Research map / Image generation / GANs
GANs Can Play Lottery Tickets Too
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Prune both the generator and discriminator instead of only the generator (ablation) | SNGAN · CIFAR-10 | extreme matching sparsity: 67.2% -> 73.8% | Helped |
| Prune both the generator and discriminator instead of only the generator (ablation) | SNGAN · CIFAR-10 | FID (best): 14.19 -> 14.59 | Hurt |
| Reset the discriminator to its initial weights rather than fine-tuning its dense weights (ablation) | IMP_G · CIFAR-10 | FID | Helped |
| Distill dense-discriminator knowledge into the pruned discriminator (ablation) | IMP_GD · CIFAR-10 | FID | Helped |
| Use iterative rather than one-shot magnitude pruning of the generator (ablation) | SNGAN · CIFAR-10 | extreme matching sparsity: 59.0% -> 67.2% | Helped |
| Use iterative rather than one-shot magnitude pruning of generator and discriminator (ablation) | SNGAN · CIFAR-10 | extreme matching sparsity: 48.8% -> 73.8% | Helped |
| Use magnitude-based masks instead of randomly generated pruning masks (ablation) | SNGAN · CIFAR-10 | extreme matching sparsity: 36.0% -> 67.2% | Helped |
| Keep the ticket's specific initialization instead of resetting to another random initialization (ablation) | SNGAN · CIFAR-10 | extreme matching sparsity: 59.0% -> 67.2% | Helped |
| Rewind weights to 5% of training rather than reset to initialization (ablation) | SNGAN · CIFAR-10 | extreme matching sparsity: 67.23% -> 86.26% | Helped |
| Rewind weights to 20% of training rather than reset to initialization (ablation) | SNGAN · CIFAR-10 | extreme matching sparsity: 67.23% -> 89.26% | Helped |
| Use a different random initialization for transfer instead of the source initialization (ablation) | IMP_GD ticket · STL-10 | FID: 121.8 -> 113.1 | Helped |
| Use the source initialization rather than a different random initialization for transfer (ablation) | IMP_G ticket · STL-10 | FID: 119.2 -> 116.7 | Helped |
Closest papers by meaning
- Data-Efficient GAN Training Beyond (Just) Augmentations: A Lottery Ticket Perspective (NeurIPS 2021 · similarity 0.93)
- Winning Lottery Tickets in Deep Generative Models (AAAI 2021 · similarity 0.90)
- Re-GAN: Data-Efficient GANs Training via Architectural Reconfiguration (CVPR 2023 · similarity 0.90)
- SMaRt: Improving GANs with Score Matching Regularity (ICML 2024 · similarity 0.88)
- Compressing Image-to-Image Translation GANs Using Local Density Structures on Their Learned Manifold (AAAI 2024 · similarity 0.88)
- RG-GAN: Dynamic Regenerative Pruning for Data-Efficient Generative Adversarial Networks (AAAI 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.