search

sutongkui / SubPhysicsDemo

Demo for Subspace Neural Physics
12 stars 1 forks source link

E:/Github/SubPhysics/ 下文件缺失 项目无法运行 #1

Open flufy3d opened 4 years ago

sutongkui commented 4 years ago

Sorry for the bug and too busy for me to maintain this repo recently. The most important thing is that i found it's not stable, also i make the traning repo public https://github.com/sutongkui/SubPhysics

ODtian commented 1 year ago

How unstable is it

sutongkui commented 1 year ago

There are some weird problems, eg, the simulation will be unpredicted if the input is not covered by the training data

ODtian commented 1 year ago

Thank you very much for your reply, the paper also talks about this limitation, but it is not explained in detail. I would love to use your code to train my own model to verify my idea, but since I don't know much about machine learning yet and my computer performance doesn't allow me to generate a large dataset, I would like to ask you a few more questions.

  1. I would like to ask you how big your dataset is and what rules are used to generate it. Are the trajectories of the balls random or averaged over space? I saw the gif you uploaded showing you interacting with the rabbit at a very fast speed and frequency, does your dataset contain this situation? Does the model work properly when it predicts the situations present in the dataset?
  2. Can the model cope with most of the model clipping problems? Have you tried the depth and shading based solutions mentioned in the paper?
  3. What is the approximate size of the model? Since this approach seems to be only for specific meshes, the space to store these models may be a problem for a real games.
  4. Does the size of the dataset and parameters affect the model size? It seems that this method uses PCA and uses a fixed number of basis, so I guess it may not be affected, but I'm just going by my gut feeling.
sutongkui commented 1 year ago

1, You can access the trainng data as a reference(Please give me an e-mail to get the link) 2, What is model clipping problems? NO for depth ... 3,Small size, it's just 10 layer or even less 4, No affect, you are right, depends on PCA dimensions

ODtian commented 1 year ago

I'm sorry I didn't make it clear, what I mean is the problem of interspersed between model meshes, will physical errors often appear? by the way my email is tianluyue2013@gmail.com.

sutongkui commented 1 year ago

If I understand it correctly, you mean the intersection or penetration between mesh models, that's the collision part in traditional physical simulation, which is very time-consuming. In the maching learnig method, inclduing this one, can't avoid the collision problems. Usually we need some post-processing to handle this.