Improving GAN Precision and Recall through GMM and Gradient Ascent
Balancing precision and recall in GANs.
Project Overview
Standard GANs struggle to balance precision and recall; improving visual quality often leads to mode collapse, while increasing diversity can degrade sample sharpness. This project explores two strategies to modify latent space structure and training dynamics on MNIST.
Methodology
Gaussian Mixture Models (GM-GAN)
- Multi-modal structure imposed in latent space
- Encourages generator to map regions to semantic modes
- Parameter σ controls precision-diversity trade-off
Discriminator Gradient Flow (DGFlow)
- Refines samples using discriminator as density ratio estimator
- Step size η controls gradient ascent magnitude
- Diffusion coefficient γ prevents mode collapse
- Refinement steps control iterations
Experimental Results
- GMM: Optimal at σ=0.1 → Precision: 0.89, Recall: 0.79, FID: 19.19
- DGFlow: Best triplet (η=5×10^-4, γ=5×10^-2, t=5)
- Confirmed importance of entropy term for diversity
Technologies
- Language: Python
- Framework: PyTorch
- Metrics: Precision, Recall, FID (Fréchet Inception Distance)
- Methods: GMM, Gradient Flow, Entropy regularization
Project Report
View or download the full project report:
GAN Precision and Recall Report