Life In 19x19

At KGS 5 dan this is the strongest free go AI, but it's the engine only so you still need an interface to play with it.

https://github.com/facebookresearch/dar ... /README.md

pookpooi wrote:At KGS 5 dan this is the strongest free go AI, but it's the engine only so you still need an interface to play with it.

https://github.com/facebookresearch/dar ... /README.md

Ehh... conundrum.

I really *really* dislike Facebook. The service, the company... everything. But I like Go.

Should I have a 'go' at compiling this engine after I'm done testing Oakfoam or not? Hmpf... it's not going to be very easy because I don't know any of the dependencies I wonder if the mentioned Torch framework needs to be installed on the computer that is going to run DarkForest, or if it's going to be built into the program. If it has to be installed, it's going to become a difficult affair.

Looks like you need to have the basic hardware first. This software must run on a GPU with a lots of CUDA cores.

splee99 wrote:Looks like you need to have the basic hardware first. This software must run on a GPU with a lots of CUDA cores.

I have a CUDA-capable card. Maybe not the most powerful card in the universe, but it has 1024 cores. I'll assume it's enough to get the program to run at least.

Ha Babel, I just PMed you to point you here and now I realized, you are already here
I have a GTX970 and i7_2600k and I think (not sure any more) in Hirataku engine both versions (GPU/CPU) were about the same fast on my computer. So it may be worth a compile.,...

Babelardus wrote:

splee99 wrote:Looks like you need to have the basic hardware first. This software must run on a GPU with a lots of CUDA cores.

I have a CUDA-capable card. Maybe not the most powerful card in the universe, but it has 1024 cores. I'll assume it's enough to get the program to run at least.

Believe it or not, for NN work, number of cores is not the primary requirement, it's memory bandwith. This said, I'm going to work on compiling this baby on my i7-4770K / GTX680 SLI machine.

Hi all!

Anyone can compile this project for Windows?

I wonder if Torch 7 can be installed on Windows. I also don't have any experience with the other dependencies this engine has. I don't even know if it can stand on its own, without having to install anything else, after it's compiled.

If someone could make this available for use by laymen such as myself, it would be much appreciated. I'm not lazy, I'm willing to do some work, but I've got no clue how to go about getting this thing playable. I'm curious to play it and see how strong it is vs Leela and Crazystone for example.

I'll take a look at this tomorrow, but I can't promise that I'll be able to get it usable on Windows.

To be really honest, I don't see a good chance to get this engine playable in a reasonable amount of time. For one, it uses dependencies I have no experience with, and I'm not even sure about some of them being able to compile and run on Windows.

This post on one of the development lists says outright that Torch 7 isn't supported on Windows. If that is true, and the package can't be just downloaded and compiled, then it's over already. I can't port Torch to Windows. I don't have the time nor the knowledge regarding machine learning. (I could, for example, try to fix Pachi, so it doesn't hang when compiled with MinGW on Windows, but even things like that require too much time to get done, without thoroughly knowing the program's code.)

While I could install a Linux virtual machine with a full development environment and compile the engine there, that would take even more time. Additionally, it would become a huge engine, as I'd need to statically compile everything into the DarkForest binary, to get it to run on all Linux systems without needing users to install a lot of libraries, which may, or may not be available for their distribution.

As Cygwin has served me well to get software running on Windows that normally will only compile on Linux, I'll try that.

Quoting from Yuangdong Tian’s latest mail to the “computer-go” mailing list:

DarkForest training code is open source now. Hopefully it will help the community.
https://github.com/facebookresearch/darkforestGo

I wonder what the full CUDA requirements are. I have a GTX 780 but that's only "compute capability" SM30 and I have already exceeded that in my own experimentation. (Specifically, SM 3.0 isn't high enough for Dynamic Parallelism - the ability to launch kernels from the GPU itself. I would not be surprised at all if this thing needed that, or newer CUDA features.)

I tried to find the answer but it isn't listed in the README and Google didn't yield it. Unfortunately, my GPU won't work under Linux on my machine, for no apparent reason, so I can't just try and find out.

Charlie wrote:I wonder what the full CUDA requirements are.

3.5 Is what I believe the required cuda compute capability is.

source:
http://timdettmers.com/2014/08/14/which ... -learning/ Note: the blog post date of 2014 is deceiving. Tim regularly updates that post as new GPUs come out and replies to the all the comments so it is up to date.

Just report to all of you guys that Facebook dark forest go is a disappointment for me. I expect it to be a strong amateur dan level AI, but instead,
I think it is quite weak. Based on my testing, it is only a strong kyu level in my setting ( dual core, GTX 1050 GPU) by playing
against 2 DGS bots internally. Facebook dark forest go lost all the games against both in my testing in 4 even games.
Here are 2 bots that I tested against dark forest go:

http://www.dragongoserver.net/userinfo.php?uid=97704

http://www.dragongoserver.net/userinfo.php?uid=97511

Shoreline wrote:Just report to all of you guys that Facebook dark forest go is a disappointment for me. I expect it to be a strong amateur dan level AI, but instead,
I think it is quite weak.

https://arxiv.org/pdf/1511.06410.pdf

"The distributed version, named darkfmcts3 in KGS Go Server, use darkfores2 as the underlying DCNN model, runs 75000 rollouts on 2048 threads and produces a move every 13 seconds with one Intel Xeon E5-2680 v2 at 2.80GHz and 44 NVidia K40m GPUs. It uses top-3 predictions in the first 140 moves and switched to top-5 afterwards so that MCTS could have more choices. Pondering is used. Dynamic komi is used only for high handicap games (≥H5). darkfmcts3 now holds a stable KGS 5d level."

I am not sure what "2048 threads" means because Xeon E5-2680v2 is a 16 thread machine. Maybe they meant 128 cores? That would make more sense as I don't see how you can fit 44 GPUs in one machine.

If that is true, it means it is 5 dan with 128 cores and 44 GPUs. So it is not surprising that with your hardware it is very weak. Maybe AlphaGo would be disappointing as well.