Leela-zero project

Isla · #1

Quote:

One reason for publishing this program is that we are setting up a public, distributed effort to repeat the work. Working together, and especially when starting on a smaller scale, it will take less than 1700 years to get a good network (which you can feed into this program, suddenly making it strong). Further details about this will be announced soon.

Anyone interested in participating in?

yoyoma · #2

Nice, this has potential! Thanks to Leela's author, Gian-Carlo Pascutto, for setting this up. I hope it can take off.

I'll look into trying it out this weekend. If anyone else tries it out let us know how it goes!

speedchase · #3

I would like to participate, however this concerns me somewhat. The license in ThreadPool.h is not compatible with the GPLV3.

Quote:

The code is released under the GPLv3 or later, except for ThreadPool.h, which has a specific license mentioned in that file.

Edit: Also, more importantly, I have no idea what this means:

Quote:

OpenCL C++ headers, https://github.com/KhronosGroup/OpenCL-CLHPP (You can just copy input_cl.hpp into CL/cl2.hpp)

jeromie · #4

The license in Threadpool.h seems sufficiently permissive to not cause me any worries about contributing to the project, and since most of the code is licensed under the GPL I don't think there's any concern about contributions to the project as a whole being stolen. Is there anything about the license that gives you pause?

speedchase · #5

My concern is that generally anything distributed alongside GPLv3 code must also be GPLv3. I am not a lawyer, so I have no idea how important this is in this context. I may just be overreacting.

Garf · #6

speedchase wrote:

The license in ThreadPool.h is not compatible with the GPLV3...My concern is that generally anything distributed alongside GPLv3 code must also be GPLv3.

That's simply completely wrong: https://www.gnu.org/licenses/quick-guide-gplv3.en.html

There are a lot of open source licenses compatible with GPLv3, including the license in that file, which seems to be the zlib license. The Khronos headers are MIT/X11 licensed - also compatible. (If it weren't possible to join them then no GPL program would be able to use OpenGL and thus no 3D on Linux!)

Krama · #7

It would probably be too silly of me to expect deepmind to release the weights of their NN.
What would they lose by doing that?
If nothing then why won't they do it?

moha · #8

Krama wrote:

It would probably be too silly of me to expect deepmind to release the weights of their NN.
What would they lose by doing that?
If nothing then why won't they do it?

I think this is not so simple. It's not just the weights that are in question, but several implementation details. DM probably experimented with quite a few NN structures / parameter tweaks before their network reached the levels published. It seems perfectly possible that even with decent expertise the first few implementations will be inferior to theirs. And while they may not lose much by publishing the weights, the implementation details are a different matter. That knowledge is basically the reason Google invested into RL the first place...

(BTW, for this reason I think the public/distributed "Zero" project will also need some kind of experimental/development channel, constantly trying new and new implementations, not just training a single net.)

Leela-zero project

Who is online