This repository reproduces the ICLR 2019 paper demonstrating that for structured network pruning, training a pruned model from scratch often achieves comparable or better accuracy than the traditional train-prune-finetune approach. The work challenges common assumptions about the value of learned “important” weights in pruning, arguing that the pruned architecture itself is more crucial to final model efficiency. The code includes implementations of several pruning methods and trained ImageNet models for reproducibility.