mseitzer/pytorch-fid
Pytorch Fid
pytorch-fid is a PyTorch-based evaluator and image distribution analysis library used to calculate the Fréchet Inception Distance. It functions as a benchmarking tool that maps image pixels to high-dimensional feature vectors using a pre-trained convolutional neural network to measure the mathematical divergence between real and synthetic datasets.
The library quantifies the quality and diversity of generative models by representing image feature sets as mean and covariance matrices. It allows for the extraction of latent representations from specific neural network layers, with configurable dimensionality to adapt evaluation metrics based on dataset sample sizes.
The framework includes utilities for batch-wise tensor processing and persistent statistics caching, which stores pre-computed dataset distributions in compressed archives to avoid redundant computations during repeated model comparisons.
Features
- Fréchet Inception Distances - Calculates the Fréchet Inception Distance to quantify the quality and diversity of synthetic images relative to real datasets.
- FID Evaluators - Implements a PyTorch-based evaluator for calculating the Fréchet Inception Distance in generative image models.
- Generative Model Evaluators - Acts as a benchmarking framework to compare synthetic image distributions against real dataset statistics.
- Synthetic Image Evaluation - Provides quantitative measurement of the fidelity and diversity of AI-generated images using statistical distribution analysis.
- Pre-trained Neural Feature Extraction - Provides the ability to use specific layers from pre-trained neural networks to extract high-dimensional feature vectors.
- High-Dimensional Distribution Analysis - Analyzes high-dimensional probability distributions of image features to measure the mathematical divergence between datasets.
- Feature Distribution Modeling - Models image dataset features as multivariate Gaussian distributions to enable statistical comparison via mean and covariance.
- Dataset Statistics Archives - Stores pre-computed image statistics in compressed archives to accelerate repeated model evaluations.
- Deep Learning Research Workflows - Integrates quantitative image metrics into deep learning research workflows to track model convergence and performance.
- Statistical Parameter Caching - Persists computed distribution parameters to disk to eliminate redundant neural network passes across multiple evaluations.
- Model Implementations - Fréchet inception distance implementation.
