I'm pre-training some midsize language models from scratch. If you tell me that I can pretrain a network with 1% drop in performance while cutting down the energy demand of the training by up to 40% and speeding inference time at the same time, I will buy it.

## Your contribution

https://arxiv.org/abs/1901.09290. I can not understand why the authors did not open source the code, since it could reduce the global warming, speedup experimentation and reduce energy consumption.