PyTorch FID

0.3.0 verified Mon Apr 27 auth: no python

A PyTorch implementation of the Frechet Inception Distance (FID) for evaluating generative models. Version 0.3.0 is the latest stable release. The package is mature with infrequent releases.

pip install pytorch-fid

Common errors

error ModuleNotFoundError: No module named 'pytorch_fid' ↓

cause Package not installed or misspelled.

fix

Run pip install pytorch-fid and verify with pip list | grep pytorch-fid.

error AttributeError: module 'pytorch_fid' has no attribute 'calculate_fid_given_paths' ↓

cause Direct import of the function from the package root is incorrect.

fix

Import via from pytorch_fid import fid_score then call fid_score.calculate_fid_given_paths(...).

error RuntimeError: CUDA out of memory. Tried to allocate ... MiB (GPU 0; ... MiB total capacity) ↓

cause Batch size too large for GPU memory.

fix

Reduce batch_size parameter (e.g., to 16 or 8) or use CPU with device='cpu'.

error ValueError: Input path ... does not exist or is not a directory ↓

cause Incorrect path to image directory.

fix

Check that both paths exist and are directories containing images.

Warnings

breaking In version 0.4.0 (unreleased), the API will change; the function `calculate_fid_given_paths` will be renamed to `compute_fid` and its argument signature will change. Check the GitHub master branch for future compatibility. ↓

fix Use the current API (0.3.0) or monitor the repo for updates.

gotcha The Inception v3 model used by pytorch-fid is not the standard torchvision one; it uses pre-trained weights from the TensorFlow implementation. If you use your own Inception model, FID scores will be inconsistent. ↓

fix Always use the bundled weights (loaded automatically) from the pytorch_fid/inception.py file.

deprecated The `dims` parameter accepts values 64, 192, 768, 2048. Using dims other than 2048 produces incomplete FID (not the true FID). This is a leftover from older experiments. ↓

fix Always use dims=2048 for standard FID.

gotcha Image directories must contain only valid image files (jpg, png). Non-image files or corrupted images cause silent failures or crashes. ↓

fix Sanitize directories to contain only supported image formats.

Imports

calculate_fid_given_paths

wrong

from pytorch_fid.fid_score import calculate_fid_given_paths

correct

from pytorch_fid import fid_score; fid_score.calculate_fid_given_paths(paths, batch_size, device, dims, num_workers)

The function is not exported directly; you must access it via the fid_score module.

Quickstart

Example of calculating FID between two directories of images.

import torch
from pytorch_fid import fid_score

paths = ['path/to/real/images', 'path/to/fake/images']

# Use GPU if available
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

# Calculate FID
fid_value = fid_score.calculate_fid_given_paths(paths, batch_size=50, device=device, dims=2048, num_workers=8)
print('FID:', fid_value)