Hi, In terms of involving simulation results of Chrono::GPU in your reinforcement learning project, I would suggest you generate the dataset using Chrono::GPU, and then use the dataset in the training. You probably have to do that anyway, even if Chrono's DEM support is wrapped in Python.
Ruochun On Thursday, November 24, 2022 at 8:23:18 AM UTC-6 Radu Serban wrote: > You are correct: Chrono::GPU is currently not wrapped in PyChrono. > > But you can certainly use smooth contact (that is construct a ChSystemSMC) > in PyChrono. Where/what did you see that made you think that’s not possible? > > > > --Radu > > > > *From:* 'yibing Yan' via ProjectChrono <[email protected]> > *Sent:* Thursday, 24 November 2022 13:37 > *To:* ProjectChrono <[email protected]> > *Subject:* [chrono] Re: the GPU module and the multicore module > > > > > Hi, > > > > Thank you for your advice, after trying for several days, I succeed in > using co-simulation, and it took only 10 minutes every one simulation. I > have another question now, as I said before, I am going to use > reinforcement learning with chrono. I think of using openAI and I also > found projectchrono <https://github.com/projectchrono>/*gym-chrono > <https://github.com/projectchrono/gym-chrono>*. However, I don't know if > I can realize what I have done with *PyChrono*, because the > Chrono::Engine Python module does not cover all the features of the C++ API. > > Meanwhile, in Project Chrono: Install the PYTHON module > <https://api.projectchrono.org/development/module_python_installation.html>, > I found maybe I can't use Chrono::GPU system and Chrono SMC system in > PyChrono. Can you give me any advice? > > Thank you very much again sincerely! > > 在2022年11月11日星期五 UTC+8 11:29:20<Ruochun Zhang> 写道： > > Hi, > > > > What you wanted is possible. There is no problem creating two systems in > one script, the only problem is their interaction. You can do that via > co-simulation. Like a said, *demo_GPU_ballcosim *may give you an idea on > how to do that. The general idea is that your complex mechanical system > involving ChBodies and joints and such, will still be managed by a Chrono > SMC system. Those ChBodies can be instructed to receive external forces and > torques via accumulators; those external forces and torques in this case > should be queried from Chrono::GPU simulations. That is, you load obj > meshes into a Chrono::GPU system, and in each time step, you let it compute > the contact forces between meshes and particles, then feed that information > to ChBodies to update the locations of those bodies, then feed the new > locations of these obj meshes back to Chrono::GPU to complete one time > step, and then repeat. > > > > If your particle system is large I wouldn't recommend doing runtime > visualization anyway, that'd be totally prohibiting in terms of > computational cost. Chrono::GPU can write particles and meshes to files, > and maybe you should generate movies based on those as a post-processing > step. > > > > Thank you, > > Ruochun > > > > On Thursday, November 10, 2022 at 6:41:44 AM UTC-6 [email protected] wrote: > > Hi, thank you for your response! > > First of all, I may have to apologise for my language skills and I really > appreciate that you are willing to listen to my ideas. I'd like to describe > my question again. I have finished a complete complex experiment, including > around 10k granular objects, two objs of my own, one container by using > *utils::AddBoxGeometry* and several links and motors, such as > *ChLinkMotorRotationSpeed*、*ChLinkMotorLinearSpeed*、*ChLinkMateFix*. I > achieved it by using *the MCORE module*, although it can speed up the > program by using multiple threads, it still took 2 hours four one time > simulation. > > > > Now I am considering using the reinforcement learning afterwards, so it is > important to improve the simulation speed. I am wondering if I can use > *Chrono::GPU* only for my granular objects, and all other content > remains the same by using *the MCORE module.* More specifically, I plan > to use both *ChSystemMulticoreSMC* and *ChSystemGpuMesh* in one cpp. The > reason I want to do this is because *Chrono::GPU* is more like a separate > module, some of the functions I need for simulation can't be found in > *Chrono::GPU*, but can be found in *the MCORE module* and are already > realized through my previous efforts. > > > > I also note that *Chrono::GPU *quote ChronoEngine_GPU、 > ChronoEngine_irrlicht、ChronoEngine_multicore、ChronoEngine_opengl、 > ChronoEngine_postprocess、ChronoEngine_robot and *ChronoEngine_multicore(* > *the MCORE > module* *) *is the one which I used now. So I can use MCORE-functions in > cpps in *Chrono::GPU*，by including the .h insteading of modify > the CMakeLists.txt. > > > > I tried and felt if this idea was unattainable. Because I need to create > two systems in one cpp. Meanwhile, for visualization, > *opengl::ChVisualSystemOpenGL* is used in *the* *MCORE module *and > *ChGpuVisualization* is used in *the Chrono::GPU *and I can't find an > interface between the two which means they can't be showed in one window. I > don't know how to solve this problem anymore and I wonder if you have any > suggestions. > > > > I really appreciate your help! > > > > > > > > 在2022年11月10日星期四 UTC+8 13:29:20<Ruochun Zhang> 写道： > > Hi, > > > > I hope someone can provide better help on the linkage issue. What I can > say is that you can try building it with cuda11.6 and the newest gcc. If > you are using cuda11.8, or an ancient version of cuda or gcc, I cannot be > sure. I've been building it with the said configuration with no problem, on > Linux or Windows. > > > > About Chrono::GPU's usage, yes it supports obj meshes. I am not sure about > what you meant by multi-core acceleration. I might, if you elaborate a bit. > And Chrono::GPU should interact with Chrono just fine, for that maybe you > can have a look at the *ballcosim *demo. > > > > If you care about polydisperse spherical particles or complex shaped > particles, then DEM-Engine is the way to go. You can start using it now. > Indeed, documentations are being added. I can drop you a message when it > becomes more accessible. Right now, I attached a snippet from one of my > previous emails, to help you understand how to build this tool on Linux. In > terms of using it, I'd start with checking out and running its demos. And > then, the methods in *API.h* are mostly commented, which for now, may > serve as an ad-hoc documentation for you to understand what some of its > basic usages are. > > Thank you, > > Ruochun > > On Wednesday, November 9, 2022 at 2:19:40 AM UTC-6 [email protected] wrote: > > Hi, thank you very much for your help! > > After last mail, I choosed to use the multicore, the simulation took 2 > hours every time now, because I was constantly adding requests and > features and the number of bodies comes to 10k. Meanwhile, I bought a 3080 > and a new computer. I am considering using the reinforcement learning > afterwards, so it is important to improve the simulation speed. > > I have studied your response several times，I am wondering if I can only > use Chrono::GPU for my particle-related content, and all other content > remains the same, such as multi-core acceleration, loading my own obj. More > specifically, I plan to use both ChSystemMulticoreSMC and ChSystemGpuMesh. > When compile the project, there are errors "LNK2019：Unresolvable external > symbols" in every functions defined in GPU and used in the Muticore and I > feel this is a deeper issue involving linkers. So I would like to ask for > guidance or is there another way to use the Chrono::GPU as a DEMsolver only > for the particles. > > And I learned something about projectchrono > <https://github.com/projectchrono>/*DEM-Engine* > <https://github.com/projectchrono/DEM-Engine>, but I found how to > Install DEM-Engine and the DEM-Engine usage are still waiting to be > added, I would love to try it if I could. > > Thank you again! > > > > 在2022年9月24日星期六 UTC+8 14:45:22<Ruochun Zhang> 写道： > > Hi, > > > > The GPU module does benefit a lot from more recent hardware. If your test > case does not feature a huge number of bodies, say some 10k, then multicore > can be a good choice. It probably requires less learning from you too. > > > > It should be noted that Chrono::GPU is not "Chrono on GPU". Most Chrono > core classes and methods cannot be used in Chrono::GPU. For all > purposes, Chrono::GPU can be seen as a standalone DEM solver for > monodisperse spherical particles, implemented on GPU. It should be used to > simulate granular materials, and it can interact with Chrono (core) so that > it becomes possible to bring a small number of more complex objects (such > as your spoon) into the simulation as well. So if you would like to use > Chrono::GPU, you have to start from its demos, to learn how to use its own > methods to instantiate and manage granular particles. Chrono::GPU's main > advantage is being fast. If your simulation has to involve millions of > granular particles, then multicore will not do and GPU is the choice. > > > > More specifically, *CreateCylindricalContainerFromBoxes* is not a > Chrono::GPU thing at all. *cohesion_ratio* is about the cohesion between > Chrono::GPU particles, and it has nothing to do with gravity, which is set > by *SetGravitationalAcceleration* in Chrono::GPU. > > > > On a different note, Chrono's DEM/granular support on GPU is moving > towards a new direction. The support for complex granular particle shapes > will be added and it will become a duo-GPU solver. It will be based on SBEL's > new DEM Engine <https://github.com/uwsbel/DEM-Engine>. Apart from being > more general and having higher efficiency, the usage of it is similar to > Chrono::GPU, as a standalone helper to Chrono core which manages the > granular part of the simulation, or work on its own as a dedicated DEM > solver. If from the previous conversation you believe Chrono::GPU is for > you, then likely this package will be of interest. More documentations and > user guides are being added to it. But again, it does benefit from recent > GPUs though. > > > > Thank you, > > Ruochun > > > > On Sunday, September 18, 2022 at 2:13:24 AM UTC-5 [email protected] wrote: > > hello there, > > I am tring to perform a simulation which I want to put an object like a > spoon and scoop the sand(granular objects) to see the force in the process. > > I find that demo_GPU_mixer.cpp demo can be referenced, but I am not sure > whether I need to add a container holding those sand and give the granular > objects gravity. If this is so, should I use > *CreateCylindricalContainerFromBoxes* to add container and use > *cohesion_ratio* in the .json to add the gravity? Also, I am sad that I > have a poor GPU so that it really take a long long time to run the gpu > module. > > Then I find that there are also some granular objects demos in the > multicore module, I find some demos about a container with granular > material. And I can run those demos faster. > > Now I am confused about what thing to do next is much better. > > > > Any help will be appreciated, thank you so much in advance. > > -- > You received this message because you are subscribed to the Google Groups > "ProjectChrono" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to [email protected]. > To view this discussion on the web visit > https://groups.google.com/d/msgid/projectchrono/8e6dcb70-2812-4b15-aa00-845f77b9c67dn%40googlegroups.com > > <https://groups.google.com/d/msgid/projectchrono/8e6dcb70-2812-4b15-aa00-845f77b9c67dn%40googlegroups.com?utm_medium=email&utm_source=footer> > . > -- You received this message because you are subscribed to the Google Groups "ProjectChrono" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To view this discussion on the web visit https://groups.google.com/d/msgid/projectchrono/dc4f9e8e-3854-4d7c-8107-020514988bb2n%40googlegroups.com.
