FFCV

FFCV (Fast Forward Computer Vision) is an open-source data loading library that dramatically accelerates machine learning training by eliminating input pipeline bottlenecks. Developed by researchers at MIT, it stores datasets in its own efficient format and uses techniques such as just-in-time compilation and fast JPEG decoding to feed GPUs at full speed. The GitHub project has roughly 3,000 stars, and its headline results include training a ResNet-50 on ImageNet in a fraction of the usual time on a single machine.

Overview

In many training jobs the GPU sits idle waiting for data: images are decoded too slowly, augmentation runs on overloaded CPUs, and storage layouts fight the access pattern. FFCV attacks all three problems at once. Datasets are converted once into a memory-friendly .beton file; loading code is compiled with Numba for near-native speed; and decoding and augmentation are fused and parallelized. Although it was designed for computer vision, FFCV works for other ML workloads too and plugs into PyTorch training loops with a drop-in DataLoader replacement, so most projects can adopt it in an afternoon and immediately see shorter epochs.

Key Features

• High-throughput data loading that keeps GPUs saturated during training
• Custom .beton dataset format optimized for fast random access
• Just-in-time compiled loading and augmentation pipelines via Numba
• Up to an order-of-magnitude faster ImageNet-scale training in published benchmarks
• Drop-in PyTorch DataLoader replacement with a familiar API
• Works beyond vision for general machine learning workloads

💳 Pricing

FFCV is free and open source under the Apache-2.0 license. There is no paid edition or hosted service — it runs entirely inside your own training infrastructure.

Who Is It For

FFCV is for ML engineers and researchers whose training runs are bottlenecked by data loading, especially teams training vision models on large image datasets who want faster experiments without buying more GPUs.