ffcv

Why we included this project

Dataloading is a classic bottleneck in vision training: the GPU sits idle waiting for images to be decoded off disk. FFCV sidesteps that by converting a dataset into a custom binary format and serving it through a highly optimized PyTorch loader, so the GPU stays fed. The speedup is dramatic in practice, with ImageNet-scale training finishing in roughly half an hour on a single GPU and benchmark costs dropping accordingly. Since you keep the same training loop and only swap in the loader, adopting it is fairly painless if you already work in PyTorch. If you run the same large vision datasets over and over, that saved iteration time adds up quickly.