Rich Sutton

March 13, 2019

The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin. The ultimate reason for this is Moore’s law, or rather its generalization of continued exponentially falling cost per unit of computation. Most AI research has been conducted as if the computation available to the agent were constant (in which case leveraging human knowledge would be one of the only ways to improve performance) but, over a slightly longer time than a typical research project, massively more computation inevitably becomes available. Seeking an improvement that makes a difference in the shorter term, researchers seek to leverage their human knowledge of the domain, but the only thing that matters in the long run is the leveraging of computation. These two need not run counter to each other, but in practice they tend to. Time spent on one is time not spent on the other. There are psychological commitments to investment in one approach or the other. And the human-knowledge approach tends to complicate methods in ways that make them less suited to taking advantage of general methods leveraging computation.  There were many examples of AI researchers’ belated learning of this bitter lesson, and it is instructive to review some of the most prominent.

In computer chess, the methods that defeated the world champion, Kasparov, in 1997, were based on massive, deep search. At the time, this was looked upon with dismay by the majority of computer-chess researchers who had pursued methods that leveraged human understanding of the special structure of chess. When a simpler, search-based approach with special hardware and software proved vastly more effective, these human-knowledge-based chess researchers were not good losers. They said that “brute force” search may have won this time, but it was not a general strategy, and anyway it was not how people played chess. These researchers wanted methods based on human input to win and were disappointed when they did not.

A similar pattern of research progress was seen in computer Go, only delayed by a further 20 years. Enormous initial efforts went into avoiding search by taking advantage of human knowledge, or of the special features of the game, but all those efforts proved irrelevant, or worse, once search was applied effectively at scale. Also important was the use of learning by self play to learn a value function (as it was in many other games and even in chess, although learning did not play a big role in the 1997 program that first beat a world champion). Learning by self play, and learning in general, is like search in that it enables massive computation to be brought to bear. Search and learning are the two most important classes of techniques for utilizing massive amounts of computation in AI research. In computer Go, as in computer chess, researchers’ initial effort was directed towards utilizing human understanding (so that less search was needed) and only much later was much greater success had by embracing search and learning.

In speech recognition, there was an early competition, sponsored by DARPA, in the 1970s. Entrants included a host of special methods that took advantage of human knowledge—knowledge of words, of phonemes, of the human vocal tract, etc. On the other side were newer methods that were more statistical in nature and did much more computation, based on hidden Markov models (HMMs). Again, the statistical methods won out over the human-knowledge-based methods. This led to a major change in all of natural language processing, gradually over decades, where statistics and computation came to dominate the field. The recent rise of deep learning in speech recognition is the most recent step in this consistent direction. Deep learning methods rely even less on human knowledge, and use even more computation, together with learning on huge training sets, to produce dramatically better speech recognition systems. As in the games, researchers always tried to make systems that worked the way the researchers thought their own minds worked—they tried to put that knowledge in their systems—but it proved ultimately counterproductive, and a colossal waste of researcher’s time, when, through Moore’s law, massive computation became available and a means was found to put it to good use.

In computer vision, there has been a similar pattern. Early methods conceived of vision as searching for edges, or generalized cylinders, or in terms of SIFT features. But today all this is discarded. Modern deep-learning neural networks use only the notions of convolution and certain kinds of invariances, and perform much better.

This is a big lesson. As a field, we still have not thoroughly learned it, as we are continuing to make the same kind of mistakes. To see this, and to effectively resist it, we have to understand the appeal of these mistakes. We have to learn the bitter lesson that building in how we think we think does not work in the long run. The bitter lesson is based on the historical observations that 1) AI researchers have often tried to build knowledge into their agents, 2) this always helps in the short term, and is personally satisfying to the researcher, but 3) in the long run it plateaus and even inhibits further progress, and 4) breakthrough progress eventually arrives by an opposing approach based on scaling computation by search and learning. The eventual success is tinged with bitterness, and often incompletely digested, because it is success over a favored, human-centric approach.

One thing that should be learned from the bitter lesson is the great power of general purpose methods, of methods that continue to scale with increased computation even as the available computation becomes very great. The two methods that seem to scale arbitrarily in this way are search and learning.

The second general point to be learned from the bitter lesson is that the actual contents of minds are tremendously, irredeemably complex; we should stop trying to find simple ways to think about the contents of minds, such as simple ways to think about space, objects, multiple agents, or symmetries. All these are part of the arbitrary, intrinsically-complex, outside world. They are not what should be built in, as their complexity is endless; instead we should build in only the meta-methods that can find and capture this arbitrary complexity. Essential to these methods is that they can find good approximations, but the search for them should be by our methods, not by us. We want AI agents that can discover like we can, not which contain what we have discovered. Building in our discoveries only makes it harder to see how the discovering process can be done.

从 70 年的人工智能研究中可以得到的最大教训是，利用计算的通用方法最终是最有效的，而且是最大的优势。其根本原因是摩尔定律，或者更确切地说是其对单位计算成本持续呈指数下降的概括。大多数人工智能研究都是在代理可用的计算是恒定的情况下进行的（在这种情况下，利用人类知识将是提高性能的唯一方法之一），但与典型的研究项目相比，在稍长的时间内，不可避免地会出现大量的计算可用。为了寻求在短期内产生影响的改进，研究人员试图利用他们对该领域的人类知识，但从长远来看，唯一重要的是利用计算。这两者不必相互矛盾，但在实践中它们往往是相互矛盾的。花在其中一个上的时间就是没有花在另一个上的时间。人们在心理上承诺投资于一种方法或另一种方法。而人类知识方法往往会使方法复杂化，使其不太适合利用利用计算的通用方法。人工智能研究人员迟迟没有吸取这一惨痛教训的例子有很多，回顾一下其中最突出的一些例子很有启发意义。
在计算机象棋中，1997 年击败世界冠军卡斯帕罗夫的方法是基于大规模深度搜索。当时，大多数计算机象棋研究人员对此感到沮丧，他们一直在寻求利用人类对象棋特殊结构的理解的方法。当一种更简单的、基于搜索的方法加上特殊的硬件和软件被证明更为有效时，这些基于人类知识的象棋研究人员就不是善于输的人了。他们说，“蛮力”搜索这次可能赢了，但这不是一种通用策略，而且无论如何它也不是人们下棋的方式。

这些研究人员希望基于人类输入的方法能够获胜，但结果却令他们失望。 计算机围棋也出现了类似的研究进展模式，只是推迟了 20 年。最初，人们付出了巨大的努力，利用人类知识或游戏的特殊功能来避免搜索，但一旦搜索被大规模有效应用，所有这些努力都被证明是无关紧要的，甚至更糟。同样重要的是使用自学来学习价值函数（就像在许多其他游戏甚至国际象棋中一样，尽管学习在 1997 年首次击败世界冠军的程序中并没有发挥重要作用）。自学和一般的学习就像搜索一样，因为它能够发挥大规模计算的作用。搜索和学习是人工智能研究中利用大量计算的两类最重要的技术。在计算机围棋中，就像在计算机国际象棋中一样，研究人员最初的努力是利用人类的理解力（这样就不需要太多的搜索），直到后来，通过采用搜索和学习才取得了更大的成功。

在语音识别方面，20 世纪 70 年代，DARPA 赞助了一场早期的竞赛。参赛者包括大量利用人类知识（单词、音素、人类声道等知识）的特殊方法。另一方面，一些较新的方法更具统计性质，并且基于隐马尔可夫模型 (HMM) 进行更多的计算。统计方法再次战胜了基于人类知识的方法。这导致了整个自然语言处理领域发生了重大变化，几十年来，统计和计算逐渐占据了主导地位。语音识别中深度学习的兴起是朝着这一一致方向迈出的最新一步。深度学习方法更少地依赖人类知识，使用更多的计算，再加上对大量训练集的学习，从而产生了更好的语音识别系统。就像在游戏中一样，研究人员总是试图制造出按照他们认为自己的想法运作的系统——他们试图将这些知识放入他们的系统中——但最终却适得其反，浪费了研究人员大量的时间，而摩尔定律让大规模计算成为可能，并找到了一种充分利用它的方法。

在计算机视觉中，也有类似的模式。早期的方法将视觉设想为搜索边缘、广义圆柱体或 SIFT 特征。但今天所有这些都被抛弃了。现代深度学习神经网络只使用卷积和某些类型的不变性的概念，而且表现要好得多。

这是一个很大的教训。作为一个领域，我们还没有彻底学会它，因为我们还在继续犯同样的错误。要看到这一点，并有效地抵制它，我们必须了解这些错误的吸引力。我们必须学会不那么痛苦.

nvidia-smi返回的是driver所能支持的最新的cuda版本
系统安装的cuda版本可以随意，torch会优先使用虚拟环境中安装的cuda版本

Conda管理Cuda

安装指定版本

1

conda install nvidia/label/cuda-12.4.0::cuda-toolkit -c nvidia/label/cuda-12.4.0  

安装最新版本

1

conda install cuda-toolkit

使用ssh作为命令行远程工具，启动远程的jupyter lab并且在本地的浏览器中打开。

远程运行：

1

jupyter lab --no-browser --port=8080

--no-broswer is very important

output:

1
2
3
4
5
6
7
8
9
10

...
[I 2024-12-10 14:30:24.585 ServerApp]     http://127.0.0.1:8080/lab?token=0061d1eb31396b1bc3cd77a7161b2084da1dedcdeca0600c
[I 2024-12-10 14:30:24.586 ServerApp] Use Control-C to stop this server and shut down all kernels (twice to skip confirmation).
[C 2024-12-10 14:30:24.603 ServerApp]

    To access the server, open this file in a browser:
        file:///home/bohanfeng/.local/share/jupyter/runtime/jpserver-11659-open.html
    Or copy and paste one of these URLs:
        http://localhost:8080/lab?token=0061d1eb31396b1bc3cd77a7161b2084da1dedcdeca0600c
        http://127.0.0.1:8080/lab?token=0061d1eb31396b1bc3cd77a7161b2084da1dedcdeca0600c

本地运行：

1

ssh -L 8080:localhost:8080 bohanfeng@192.168.2.102

本地浏览器访问：
http://127.0.0.1:8080/lab?token=0061d1eb31396b1bc3cd77a7161b2084da1dedcdeca0600c

Repository: https://github.com/owkin/FLamby

Installation

1
2
3
4
5
6
7
8

git clone https://github.com/owkin/FLamby.git
cd FLamby
conda env create -f environment.yml
conda activate flamby
pip install -e .[all_extra]
pip install wget
pip install lifelines
pip install jupyterlab

Dataset

Fed-TCGA-BCRA
https://owkin.github.io/FLamby/fed_tcga_brca.html

Baseline Learning

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

import torch
from flamby.utils import evaluate_model_on_tests

# 2 lines of code to change to switch to another dataset
from flamby.datasets.fed_tcga_brca import (
    BATCH_SIZE,
    LR,
    NUM_EPOCHS_POOLED,
    Baseline,
    BaselineLoss,
    metric,
    NUM_CLIENTS,
    Optimizer,
)
from flamby.datasets.fed_tcga_brca import FedTcgaBrca as FedDataset

Import several macros, datasets and metrics.

1
2
3
4
5
6

# Instantiation of local train set (and data loader)), baseline loss function, baseline model, default optimizer
train_dataset = FedDataset(center=0, train=True, pooled=False)
train_dataloader = torch.utils.data.DataLoader(train_dataset, batch_size=BATCH_SIZE, shuffle=True, num_workers=0)
lossfunc = BaselineLoss()
model = Baseline()
optimizer = Optimizer(model.parameters(), lr=LR)

In this script, the pooled parameter is set to False when creating the FedDataset instances. This indicates that the dataset is not pooled, meaning that the data is kept separate for each client or center. Each client or center has its own local dataset, which is a common setup in federated learning to simulate real-world scenarios where data is distributed across different locations or devices.

1
2
3
4
5
6
7
8

# Traditional pytorch training loop
for epoch in range(0, NUM_EPOCHS_POOLED):
    for idx, (X, y) in enumerate(train_dataloader):
        optimizer.zero_grad()
        outputs = model(X)
        loss = lossfunc(outputs, y)
        loss.backward()
        optimizer.step()

正常的训练流程

1
2
3
4
5
6
7
8
9
10
11
12
13
14

# Evaluation
# Instantiation of a list of the local test sets
test_dataloaders = [
            torch.utils.data.DataLoader(
                FedDataset(center=i, train=False, pooled=False),
                batch_size=BATCH_SIZE,
                shuffle=False,
                num_workers=0,
            )
            for i in range(NUM_CLIENTS)
        ]
# Function performing the evaluation
dict_cindex = evaluate_model_on_tests(model, test_dataloaders, metric)
print(dict_cindex)

使用的evaluation metric是lifelines.utils.concordance_index，返回的是c_index

Federated Learning

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18

import torch
from flamby.utils import evaluate_model_on_tests

# 2 lines of code to change to switch to another dataset
from flamby.datasets.fed_tcga_brca import (
    BATCH_SIZE,
    LR,
    NUM_EPOCHS_POOLED,
    Baseline,
    BaselineLoss,
    metric,
    NUM_CLIENTS,
    get_nb_max_rounds
)
from flamby.datasets.fed_tcga_brca import FedTcgaBrca as FedDataset

# 1st line of code to change to switch to another strategy
from flamby.strategies.fed_avg import FedAvg as strat

1
2
3
4
5
6
7
8
9
10
11
12
13

# We loop on all the clients of the distributed dataset and instantiate associated data loaders
train_dataloaders = [
            torch.utils.data.DataLoader(
                FedDataset(center = i, train = True, pooled = False),
                batch_size = BATCH_SIZE,
                shuffle = True,
                num_workers = 0
            )
            for i in range(NUM_CLIENTS)
        ]

lossfunc = BaselineLoss()
m = Baseline()

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28

# Federated Learning loop
# 2nd line of code to change to switch to another strategy (feed the FL strategy the right HPs)
args = {
            "training_dataloaders": train_dataloaders,
            "model": m,
            "loss": lossfunc,
            "optimizer_class": torch.optim.SGD,
            "learning_rate": LR / 10.0,
            "num_updates": 100,
# This helper function returns the number of rounds necessary to perform approximately as many
# epochs on each local dataset as with the pooled training
            "nrounds": get_nb_max_rounds(100),
        }
s = strat(**args)
m = s.run()[0]

# Evaluation
# We only instantiate one test set in this particular case: the pooled one
test_dataloaders = [
            torch.utils.data.DataLoader(
                FedDataset(train = False, pooled = True),
                batch_size = BATCH_SIZE,
                shuffle = False,
                num_workers = 0,
            )
        ]
dict_cindex = evaluate_model_on_tests(m, test_dataloaders, metric)
print(dict_cindex)

FedAvg vs FedAvgFineTuning

FedAvg

FedAvgFineTuning

Convert Raw RGB-D to tree-structure scene(maybe in unity), for more

1. Raw point cloud (voxel) to semantic segmented cloud
2. classify the segmented object into different structure and establish parent-child relationship
3. identify the state of object
   1. pose like, numbers (\\)
   2. state like, open/close (LLM)

发现和lff近期发表的一篇文章思想非常一致 https://arxiv.org/html/2410.07408v1

和场景理解的对比

Setup

仓库: https://github.com/Simple-Robotics/cosypose

1
2
3

git clone --recurse-submodules https://github.com/Simple-Robotics/cosypose.git
cd cosypose
conda env create -n cosypose --file environment.yaml

注意执行这一步的时候pip 会提示setuptools 和matplotlib-inline不符合3.7.6的python，到环境中手动安装适配的版本

1
2
3

conda activate cosypose
pip install setuptools==63.4.1
pip install matplotlib-inline==0.1.6

1
2
3

git lfs pull
python setup.py install
python setup.py develop

根据README下载数据
注意第一块指令无法下载成功，由 https://bop.felk.cvut.cz/datasets/ 得知下载链接迁移到了huggingface, https://huggingface.co/datasets/bop-benchmark/datasets/tree/main/ycbv 可以从这里手动下载测试集并放置到local_data/bop_datasets/ycbv/test

设置测试使用的models

1

cp ./local_data/bop_datasets/ycbv/model_bop_compat_eval ./local_data/bop_datasets/ycbv/models

Debug

np.where(mask)[0].item()

运行

1
2

export CUDA_VISIBLE_DEVICES=0 
python -m cosypose.scripts.run_cosypose_eval --config ycbv

时出现报错

1
2
3
4
5
6
7
8
9
10
11
12

Traceback (most recent call last):
  File "/home/cyl/.conda/envs/cosypose/lib/python3.7/runpy.py", line 193, in _run_module_as_main
    "__main__", mod_spec)
  File "/home/cyl/.conda/envs/cosypose/lib/python3.7/runpy.py", line 85, in _run_code
    exec(code, run_globals)
  File "/home/cyl/cosypose/cosypose/scripts/run_cosypose_eval.py", line 491, in <module>
    main()
  File "/home/cyl/cosypose/cosypose/scripts/run_cosypose_eval.py", line 332, in main
    scene_ds = make_scene_dataset(ds_name)
  File "/home/cyl/cosypose/cosypose/datasets/datasets_cfg.py", line 68, in make_scene_dataset
    ids.append(np.where(mask)[0].item())
ValueError: can only convert an array of size 1 to a Python scalar

添加debug输出，得到

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19

Debug - scene_id: 48, view_id: 1
Debug - mask matches: 1
Debug - where result shape: (1,), values: [225]
Debug - scene_id: 48, view_id: 36
Debug - mask matches: 1
Debug - where result shape: (1,), values: [226]
Debug - scene_id: 48, view_id: 47
Debug - mask matches: 1
Debug - where result shape: (1,), values: [227]
Debug - scene_id: 48, view_id: 83
Debug - mask matches: 1
Debug - where result shape: (1,), values: [228]
Debug - scene_id: 48, view_id: 112
Debug - mask matches: 1
Debug - where result shape: (1,), values: [229]
Debug - scene_id: 48, view_id: 135
Debug - mask matches: 0
Debug - where result shape: (0,), values: []
0:00:00.912023 - Expected exactly one match, got 0 matches for scene_id=48, view_id=135

发现是下载的测试数据集并不包含数据集keyframe.txt中所有的帧，导致一些关键帧识别不到

运行到一半被终止的情况

如果想重新开始新的训练： 清空local_data/joblib_cache

Framework

prediction script cosypose.scripts.run_cosypose_eval

AI explanation

The script predicts object poses based on multi-view input by following these steps:

1. Dataset Loading: It first loads the dataset using the make_scene_dataset function, which prepares the scene data for evaluation. The dataset is wrapped in a MultiViewWrapper to handle multiple views.

2. Model Loading: The script loads pre-trained models for pose prediction using the load_models function. It loads both coarse and refiner models based on the configuration specified in the command-line arguments.

3. Prediction Setup: The script sets up the prediction parameters, including the number of iterations for coarse and refiner models, and whether to skip multi-view processing based on the number of views specified.

4. Multi-view Prediction: The MultiviewScenePredictor is initialized with the mesh database, which is used to predict poses across multiple views. The MultiviewPredictionRunner is then used to run predictions on the dataset, leveraging the multi-view setup to improve pose estimation accuracy.

5. Pose Estimation: The script uses the loaded models to predict object poses. It processes detections from either pix2pose or posecnn depending on the dataset, and refines these predictions using the refiner model.

6. Evaluation: After predictions, the script evaluates the predicted poses using the PoseEvaluation class. It calculates various metrics like ADD-S and AUC to assess the accuracy of the pose predictions.

7. Results Logging: Finally, the script logs the results, including evaluation metrics, and saves them to a specified directory.

The multi-view approach allows the script to leverage information from different viewpoints, which can help resolve ambiguities and improve the robustness of the pose estimation.

Model dataset for the initialization of MultiviewScenePredictor

In run_cosypose_eval we initialize MultiviewScenePredictor in this way:

1

mv_predictor = MultiviewScenePredictor(mesh_db)

In the MultiviewScenePredictor we use the mesh_db to initialize MultiviewRefinement and solve:

1
2
3
4
5
6
7
8
9

problem = MultiviewRefinement(candidates=candidates_n,
                    cameras=cameras,
	                pairs_TC1C2=pairs_TC1C2,
	                mesh_db=self.mesh_db_ba)
ba_outputs = problem.solve(
	n_iterations=ba_n_iter,
	optimize_cameras=not use_known_camera_poses,
)

The solve function of MultiviewRefinement:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31

def solve(self, sample_n_init=1, **lm_kwargs):
	timer_init = Timer()
	timer_opt = Timer()
	timer_misc = Timer()

	timer_init.start()
	TWO_9d_init, TCW_9d_init = self.robust_initialization_TWO_TCW(n_init=sample_n_init)
	timer_init.pause()

	timer_opt.start()
	TWO_9d_opt, TCW_9d_opt, history = self.optimize_lm(
		TWO_9d_init, TCW_9d_init, **lm_kwargs)
	timer_opt.pause()

	timer_misc.start()
	objects, cameras = self.make_scene_infos(TWO_9d_opt, TCW_9d_opt)
	objects_init, cameras_init = self.make_scene_infos(TWO_9d_init, TCW_9d_init)
	history = self.convert_history(history)
	timer_misc.pause()

	outputs = dict(
		objects_init=objects_init,
		cameras_init=cameras_init,
		objects=objects,
		cameras=cameras,
		history=history,
		time_init=timer_init.stop(),
		time_opt=timer_opt.stop(),
		time_misc=timer_misc.stop(),
	)
	return outputs

Adaption

不急

前摇部分

基本原理

本地增添博客内容(markdown文件)->hexo根据文件内容生成网页源码->上通过指令上传(push)到github->github自行部署静态页面

基本准备

安装git

https://www.cnblogs.com/xueweisuoyong/p/11914045.html

Github shh key

因为把本地写的内容传到github，需要绑定一个ssh密钥
参见：https://docs.github.com/en/authentication/connecting-to-github-with-ssh/generating-a-new-ssh-key-and-adding-it-to-the-ssh-agent

1

ssh-keygen -t ed25519 -C "guanshengyuanlu@163.com"

把这串公钥添加到github ssh settings里面

安装npm

https://blog.csdn.net/lizhong2008/article/details/133844070
最新版本即可

本地部署

设定本地的git config

1
2

git config --global user.email "guanshengyuanlu@163.com"
git config --global user.name "Draumurvakna"

克隆仓库

1
2

git clone git@github.com:Draumurvakna/MIGAO-Blog-Src.git
cd MIGAO-Blog-Src

安装环境

1
2
3
4

git submodule update --recursive --init   
npm update
cd themes/icarus
npm update

更新网站

1

./show.sh #预览

1

./deploy.sh #可以直接通过网页访https://draumurvakna.github.io/

正文

网站上的每一篇文章在本地都是一份markdown文本文件，存在source/_posts中

通过指令hexo new "article title" 来创建一篇新博客

然后到对应文件里面编辑就行了
markdown的编辑器推荐用typora，当然如果足够硬核的话用txt文本编辑器也毫无问题！

如果想添加图片的话，就往同文件夹下的资源文件夹（和这篇博客名字相同的文件夹）中添加照片然后在文中输入

1

可以参考Sample Blog

添加完自己想要的内容之后用./deploy.sh部署一下网站，稍等片刻，进到网站里就可以看到最新的变化了。

后

我想给博客换个背景

进到themes/icarus/source/img文件夹📁

如果重新部署后发现没有更改，那就网页里按一下<Ctrl>+F5

我想换个头像👤

如上图，改avatar.png

关于评论系统的话需要自己搞定o

参考 https://chen-yulin.github.io/2024/09/03/%5BOBS%5Dhexo-Hexo%20Comment%20System%20--%20Twikoo/

研究生层面的罗斯商学院双学位项目
密大方面提供的硕士学位（授课型）：

• 管理 30 credits
• 供应链（管理）30 credits
• 商务分析 (BA) 36.5 credits 挺适合转商科，量化分析（programming required）

密大的一年放在三年学制的最后一学年的6月~来年5月，我们认对方6学分，对方认我们Gateway
准入条件：

• 密院研究生
• 对方的线上面试方式

学费，5w$~6w$ （如果承认学分学费可以打折），生活安娜堡预计1000$/m，饮食1000$/m， 杂项500$/m

可以提供找工作的签证机会（利好留美发展者）
但是时间会和秋招冲突，会给国内找工作面试带来困难

发学位证书的时间在两边并不统一

26级包括专硕招生（和双学位挂钩）

Useful Link

https://www.youtube.com/watch?v=cLWgjienc_s
https://github.com/habamax/vim-godot

Implementation

Nvim

安装vim-godot, 并且不用手动在配置中开启

1
2

{ "habamax/vim-godot", event = "VimEnter" },

在init.lua中，自动识别godot项目并且开启服务器端口

1
2
3
4
5
6
7

-- for godot
local gdproject = io.open(vim.fn.getcwd() .. "/project.godot", "r")
if gdproject then
	io.close(gdproject)
	vim.fn.serverstart("./godothost")
	vim.notify("Open godot project")
end

在dap/nvim-dap/dap-godot.lua中配置了dap debugger

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

local dap = require("dap")
dap.adapters.godot = {
	type = "server",
	host = "127.0.0.1",
	port = 6006,
}
dap.configurations.gdscript = {
	{
		type = "godot",
		request = "launch",
		name = "Launch scene",
		project = "${workspaceFolder}",
		launch_scene = true,
	},
}

其余dap的默认配置沿用之前的。
lsp:

1
2
3
4

lspconfig.gdscript.setup({
	capabilities = capabilities,
	on_attach = on_attach,
})

Godot Editor

参考视频中在编辑器配置中设置外部编辑器即可。

Reference: http://nilssondev.com/posts/my-journey-with-physics

Things Jolt Does Better

• Upfront about allocations. Jolt expects that you tell it how much memory will be needed for the simulation, which definitely results in better performance since Jolt won’t be reallocating a lot of memory during the simulation. (Jolt does provide a way of doing allocations dynamically, so upfront allocations isn’t a requirement)
• Better multi-core support. Out of the box Jolt will utilize multiple cores very well, whereas I’ve found that PhysX doesn’t utilize multiple cores all too well. Jolt was designed with this in mind whereas PhysX wasn’t.(多线程原生支持)
• Explicitly reference counted objects. When you create a new object in Jolt you’ll most likely get a “Ref” instance back. PhysX does utilize reference counting, but it’s not as apparent as it is with Jolt because PhysX returns pointers to most objects.
• Multiple collision shapes per body. Jolt handles this by providing a “compound” shape, which you construct from multiple other shapes, whereas PhysX just lets you add an arbitrary amount of shapes individually to any body. I’m not entirely sure how PhysX handles this internally, but I do know that Jolt computes these shapes into a single more optimal collision shape, meaning that collision tests are most likely more optimal.（最佳碰撞划分）
• Better Multi-Threaded access. Jolt was designed with this in mind, and most methods in Jolt supports being accessed by multiple threads at once, whereas PhysX requires you to make use of a “Task” system, which is poorly designed in my opinion.（多线程）
• Jolt achieves a higher simulation fidelity / better simulation quality out of the box. Jolt gives us some really good simulation values out of the box, whereas PhysX requires us to play around with a lot of parameters in order to get a high quality simulation.

Things PhysX Does Better

• PhysX has better simulation performance out of the box, whereas Jolt requires more careful setup and usage of the API in order to achieve good performance. Jolt does however gives more opportunities for optimizations, and can achieve better performance overall, but it does require more work（物理准确性方面更好）
• PhysX supports GPU-based simulation by utilizing the CUDA architecture, sadly this is limited to Nvidia GPUs only （可以显卡加速）
• The PhysX Visual Debugger (PVD) supports recording the simulation over the network, Jolt requires us to record the simulation to a file
• The PVD doesn’t just visually record our simulation, but also records the parameters of the simulation at any given moment, allowing us to inspect the parameters more easily. Jolt on the other hand only records the simulation visually.
• PhysX offers us more control over mesh collider cooking, giving us more parameters to mess around with
• PhysX has far better documentation overall, since it’s an older and more widely used API there’s more resources on it online, whereas Jolt lacks online documentation. The developer of Jolt is very active on the GitHub repository though, and has been very helpful with any issues I’ve encountered, or any questions I’ve had about the usage. This isn’t the case with the PhysX developers.（文档更完备）
• Achieves better performance when adding a lot of new bodies during the simulation. Jolt requires us to manually optimize the broadphase structure, and doing that when there’s already a lot of bodies in the broadphase can result in slowdowns temporarily.（在运行时创建物理物体更方便（？））