OrbisLumenの部屋

Blog Log

Sat, 11 Apr 2026 00:00:00 GMT

Update log 5 (2026-09-24)

Transfer the mizuki to the shirone repo, seperate the content to private repo. Employ my blog using vercel, with a new domain name orbislumen.blog. Add comment module twikoo with mongodb.

Update log 4 (2026-04-12)

Import music from Netease Music.

Update log 3 (2026-04-12)

Change the diary function's work style.

Here is the detail update log

Update log 2 (2026-04-12)

Open the comments function.

Update log 1 (2026-04-12)

Open the diary function, and write an example for diary.

Initial (2026-04-11)

Finish the post and animate functions, change the whole blog's style for myself and close diary, projects, skills, timeline, albums and devices functions.

机器学习（暑假）

Sun, 28 Jun 2026 00:00:00 GMT

项目介绍

::github{repo="OrbisLumen/ml-tenweeks"}

学习日志

第一周

看李宏毅教授视频公开课复习机器学习部分内容，包括 ChatGPT 原理剖析、机器学习模型的基本概念、了解生成式人工智能的基本工作方式。复习git相关内容，整理部分git命令。

第二周

机器学习基本原理学习，生成式学习与工具型 AI，深度学习基础，梯度下降与优化，CNN 与图像分类。 Linux 基本命令学习。

第三周

学习大语言模型的 Finetuning 与 Prompting，了解 In-context Learning、Instruction-Tuning、Chain of Thought、Soft Prompt 等方法。入门 PyTorch，梳理数据加载、模型训练、验证、测试与参数保存流程；阅读并修改 HW1—HW3，成功运行 HW1，HW2 与 HW3 因数据集链接失效未能完成训练。

第四周

学习大模型、大数据与能力涌现，理解模型规模、训练数据和算力分配对能力的影响。了解 VAE、Flow-based Model、Diffusion Model、GAN 与 Stable Diffusion 等图像生成模型，并整理 Self-attention、Transformer 和生成模型相关课程作业材料。

第五周

学习 NLP 基础、词向量与 Word2vec，理解从 one-hot 表示到连续向量表示的变化，以及基于上下文共现学习词义的基本思想。了解 Constituency Parsing 与 Dependency Parsing，并继续梳理语言模型、RNN、Transformer、预训练模型和多模态学习等方向。

第六周

学习翻译、问答、自然语言生成与共指消解任务，结合翻译场景理解 Seq2Seq 和 Attention 的作用。围绕词嵌入、序列模型、情感分析、关系抽取与大模型应用规划 NLP 实战演练，后续仍需补充代码与实验记录。

第七周

学习 T5 与大语言模型，进一步理解预训练、指令微调和人类反馈对齐之间的关系。了解 LangChain、RAG，以及全量微调、P-Tuning、Prompt Tuning、LoRA 等适配方法，并以 Hugging Face 文本分类和问答任务作为后续实践方向。

第八周

学习知识增强语言模型、模型分析与可解释性，以及 NLP 应用中的偏见、幻觉、隐私和安全问题。了解语义检索、智能问答与情感分析系统，梳理外部知识接入、检索增强和 PaddleNLP 实践的基本流程。

第九周

学习模型压缩基础，了解参数量、计算量、存储开销和推理延迟之间的关系。梳理知识蒸馏、结构化与非结构化剪枝、参数量化等方法，并阅读将模型压缩视为约束优化问题的相关综述论文。

第十周

学习模型压缩代码的整体实操流程，包括模型剪枝、知识蒸馏、低秩分解与参数量化。重点梳理基线模型、评价指标、压缩方法和结果对比之间的关系，后续实验需记录压缩前后的精度、参数量、计算量与推理速度。

动手学深度学习v2

Tue, 28 Apr 2026 00:00:00 GMT

Notice

::github{repo="OrbisLumen/ml-Li-Mu-"}

李沐动手学机器学习

Learning Track

5.6

学习了数据操作，数据预处理，对数据有了初步认识

5.7

复习了线性代数和微积分相关知识

5.8

学习了矩阵计算，这个分子布局和分母布局的不同有点搞心态哦

学习了自动求导相关知识，了解了计算图，和相关函数

5.9

学习了线性回归的数学问题

5.10

学习了线性回归的代码实现，还是比较困难的，借助LLM理解了代码

对与线性回归的最底层的数学原理暂时没有深入理解，但是至少知道了根据梯度更新参数的过程

(如果当前参数让损失变大，就往相反方向调整一点)

5.17

这周事情好多，我机器学习进度推进就放在了周末，结果昨天没commit，今天中午想着discard jupyter notebook的更改时手贱把所有更改搞没了，现在是十一点整，我有点绝望地看着电脑，然后重写笔记

下午把笔记补回来了，然后学习了多层感知机

5.18

学习了模型选择，欠拟合和过拟合

5.19

学习了权重衰退和丢弃法

5.20

学习了数值稳定性

5.21

完成了从全连接到卷积的学习，包括通道，填充，步长，池化层，LeNet

5.22 - 5.24

学习了现代卷积神经网络，包括AlexNet，VGG，NiN，GoogLeNet和ResNet 学习了批量归一化的知识，同时需指出残差思想的重要性，利于快速更新数据端参数

5.25

学习了硬件相关知识，知道了并行加速之类的问题

5.26 - 5.29

学习了计算机视觉的微调以及基于锚框的目标检测系列算法

5.30 - 5.31

学习了循环神经网络

6.1 - 6.3

学习了现代循环神经网络，包括 GRU LSTM DeepRNN BiRNN 学习了机器翻译的实现，使用seq2seq实现和编码器解码器架构

6.4 - 6.5

学习了注意力机制，Transformer 学习了 BERT 和 BERT fine-tune

6.6

学习回顾了优化算法，完结撒花 🎉🎉🎉

Summary

首先这边感谢我们学院的戴瑞婷导师给我推荐了机器学习的入门课，其次感谢李沐大神的录制，这个动手学机器学习的课程我总共耗时一个月多一点看完，期间手敲了大量的笔记，至于代码实现虽说是在md里敲的，但是也基本都过了一遍，但是有些过于难的还是放弃了。

总体来说，这个课程对于学习了线性代数，微积分和python基础（cs61a）的我来说确实有一点吃力，概率论没学很多问题都是请教llm老师，很多理论推导部分都是请llm帮我解释直观理解的，还有代码实现部分有很多包都没有学过怎么用，通过读文档不会什么找什么的方式坚持啃了下来，收获也算颇丰。

这个课程的学习让我了解了机器学习在干什么，解决什么问题，以及为什么机器学习处于时代风口上，让我初步认识了一些模型和理念，虽然李沐大神常说应用先于理论，很多东西都是实践探知出来的，但是我也清楚地认识到了理论基础的重要性。

后续我会继续在这方面跟着导师努力学习，说句实话，虽然有些地方暂时没有搞懂，但是我对这个研究领域还是颇有兴趣的。

Building Blog

Sat, 11 Apr 2026 00:00:00 GMT

这是我的第一个个人博客，花了我一整天的时间...

心路历程

已经学了不少东西了，所以就想搞一个个人博客，之前碰过一点前端的东西，不过已经忘了差不多了，于是我就去b站上找了找，找到了这个伟大的开源项目，搞出来了属于我自己的个人博客，还是非常开心的

具体实现

环境配置

需要node.js，然后又跟着教程配了一堆我看不懂的框架什么的，终于可以在localhost实时观看页面修改（热修改）了

页面修改

跟着教程和文献看了一下这整个项目使用的框架规则，还是比较容易上手的，基本上会写md就行，其他格式遵照项目提供的规则即可生成一个基本还不错的个人博客

公网部署

Cloudflare牛逼！这边也是直接push到fork的github个人仓库，然后cloudflare连接，就可以部署公网，生成一个静态页面

一些感想

属于是听从学长的意见，一个学软工的，怎么能没有个人博客？

所以花了这么些个时间折腾出一个还算可以看的东西，后续其他功能我也没关，只不过404了，有时间以后就会实现，现在只实现了post功能

现在才发现，不能随波逐流啊，所谓“前端已死”，实则并非，你可以不学，但不能不会，全栈依然生命力旺盛，至少在我这个半入流的SE学生看来，只使用前端基础加vibecoding根本不可能写出如此优雅的网页

最后的最后

啊啊啊啊，期中，啊啊啊，绩点，啊啊，六级，啊，CS61B

好累，好累，好累...

CS61B

Fri, 10 Apr 2026 00:00:00 GMT

Notice

::github{repo="OrbisLumen/cs61b"}

CS61B 2024 Spring

Learning Track

4.10

Wasted my whole afternoon and evening setting up the IntelliJ environment...

Finally found that it was neither an IntelliJ version problem nor a JDK version problem...

I only needed to open one project at a time to build the correct debug environment.

->_<-

4.12

The autograder was annoying and made me change my GitHub repository structure several times.

4.13

Finished Project 0.

Since it was a FA19 project, some libraries were deprecated, which made it annoying to deal with.

What a pity — I couldn't see the universe animation...

Er...

Finally decided to switch from the FA19 skeleton to the SP24 skeleton.

The reasons were mentioned above.

4.15

Lab 3 made me speechless...

Tell me how to rewrite a pile of shit?!

4.19

Finished Project 1A and got a glimpse of test-driven programming.

4.20

Finished Project 1B and built a deque with an array list.

I improved my ability to write tests and even found a bug that GPT didn't notice at first.

4.21

Finished Project 1C, extended the deque methods, and used deques in practice.

Started the data structures part of CS61B.

4.24

Finished learning the lectures on asymptotics, disjoint sets, BSTs, B-trees, and red-black trees.

Started the homework and labs to review these ADTs.

Skipped Lab 4, which was about GitHub.

4.26

Finished HW2, Percolation, and learned the application of Quick Union.

Finished Lab 5, Disjoint Sets, and implemented Weighted Quick Union with Path Compression.

4.28

Finished Lab 6, BSTMap, and completed the remove method with GPT's help.

4.30

Finished Lab 7, LLRB Trees. The base case was difficult to figure out, so I completed it with GPT's help.

5.9

Finished lec21, started the graph part.

5.11

Finished lec22, 23, 24, 25, learned the dijkstra's algorithm.

5.15

Finished lec26, 27, 28.

5.16

Started the last unit with sorting and algorithm.

5.18

Finished selection, heap, merge, insertion sort.

5.19

Learned quick sort and the theoretical bounds of compare sort.

5.20

Reminder: skip the software engineering III and IV, plz watch it after proj 2.

Reminder: skip the last three lec.

Reminder: proj 2, 3 remained uncompleted.

acm journey

Sat, 04 Apr 2026 00:00:00 GMT

Overview

::github{repo="OrbisLumen/acm-quick-exprience"}

This is a brief record of my first exposure to ACM and ICPC through the “easy course” ACM-ICPC Algorithms and Programming, which I took to earn two credits.

During the course, I learned a lot and met a master named c1ampy.

Thoughts

Through this experience, I learned what ACM is and got a glimpse of competitive programming.

I was initially ambitious about achieving something in this field. However, considering the time investment required and my own priorities, I decided to give it up.

This is my short journey with ACM.

diary

3.2

After talking with an upperclassman named Izumi Sagiri, I decided to take this course to earn the credits.

I learned that everyone must participate in the university contest, and if you qualify for the finals, you can get 20% of your course grade from it (which basically guarantees the credits). So I invited c1ampy to form a team. To my surprise, he agreed.

From then on, I started learning the basics of C++, getting a quick overview of STL, as well as topics like Data Structures, Graph Theory, Dynamic Programming, and some CS mathematics.

3.6

Learned the basics of C++ (mainly STL).

3.10

c1ampy brought in another strong teammate, fanjunwen, and our team was officially formed.

3.13

c1ampy realized that as a beginner, I couldn’t quickly master full implementations. So instead, he focused on explaining common algorithms at a high level.

This way, I could still contribute ideas during the competition.

3.16

We decided on our team name:
「想不出队名还要去问 LLM 的屑队」

3.27

We participated in the warm-up contest.

Oops! c1ampy noticed there was an interactive problem and taught me how to approach that type of question.

3.28

The preliminary contest began!

08:55 — c1ampy asked me to start from the beginning of the problem set, while fanjunwen started from the end.
09:00 — I found Problem A too difficult, so I switched to B.
09:20 — I decided to solve B using recursion.
09:31 — Boom! Time Limit Exceeded. I changed my approach and used memoization (DFS + cache).
10:08 — Nice! Accepted. Then I moved on to the problem assigned by c1ampy.
10:20 — I realized I could use binary search for the case of 2n, but didn’t know how to handle 2n+1. I told c1ampy, and he solved it.

After that, I didn’t contribute much. But c1ampy coded a lot and carried us into the finals.

Congratulations! Not only did I get 20% of the course grade, but I also got the chance to visit another UESTC campus and spend a whole day there — basically just having fun. Haha~

4.5 — The Final

We made it to the finals.

It was, without exaggeration, one of the happiest days in this short ACM journey.

Before everything began, c1ampy told me that a formal ACM/ICPC-style contest usually runs from around 9:00 to 14:00. That means you get a few intense hours to think with a short window to eat.

So, in preparation for both the contest and the inevitable hunger, I bought some McDonald’s. (And yes, I still firmly believe KFC should just hand over egg tarts to McDonald’s and retire from that competition.)

After picking up our contest T-shirts and checking in, we sat down.

The finals were about to begin.

Timeline

11:30 — c1ampy started setting up the coding environment, completely ignoring the rule:
“Do not touch the mouse or keyboard before 11:55.”
Honestly, how is anyone supposed to get ready in just five minutes?
12:00 — We split our strategy.
I started from the beginning of the problem set, while fanjunwen attacked from the end.
12:30 — I had an idea for Problem C.
I shared it with them, and they asked me to double-check its correctness.
12:40 — I became reasonably confident in my approach.
Not perfectly rigorous—but convincing enough.
Then fanjunwen casually found an O(1) solution.
12:50 — Problem C: accepted.
While reviewing submissions, I realized we had also solved Problem I.
That gave me a burst of confidence, and I moved on.
13:30 — After hitting a wall on several problems, I finally gave in and took a break to eat.
Survival first.
14:00 — Problem A looked approachable.
I committed to it.
One hour later, I gave up… (It was one of the hardest problems in the set.) >_<
~15:00 — Problem F looked easy.
But we just couldn’t get it accepted.
Something felt wrong.
I went through c1ampy’s code line by line—
and it looked absolutely correct.
16:00 — The scoreboard froze.
And we still hadn’t solved F.
At that point, I decided:
if I couldn’t spot the bug directly, I would simulate inputs manually and compare outputs to hunt down the edge case.
16:30 — It worked.
Pen-and-paper debugging saved us.
We finally fixed the issue.
16:55 — I felt like I had done everything I could.
Meanwhile, c1ampy kept pushing on Problem J…
but in the end, we had to let it go.
17:00 — Contest over.
17:10 — Editorial session.
And yes—Problem A was exceedingly brutal.
17:30 — Rank reveal and awards ceremony:
- 1 problem → Third Prize
- 3 problems → Second Prize
- 5 problems → First Prize

Aftermath

It was a final full of pressure, frustration, breakthroughs, and small moments of clarity.

Not perfect.
Not easy.
But deeply satisfying.

And to end the day, we treated ourselves to a bowl of Japanese noodles.

Absolutely worth it.

CS61A

Mon, 30 Mar 2026 00:00:00 GMT

Notes

CS61A 2024 Fall

I have viewed all the lec. For some lec's practice, I have coded it.

I have finished all the project. But HW after 09 (included), LAB after 11 (included) and DISC after 11 (included) haven't been done yet. Since it is for SQL.

From 2026.1 to 2026.3

Overview

This repository is used for self-learning.

::github{repo="OrbisLumen/cs61a"}

This repository contains all homework (hw), labs, and projects for the UC Berkeley CS61A course.

Forked from here. Thanks!

For reference and study, you can also access the CS61A website archive mirrored here.

Features

Homework (hw): Reinforce lecture topics with conceptual and coding exercises.
Labs: Guided practice to understand core concepts.
Projects: Comprehensive assignments to apply knowledge in larger-scale coding challenges.

サクラノ詩－櫻の森の上を舞う－

Wed, 14 Jan 2026 00:00:00 GMT

序

如是一切是我中的众生，故也是众生各自中的一切 ——春与修罗·序

随笔

（剧透向）

PicaPica

——天才的背面，是用一生练习“普通”的演技

“草薙，你要记住。正因为言语有着期限才具有价值。无限期的言语毫无意义，就跟不存在一样。所以约定才有其价值啊。”鸟谷真琴是舞台上的天才演员，却也是生活里最疲惫的伪装者。她的“演技”并非为了扮演他人，而是为了扮演一个“普通”的自己，以此掩盖那份让她与世界产生距离的敏锐与孤独。她精准地模仿着“应有的反应”，却恐惧于流露真实的“自我”。

《樱之诗》中，她与直哉的关系，本质上是两个“天才的囚徒”在互相试探：她看穿了他用懒散掩盖的伤痕，他也触及了她完美笑容下的裂痕。在共通线中，她的直觉总是领先于任何人，她的艺术作品总能带给直哉惊喜。她的角色迫使我们去思考，当“成为自己”本身就需要耗尽毕生的演技时，那份被小心翼翼藏起来的、名为“真琴”的真实，究竟有多么珍贵，又有多么脆弱。

Olympia

——她的世界是一幅已完成的名画，而我愿意永远驻足凝视

“直哉给了我的这条命，要我轻易地就放弃它，想都不要想。”御樱禀的存在本身，就是一件完美的艺术品。她仿佛活在时间之外，心境澄澈如古池，她的世界自给自足，圆满无缺，如同画廊里一幅业已完成的传世名画。但是她不知道的是，名为Olympia的玩偶，名为梦的世界，是直哉，雫，父亲，甚至是她自己一手搭建的。

禀她生来就是天才，“正因为有才能，便能够轻易到达那个地方。正因为能够轻易做到，所以才轻易放弃。”也许她刚出生出生就能看透世界的本质，但在母亲离世的一刹，她选择用奇迹来覆盖奇迹，而奇迹的千年樱扭曲了一切，让禀永远活在了千年樱的幻梦中。她找到了记忆，找到了真相，找到了意志，在一个世界中，她用尽一生来感谢拯救她的直哉。

但是在另一个世界，她深知，为了直哉，为了夏目圭，她不能止步不前。但是，正如others最后自己所说的那样“不落地的翅膀，无论多么强壮多么高洁多么美丽，终究是虚幻之物。就算存在着神明，若是没有大地的话，也会失去翅膀吧。”“平凡地天才展示才能，真正的天才让人忘却才能。”她最终离开了直哉身边，走向了世界，承担起了作为天才的使命。

ZYPRESSEN

——以科学的轨迹计算未来，却将心动变成了最大的变量

“爱上一个人即意味着会伤害到一个人，因为爱就是这么一种庞大的感情。所以说，不能因为害怕会伤害到某人，就逃避自己的真实心情，必须要好好回应这份爱。”冰川里奈是理性的化身，她相信星空可测，万物有律，人生亦当如程序般按部就班地走向光明的未来。她是直哉与那个充满诗性与癫狂的艺术世界之间，最坚实的“地心引力”。

然而，科学无法计算的，是名为“心动”的变量。她对直哉的感情，如同一道无法纳入公式的奇异吸引子，打乱了她所有预设的人生轨迹。她的挣扎与可爱，正在于这种“失控”——用最严谨的态度，去分析最不严谨的情感。最终，她或许没能完全理解直哉艺术的深渊，但她用自己笨拙而真诚的方式，证明了爱本身就是一种最伟大的非理性科学，它的结果无法预测，却能真实地改变两个人（甚至三个人）的宇宙。

Marchen

——最轻盈的脚步，踏出了命运交响曲中最坚定的节拍

“挚爱之人死去的时候，我必须杀死自己。挚爱之人死去的时候，除此之外，别无他法。”川内野优美，这个总是带着童话般轻盈气息的女孩，却踏入了最沉重的命运乐章。“Marchen”即德语“童话”，而她本人也像一则闯入现实的美好故事。

“人是肮脏的生物，所以喜欢美丽的事物，讨厌丑陋的事物。”她以舞蹈般的姿态生活，用积极与善良感染着周围，仿佛命运的悲怆交响与她无关。但正是在这份轻盈之下，藏着惊人的坚定。她证明了，对抗命运的未必是怒吼，也可以是一首不曾停歇的、温柔而充满生命力的舞曲。千年樱闪烁的那一刻，或许是对直哉和冰川的伤害，但那让我们看到了她作为一匹被毒蘑菇诱惑的狼的全部野性，蕾丝边最终得到了最心爱女孩最诚挚的回应。

“那么诸位，乐而不淫哀而不伤，按照正确的拍子，握手吧。即是说，我等欠缺之物，正是忠厚正直与相互理解。嗨，那么诸位，嗨，请一起——按照正确的拍子，握手吧……”

A Nice Derangement of Epitaphs

——将全部温柔炼成钢铁，只为守护那个拒绝世界的背影

“如果什么都无法感觉，就会连是生是死都不知道了，正因为活着，所以才会痛苦，正因为身体想要活，所以才会痛苦。”这章讲述的是草薙一族，夏目一族守护雫的故事，直哉放弃9亿遗产，制造6幅赝作，作为老爹的墓志铭，亦作为保护身为伯奇的雫的最后一道盾牌。

看似雫只是个被保护的对象，看似与雫毫无关系，实则并非如此。夏目雫是将“温柔”这门艺术，淬炼到极致乃至变形的人。雫是冷却的钢，沉默地吸收所有高温，隔绝所有伤害。她的爱是绝对的、排他的、甚至带有一丝非理性的献身感。“A Nice Derangement of Epitaphs”（一句优美的墓志铭错乱）这个充满悖论的标题，同样也精准描述了她：她为自己书写的人生墓志铭，注定与“自我”无关，全部是关于她所守护的那个用尽一切守护她的人。她的故事令人心碎地询问：当温柔变得如钢铁般坚硬而偏执，它究竟是爱的极致，还是对自我的一种悲壮埋葬？

What is mind? No matter. What is matter? Never mind.

——我将终我一生爱着你

“虽然让你感到痛苦的，是你的心，但是让你感到幸福的，也是你的心。不能嫌弃心灵，要是做了这种事，心灵会消失不见的。也不要嫌弃自己的身体，做那种是只会使你自己感到痛苦。”

“心灵不是身体，身体也不是心灵，但是，心灵便是身体、身体便是心灵。虽然不同，但却相同。”奥林匹亚中妖艳的妓女，是对水菜身体与心灵的救赎。谜语般的标题，完美概括了草薙健一郎与夏目水彩之间那场焚烧灵魂的爱情。对他们而言，艺术的本质（mind）与现实的桎梏（matter）在绝对的激情面前，皆可抛却。健一郎是掠夺美的“火焰”，而水菜是他唯一想用来盛放这火焰的“容器”。他们的爱建立在对“美”的极致共感与共同毁灭的倾向上，超越了世俗伦理、家庭责任甚至自我存续。水彩理解并接纳了健一郎艺术中那残酷的、吞噬一切的本质，并自愿成为他艺术的祭品与燃料。

他们的故事叩问：当爱情与艺术狂热交融至不分彼此时，它带来的究竟是创造还是毁灭？他们的答案是：创造即毁灭，毁灭即永恒。最终，水彩的逝去成为健一郎艺术中无法愈合的“绝对缺失”，而这缺失本身，构成了他作品中令人战栗的“绝对存在”。他们的爱情，是一场用生命实践的美学殉道。于是乎，《卧樱》问世……

“假象中铺满天空的春色，仅仅是无拘无束的因果交流电灯，在明亮的、明亮的、明亮的闪耀着。Watermelon的电气石，音与语言的交叉点，在这日益增色的世界中。七朵樱花于眼前飘过，追寻着我的目标，追着追着便消失不见。轻纱笼罩的世界在消失。轻松飘摇的樱之森中鸣奏着世界，是用美妙的音色所鸣奏着的世界。它若是虚无的话，虚无本身即是如此。”

The Happy Prince and Other Tales.

——小鸟不会飞向南方，它会死在王子身旁

“即便如此，我也会一直等下去。所以哪怕不是现在，也没有关系。我想和直哉一起画画，仅此而已。”圭之于直哉，正如那只放弃南飞、最终冻死在快乐王子脚下的燕子。圭拥有敏锐的感知与温柔的灵魂，他最早察觉直哉才华的重量与其下的伤痕，并试图用自己全部的陪伴去温暖那颗被天才与宿命冻结的心。

然而，直哉是那座不断剥落自己（承受父亲阴影、艺术使命、情感纠葛）去试图“照亮”或“救赎”他人的“王子”。圭的守护是静默而无力的，他无法真正将直哉从命运的旋涡中拖离，只能一次次目睹他走向更深的痛苦与更炽烈的燃烧。

这则童话的悲剧性在于：燕子（圭）的牺牲，并未改变王子（直哉）必然倾覆的命运，反而成为这命运中最令人心碎的一笔注脚。圭的“不会飞向南方”，意味着他放弃了属于自己的、可能更轻盈温暖的人生轨迹，选择留在直哉沉重而寒冷的命运现场。他的情谊，是一场明知徒劳却依然献上的、关于“理解”与“陪伴”的壮烈仪式，最终成为直哉心中另一道无法弥合的、温柔的伤口。

当那两朵娇艳的向日葵登上展台亮相的一刹，直哉发自内心地共鸣，“艺术真是残酷呢。梦会吃人，吃下的梦会结出名为艺术的果实。成熟的果实落在地上，给人们带来滋润。人们都不知道那是怎样的果实，却只称赞它的甘甜。”

櫻の森の下を歩む

——你的世界没有尽头，我的停留便是全部

“我认为，人与人的机缘交会，要比任何事情都重要，因为一个人能做到的事情是有限的。”

在真结局樱花覆盖的螺旋坡道上，夏目蓝与草薙直哉完成了这场长达十年的漫长共行。这不是传统意义上的爱情终点，而是两个被命运灼伤的灵魂，最终找到了彼此作为容器的形状。蓝曾是最炽烈的火焰，以燃烧自身为代价向世界投掷艺术的匕首；直哉曾是最沉默的容器，承载着父辈的罪与美却不敢拥有自己的形状。而在这条坡道上，火焰学会了不熄灭地温暖，容器学会了不破碎地盛放。他们之间的陪伴，是对“天才必须孤独”这一诅咒的温柔反驳。

直哉没有成为第二个健一郎去“掠夺”蓝，而是成为了第一个草薙直哉——选择“守护”这份危险而璀璨的美，允许它存在，却不占有；注视它燃烧，却不逃离。

“浪漫是最难言的停留，陪伴是最长情的告白”——蓝最终停留在了这个世界，不再将死亡视为艺术的必然归宿；直哉停留在了她的身旁，不再将逃离视为自我救赎的唯一路径。他们的浪漫，在于接受了彼此生命中的残缺与疯狂，并将这种接受本身，变成了比任何艺术宣言都更坚实的日常。樱花年年飘落，而他们选择年复一年并肩走过，这便是对无常命运最平静、也最有力的胜利。在樱之森下，所有激烈的情感和痛苦的过去，最终都沉淀为脚下沙沙的脚步声。没有拯救，只有共同承担；没有治愈，只有相互容纳。这或许就是《樱之诗》最终想诉说的——爱不是奇迹的解药，而是与伤痕共存的勇气。

“欢迎回来，蓝。”

pvz-for-learning

Mon, 22 Dec 2025 00:00:00 GMT

Note

This is a cpp program using visual studio

::github{repo="OrbisLumen/pvz-for-learning"}

Introduction

This is my very first cpp project.<br>

To say honestly, it's just a homework for my college course Fundamentals of Programming and Algorithms Ⅰ in the first year in UESTC.<br>

By the way, we are required to finish this hw in corporation. However, as a freshmen both in college and in the field of software engineering, I haven't learned that team collabration is very important.<br>

As a result, I finish this by myself while watching the videos by rock.

Self-thinking

For first project, I think it is okay.<br> It can be seen as the start of my cs journey

Content

knowledge

basic knowledge of cpp
process control
function calls

Update Log

change all the sunshine animation control from integer to float for smoother animation
change sunshine ball animation implementation model from offsite to bezier curve abandon unnecessary sunshineball struct menbers (x,y,xoff,yoff,destY)
change sound playing method several times,finally use the method alias
fix there is a ghost image in last mouse clicking position when dragging plants from banner (need to update msg.x and msg.y when left button down in the function "userClick()")
fix only one Peashooter shoot if increase createZombies frequency (add shootTimer for every plants instead of using one "count" for all)
change when to minus sunshine
add most audio by PlaySoundAsync
add zombies eating sound by PlaySound, also by adding soundFlag and soundCount to zombie struct to control
add eatingfre to zombie struct to slow zombies eating animation
fix when you win the game,but the zombies are still created (this is actually a variable conflict,the global zmCount meet the static zmCount) now change the global zmCount to zmCnt
adjust zombies create frequency if (zmCnt <= ZM_MAX * 0.15) {zmFre = 600;} else if (zmCnt <= ZM_MAX * 0.75) {zmFre = rand() % 200 + 100;} else {zmFre = 10;}