TorchMetrics

A standardized collection of 100+ PyTorch metrics designed for effortless distributed evaluation and seamless Lightning integration.

TorchMetrics provides a rigorous implementation of machine learning evaluation tools (ranging from Accuracy and R2 Score to specialized domains like FID for GANs or WER for speech). Built on the PyTorch ecosystem, it handles the complexities of distributed training (DDP) by automatically synchronizing metric states across multiple GPUs. This library replaces custom boilerplate code with a modular API, ensuring that performance benchmarks remain consistent, reproducible, and optimized for high-performance computing environments.

https://torchmetrics.readthedocs.io/