How to beginner · 3 min read

How to calculate accuracy in PyTorch

Q: How to calculate accuracy in PyTorch

In PyTorch, calculate accuracy by comparing predicted labels with true labels using tensor operations like torch.argmax and torch.eq. Then compute the mean of correct predictions as accuracy.

Quick answer

In PyTorch, calculate accuracy by comparing predicted labels with true labels using tensor operations like torch.argmax and torch.eq. Then compute the mean of correct predictions as accuracy.

PREREQUISITES

Python 3.8+
pip install torch>=2.0

Setup

Install PyTorch if not already installed. Use the following command to install the latest stable version:

bash

pip install torch torchvision

Step by step

This example shows how to calculate accuracy for a classification model's output logits and true labels tensors.

python

import torch

# Example logits from model (batch_size=4, num_classes=3)
logits = torch.tensor([[2.0, 1.0, 0.1],
                       [0.5, 2.5, 0.3],
                       [1.2, 0.7, 1.8],
                       [0.1, 0.2, 0.3]])

# True labels
labels = torch.tensor([0, 1, 2, 2])

# Get predicted class indices
preds = torch.argmax(logits, dim=1)

# Compare predictions with true labels
correct = torch.eq(preds, labels)

# Calculate accuracy as mean of correct predictions
accuracy = correct.float().mean().item()

print(f"Accuracy: {accuracy * 100:.2f}%")

output

Accuracy: 75.00%

Common variations

You can calculate accuracy during training or evaluation loops. For multi-label classification, use thresholding instead of argmax. To compute accuracy on GPU, ensure tensors are on the same device.

python

def accuracy(logits, labels):
    preds = torch.argmax(logits, dim=1)
    return (preds == labels).float().mean().item()

# Example usage in training loop
for batch in dataloader:
    inputs, labels = batch
    outputs = model(inputs)
    acc = accuracy(outputs, labels)
    print(f"Batch accuracy: {acc * 100:.2f}%")

Troubleshooting

If accuracy is always zero, verify that predictions and labels have matching shapes and are on the same device.
For multi-class problems, ensure you use torch.argmax on logits, not probabilities.
Check that labels are integer class indices, not one-hot encoded vectors.

✅

Key Takeaways

Use torch.argmax on model outputs to get predicted classes for accuracy calculation.
Compare predictions with true labels using torch.eq and average correct predictions.
Ensure tensors are on the same device and have compatible shapes to avoid errors.
For multi-label tasks, adapt accuracy calculation with thresholding instead of argmax.
Wrap accuracy calculation in a reusable function for clean training and evaluation code.

Verified 2026-04

Verify ↗