torchmetrics

What it is and what it does

Torchmetrics is a PyTorch library that standardizes metric computation for machine learning models, particularly in distributed and multi-device training scenarios. It provides a collection of pre-built metrics (accuracy, F1, precision, recall, and many others) alongside an API for defining custom metrics. The library handles the boilerplate of accumulating metric state across batches and synchronizing results across multiple GPUs or nodes automatically, so you can focus on model logic rather than metric plumbing.

The package works as a PyTorch module—you instantiate a metric, move it to your device, call it with predictions and targets in a training loop, and it tracks state internally. It supports single-GPU, multi-GPU with DDP, and multi-node setups without code changes. It also integrates tightly with PyTorch Lightning, where metrics are automatically placed on the correct device and logged with minimal extra code.

Use it for

• Evaluate classification models on multi-GPU training runs without manually synchronizing accuracy or F1 scores across devices.
• Track cumulative metrics over an entire epoch or dataset without storing all predictions in memory.
• Implement custom metrics by subclassing Metric and defining update() and compute() methods for domain-specific evaluation.
• Log metrics to PyTorch Lightning training runs with automatic device placement and distributed synchronization.
• Benchmark model performance reproducibly using a standardized metric interface across different training frameworks.

Worth the install?

AI-flagged interpretation of the facts on this page. Verify before relying on it.

Install

torchmetrics on PyPI

Before you install

Low install friction with a pure-Python wheel distribution. Active maintenance with recent releases and 2458 repository stars. Requires torch and three lightweight dependencies (numpy, packaging, lightning-utilities).

Requires PyTorch (torch) to be installed; Python 3.10 or later.

License in practice

Apache-2.0 permissive license allows commercial and private use with minimal restrictions—standard for ML tooling.

Quickstart

pip install torchmetrics

import torch
import torchmetrics

metric = torchmetrics.classification.Accuracy(task="multiclass", num_classes=5)
device = "cuda" if torch.cuda.is_available() else "cpu"
metric.to(device)

preds = torch.randn(10, 5).softmax(dim=-1).to(device)
target = torch.randint(5, (10,)).to(device)
acc = metric(preds, target)

Package facts

Evidence: torchmetrics-1.9.0-py3-none-any.whl

Tags

More Artificial Intelligence packages

See also torcheval · entmax · torchrunx · ema-pytorch · nv-one-logger-pytorch-lightning-integration · lightning · motmetrics · spmd-types · litdata · torch