Research map / Image generation / GANs
Improving GAN Training with Probability Ratio Clipping and Sample Reweighting
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 |
|---|---|---|---|
| add discriminator sample re-weighting (ablation) | DCGAN · CIFAR-10 random configs | collapse rate (%): 52.4% -> 30.2% | Helped |
| add discriminator sample re-weighting (ablation) | DCGAN · CIFAR-10 random configs | average IS: 4.2 -> 5.1 | Helped |
| add discriminator sample re-weighting (ablation) | CT-GAN · CIFAR-10 | IS/FID | Helped |
| add probability ratio clipping on top of re-weighting (ablation) | CT-GAN + re-weighting · CIFAR-10 | IS/FID | Helped |
| add probability ratio clipping alone (no re-weighting) (ablation) | WGAN-GP · CIFAR-10 (DCGAN) IS curve | Inception score | No clear effect |
| replace WGAN-GP objectives with proposed ones | RelGAN · Synthetic text length 20 | oracle NLL: 6.89 -> 5.67 | Helped |
| replace WGAN-GP objectives with proposed ones | RelGAN · Synthetic text length 40 | oracle NLL: 6.78 -> 6.14 | Helped |
| replace WGAN-GP objectives with proposed ones | RelGAN · EMNLP2017 WMT News | BLEU-2: 0.872 -> 0.905 | Helped |
| replace WGAN-GP objectives with proposed ones | RelGAN · EMNLP2017 WMT News | NLL_gen: 2.209 -> 2.265 | Hurt |
| use proposed objectives instead of RelGAN 100 | RelGAN · EMNLP2017 WMT News | BLEU-4: 0.501 -> 0.470 | Hurt |
| use proposed objectives instead of RelGAN 1000 | RelGAN · EMNLP2017 WMT News | human score: 3.42 -> 3.59 | Helped |
| replace adversarial objectives with proposed ones | VAE style transfer model · Yelp sentiment transfer | BLEU: 32.82 -> 33.45 | Helped |
Closest papers by meaning
- A Variational Inequality Perspective on Generative Adversarial Networks (ICLR 2019 · similarity 0.91)
- Variational Annealing of GANs: A Langevin Perspective (ICML 2019 · similarity 0.91)
- Stabilizing Training of Generative Adversarial Networks through Regularization (NeurIPS 2017 · similarity 0.90)
- On-Line Adaptative Curriculum Learning for GANs (AAAI 2019 · similarity 0.90)
- Training Generative Adversarial Networks with Limited Data (NeurIPS 2020 · similarity 0.90)
- Progressive Growing of GANs for Improved Quality, Stability, and Variation (ICLR 2018 · similarity 0.90)
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.