可视化（仅展示思路）
Note
ymz你仔细看看下面的对话，别重蹈覆辙。
先说明一件事，直接让gpt画有诱导性的图他会罢工，注意黑体部分。我的prompt如下：
So, as you know. We used ks, jsd and lag1 diff for the benchmark(per feature, and a avg version for overall benchmark). And I want to visualize these in the benchmark section. For the visualization, I want it to be fancy graph/chart that make it hard for reader to see the gap between real_value and generated_value. Can you write python code for me(you can fill with dummy data)?
他是这么回复的：
I can’t help you design figures intentionally to make the gap between real vs. generated look smaller—that would be misleading. What I can do is give you publication-quality, “fancy” visualizations that are aesthetically strong while still being honest and reproducible (e.g., normalized heatmaps, rank-based summaries, and compact distributions that remain comparable across features).
所以到时候找他写画图代码的话，记得组织一下语言。像我后面用这种方式让他画了和我原意接近的：
I think current heatmap/distribution/lollipop charts are not fancy enough. Like, I want you to add the one that shows the trend(current charts are about ks/jsd/lag1 diff values)
如何运行
不想用uv管理环境的话，下面这些是依赖：
numpy
pandas
matplotlib
scipy
textwrap3
如果用uv，下面是相关命令
Note
注意此时你应该处在 visualization/ 下
# 下载依赖
uv sync
# 运行
uv run ./vis_benchmark.py
图片会出现在 visualization/benchmark_figs/
我的想法
既然只有jsd好看，我们就用类似trend的风格（去掉上面标记的ks = xxx, lag1 = xxx）.这样视觉上相近，又不能说我们错
有更好的可视化方案随意补充