marcojira/fld
PyTorch code for FLD (Feature Likelihood Divergence), FID, KID, Precision, Recall, etc. using DINOv2, InceptionV3, CLIP, etc.
### Overview:
FLD is a comprehensive, sample-based metric designed to evaluate the generalization of generative models. It is sensitive to sample fidelity, diversity, and novelty, making it particularly useful for detecting overfitting and memorization. The metric is sensitive to perceptual likelihood in a feature space, providing a nuanced evaluation of generative models.
### Key Features:
- **FLD (Feature Likelihood Divergence)**: A new metric for generative models that evaluates sample fidelity, diversity, and novelty.
- **Support for Multiple Metrics**: Includes FID (Feature Independence Divergence), KID (Kernel Divergence), Precision, Recall, and others.
- **Feature Extraction**: Uses DINOv2, InceptionV3, and CLIP feature spaces for accurate evaluation.
- **Computation from Code**: Computes metrics directly from the generative model within Python code, allowing for easy integration.
### Tech Stack:
- **Python**: The primary programming language used for implementation.
- **PyTorch**: Framework used for implementing the generative models and computing metrics.
### Why It Matters:
- **Comprehensive Evaluation**: Provides a comprehensive evaluation of generative models, sensitive to overfitting and memorization.
- **Sample-Based Evaluation**: Evaluates generative models using samples from the test set, making it effective for detecting overfitting.
- **Versatile Application**: Can be extended to other modalities like audio, video, and tabular data using appropriate feature extractors.
### Repository:
- **Repository**: `marcojira/fld`
- **