3.1 KiB
3.1 KiB
QA Report — GNN-AMR Correction Presentation
PPTX Status
- File:
output/final_presentation_cn.pptx - Creation: ✅ Success
- Slides: 17
- Dimensions: 13.3" × 7.5" (16:9 widescreen)
- Images inserted: 6 (from existing result files)
Slide Summary
| Slide | Title | Type | Images |
|---|---|---|---|
| 1 | GNN 引导的自适应网格加密 | Title | 0 |
| 2 | 研究背景：为什么这个问题重要 | Background | 0 |
| 3 | 技术瓶颈：传统残差驱动 AMR 的局限 | Gap | 0 |
| 4 | 核心思路：GNN 学习修正信号 | Approach | 0 |
| 5 | 全流程概览：5 步流水线 | Pipeline | 0 |
| 6 | Step 1：数据生成与标记策略 | Method | 1 |
| 7 | Step 1：15 维特征工程 | Method | 0 |
| 8 | Step 3：CorrectionGNN 架构 | Architecture | 0 |
| 9 | Step 3：训练结果 | Result | 0 |
| 10 | Step 4a：离线评估 | Result | 1 |
| 11 | Step 4b：Rollout 评估 | Result | 1 |
| 12 | Step 5：端到端 AMR 可视化 | Visualization | 1 |
| 13 | 单步对比：GNN vs Teacher vs Physics | Visualization | 1 |
| 14 | 验证与稳健性 | Validation | 1 |
| 15 | 创新点与可复用价值 | Innovation | 0 |
| 16 | 局限性与未解决问题 | Limitations | 0 |
| 17 | 总结 | Summary | 0 |
Figure Assets Used
| Asset | Source | Slide |
|---|---|---|
data_correction_check/sample0011_step002.png |
check_correction_data.py | 6 |
result/correction/vis/sample0000_step004_vis.png |
test_correction.py | 10 |
result/correction/rollout/aw_rel_vs_elements.png |
eval_correction.py | 11 |
result/correction/viz/amr_overview.png |
viz_correction.py | 12 |
result/correction/viz/marks_sample0005_step010.png |
viz_correction.py | 13 |
result/correction/viz/compare_sample0005_step010.png |
viz_correction.py | 14 |
Data Sources
- Training log:
outlook/ckpt/correction_train_log.json(best epoch 40, AUC=0.950) - Rollout results:
outlook/result/correction/rollout/eval_results.json - Dataset:
outlook/data_correction/(100 samples, 1236 step files)
Verification
- ✅ PPTX reopens successfully
- ✅ 17 slides created
- ✅ 6 images embedded
- ✅ All slide titles in Chinese
- ✅ Consistent color scheme (blue/red/green/orange accents)
- ✅ 16:9 widescreen format
- ⚠️ No speaker notes added (lightweight deck)
- ⚠️ No headless renderer available for visual QA
Presentation Arc
Problem-to-Solution (methods paper):
- Background: Helmholtz scattering + FEM
- Gap: Traditional AMR needs per-step FEM solve
- Approach: Correction learning (teacher - physics)
- Pipeline: 5-step workflow
- Architecture: CorrectionGNN
- Training: AUC=0.950, top-k 46.6%
- Evaluation: Offline + Rollout
- Visualization: AMR + step comparison
- Validation: 5-check data quality
- Innovation + Limitations + Summary
Known Limitations
- No speaker notes (can be added manually)
- Some text-heavy slides (features, architecture) use structured layouts rather than figures
- Rollout improvement is modest (~2%) — presented honestly
- No rendered slide previews (no headless renderer available)