leela-zero /leela-zero
Go engine with no human-provided knowledge, modeled after the AlphaGo Zero paper.
★ 5,594⑂ 1,013👁 283⑃ 659● 2,007▤ 18
+0
日增长
5,594
stars
1,013
forks
283
watchers
659
prs
2,007
issues
18
releases
README
What A Go program with no human provided knowledge. Using MCTS (but without Monte Carlo playouts) and a deep residual convolutional neural network stack. This is a fairly faithful reimplementation of the system described in the Alpha Go Zero paper "Mastering the Game of Go without Human Knowledge". For all intents and purposes, it is an open source AlphaGo Zero. Wait, what? If you are wondering what the catch is: you still need the network weights. No network weights are in this repository. If you manage to obtain the AlphaGo Zero weights, this program will be about as strong, provided you also obtain a few Tensor Processing Units. Lacking those TPUs, I'd recommend a top of the line GPU - it's not exactly the same, but the result would still be an engine that is far stronger than the top humans. Gimme the weights Recomputing the AlphaGo Zero weights will take about 1700 years on commodity hardware. One reason for publishing this program is that we are running 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). I want to help Using your own hardware You need a PC with a GPU, i.e. a discrete graphics card made by NVIDIA or AMD, preferably not too old, and with the most recent drivers installed. It is possible to run the program without a GPU, but performance will be much lower. If your CPU is not very recent (Haswell or newer, Ryzen or newer), performance will be outright bad, and it's probably of no use trying to join the distributed effort. But you can still play, especially if you are patient. Windows Head to the Github releases page at https://github.com/leela-zero/leela-zero/releases, download the latest release, unzip, and launch autogtp.exe. It will connect to the server automatically and do its work in the background, uploading results after each game. You can just close the autogtp window to stop it. macOS and Linux Follow the instructions below to compile the leelaz and autogtp binaries in the build subdirectory. Then run autogtp as explained in the contributing instructions below. Contributing will start when you run autogtp. Using a Cloud provider Many cloud companies offer free trials (or paid solutions, not discussed here) that are usable for helping the leela-zero project. There are community maintained instructions available here: Running Leela Zero client on a Tesla V100 GPU for free (Google Cloud Free Trial) Running Leela Zero client on a Tesla V100 GPU for free (Microsoft Azure Cloud Free Trial) I just want to play with Leela Zero right now Download the best known network weights file from here, or, if you prefer a more human style, a (weaker) network trained from human games here. If you are on Windows, download an official release from here and head to the Usage section of this README. If you are on macOS, Leela
Top issues
#78How to find Information about the new best network1992 comments · 2017-11-17#591Version 0.10 released - Next steps732 comments · 2018-01-07#1311Facebook open sources elf opengo 416 comments · 2018-05-02#1030Bigger, stronger, not faster (4dc12a8e etc)378 comments · 2018-03-13#696Tune first play urgency ?369 comments · 2018-01-20#168140x256 training note.305 comments · 2018-08-02#965128x10 transition (6615567e)280 comments · 2018-03-02#1599Dynamical "komi" to keep winrate within a certain range215 comments · 2018-07-03
Latest release
Leela Zero 0.17 + AutoGTP v18 v0.17 2019-04-04Implemented batching for GPUs, increasing speed. Added support for Tensor Cores on NVIDIA cards. Optimized Winograd code for CPU and GPU. Improved UCT formula. Improved root move selection by using Lower Confidence Bounds. Improved passing heuristics (disabled in self-play). Improved autotuner behavior in edge cases. The training code now supports mixed precision fp32/fp16 training. AutoGTP can now start from an arbitrary SGF position. AutoGTP can now play matches with different white/black parameters. AutoGTP can now send initial commands to the engine. AutoGTP now provi