Rethinking Augmentation: Experiment Results
小样本增强策略再思考：实验结果库
Official experiment results repository for the paper:
"When More is Not Better: Rethinking Data Augmentation under Small-Sample Regimes"
TL;DR: We find that in small-sample regimes (e.g., CIFAR-100 100-shot), complex augmentation strategies (like RandAugment) often yield diminishing returns and high instability. A single, well-tuned operation can achieve comparable accuracy with significantly lower variance.
核心发现：在小样本场景下，盲目增加数据增强的复杂度（如 RandAugment）往往收益递减且带来极大的不稳定性。我们发现，单一且经过调优的增强操作能在保持精度的同时显著降低方差。
📊 Key Findings / 核心发现
1. Stability-Accuracy Trade-off / 稳定性与精度的权衡 (CIFAR-100 100-shot)
| Policy / 策略 | Val Acc (%) | Stability (Std) | Note |
|---|
| RandAugment (N=2,M=9) | 42.24% | 1.17 | High Variance (Unstable) |
| Single-Op (Ours) | 40.74% | 0.78 (Lowest) | Stable & Reliable |
| Baseline | 39.90% | 1.01 | - |
2. Zero-Variance Generalization / 零方差泛化 (CIFAR-10 50-shot)
| Metric / 指标 | Result / 结果 |
|---|
| Top-1 Accuracy | 50.0% |
| Stability (Std) | 0.0 (Zero Variance across 3 seeds) |
📂 Repository Structure / 仓库结构
checkpoints/: PyTorch model weights (best_model.pth). / 官方模型权重。figures/: Paper visualizations (Heatmaps, Trade-off plots). / 论文可视化图表（热图、权衡对比图等）。- phase_c_final_policy.json: The discovered optimal single-operation policy. / 搜索出的最优单一增强策略。
- cifar10_50shot_results.csv: CIFAR-10 generalization experiment data. / CIFAR-10 泛化实验数据。
- stability_seeds_results.csv: Raw data verifying the stability claim. / 验证“稳定性”结论的原始数据。
- destructiveness_metrics.csv: LPIPS/SSIM analysis for semantic preservation. / 语义保真度分析数据。
📜 Citation / 引用
If you find this study useful, please cite our work:
如果您觉得这项研究对您有启发，请引用我们的工作：
(Citation will be updated upon acceptance / 论文接收后更新)
🔗 Links / 相关链接
