fidax 0.2

An sklearn-style FID metric class for Jax

FIDax

A JAX implementation of the Fréchet Inception Distance (FID) metric for evaluating generative models in form of a sklearn-compatible metric.

Features

• Pure JAX Implementation: Leverages JAX's JIT compilation for fast computation
• Pre-computed Statistics: Supports using pre-computed real image statistics for faster evaluation
• GPU Accelerated: Optimized for CUDA-enabled GPUs
• Torchmetrics Compatible: Results match torchmetrics implementation up to 1e-1 absolute tolerance with FP32 execution of the InceptionV3 model and FP64 for the metric computation on CIFAR10 tests

Installation

# Clone the repository
git clone [e-mail]:wittenator/fidax.git
cd fidax

# Install dependencies using uv
uv sync --frozen

or install it directly as a dependency with e.g. uv:

uv add fidax

Quick Start

import jax 
jax.config.update("jax_enable_x64", True)
import jax.numpy as jnp
from fidax.fid import FrechetInceptionDistance

# Initialize FID metric
fid = FrechetInceptionDistance(max_samples=100)

# Update with real images (shape: [N, 299, 299, 3], range: [-1, 1])
real_images = jnp.random.uniform(-1, 1, (100, 299, 299, 3))
fid.update(real_images, real=True)

# Update with generated/fake images
fake_images = jnp.random.uniform(-1, 1, (100, 299, 299, 3))
fid.update(fake_images, real=False)

# Compute FID score
fid_score = fid.compute()
print(f"FID Score: {fid_score}")

Advanced Usage

Pre-computed Statistics

# Use pre-computed real statistics for faster evaluation
real_stats = {
    "mu": mu_real,      # Mean of real activations
    "sigma": sigma_real # Covariance of real activations
}

fid = FrechetInceptionDistance(max_samples=1000, real_stats=real_stats)
# Only need to update with fake images
fid.update(fake_images, real=False)

Requirements

• Python ≥ 3.12
• JAX with CUDA support
• Flax
• NumPy

See pyproject.toml for complete dependency list.

Development

This project uses a development container with GPU support. To set up the development environment:

# The dev container will automatically install dependencies
# Run tests
uv run pytest src/fidax/test_fid_metric.py

Testing

The implementation includes tests against torchmetrics:

uv run pytest src/fidax/test_fid_metric.py -v

Tests verify:

• Equivalence with torchmetrics implementation
• Pre-computed statistics functionality
• Real-world performance on CIFAR-10 dataset

License

Apache 2.0 License - see LICENSE for details.

Related Projects

• jax-fid-parallel - Parallel implementation of FID computation in JAX
• jax-fid - Original JAX implementation of FID that inspired this project