lyrids/PhoenixGo: Go AI program which implement the AlphaGo Zero paper · GitHub

PhoenixGo is an Go AI program which implement the AlphaGo Zero paper "Mastering the game of Go without human knowledge". It is also known as "BensonDarr" in FoxGo, "cronus" in CGOS, and the champion of "World AI Go Tournament 2018" held in Fuzhou China.

If you use PhoenixGo in your project, please consider mentioning in your README.

If you use PhoenixGo in your research, please consider citing the library as follows:

@misc{PhoenixGo2018,
  author = {Qinsong Zeng and Jianchang Zhang and Zhanpeng Zeng and Yongsheng Li and Ming Chen}
  title = {PhoenixGo},
  year = {2018},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/Tencent/PhoenixGo}}
}

• GCC with C++11 support
• Bazel (0.11.1 is known-good)
• (Optional) CUDA and cuDNN (for GPU support)
• (Optional) TensorRT (for accelerating computation on GPU, 3.0.4 is known-good)

Clone the repository and configure the building:

git clone https://github.com/Tencent/PhoenixGo.git
cd PhoenixGo
./configure

./configure will ask where CUDA and TensorRT have been installed, specify them if need.

Then build with bazel:

bazel build //mcts:mcts_main

Dependices such as Tensorflow will be downloaded automatically. The building prosess may take a long time.

Download and extract the trained network:

wget https://github.com/Tencent/PhoenixGo/releases/download/trained-network-20b-v1/trained-network-20b-v1.tar.gz
tar xvzf trained-network-20b-v1.tar.gz

Run in gtp mode with a config file (depend on the number of GPUs and using TensorRT or not):

bazel-bin/mcts/mcts_main --config_path=etc/{config} --gtp --logtostderr --v=1

The engine supports the GTP protocol, means it could be used with a GUI with GTP capability, such as Sabaki.

--logtostderr let mcts_main log messages to stderr, if you want to log to files, change --logtostderr to --log_dir={log_dir}

You could modify your config file following #configure-guide.

PhoenixGo support running with distributed workers, if there are GPUs on different machine.

Build the distribute worker:

bazel build //dist:dist_zero_model_server

Run dist_zero_model_server on distributed worker, one for each GPU.

CUDA_VISIBLE_DEVICES={gpu} bazel-bin/dist/dist_zero_model_server --server_address"0.0.0.0:{port}" --logtostderr

Fill ip:port of workers in the config file (etc/mcts_dist.conf is an example config for 32 workers), and run the distributed master:

bazel-bin/mcts/mcts_main --config_path=etc/{config} --gtp --logtostderr --v=1

Work in progress.

Here are some important options in the config file:

• num_eval_threads: should equal to the number of GPUs
• num_search_threads: should a bit larger than num_eval_threads * eval_batch_size
• timeout_ms_per_step: how many time will used for each move
• max_simulations_per_step: how many simulations will do for each move
• gpu_list: use which GPUs, sperated by comma
• model_config -> train_dir: directory where trained network stored
• model_config -> checkpoint_path: use which checkpoint, get from train_dir/checkpoint if not set
• model_config -> enable_tensorrt: use TensorRT or not
• model_config -> tensorrt_model_path: use which TensorRT model, if enable_tensorrt
• max_search_tree_size: the maximum number of tree nodes, change it depends on memory size
• max_children_per_node: the maximum children of each node, change it depends on memory size
• enable_background_search: pondering in opponent's time
• early_stop: genmove may return before timeout_ms_per_step, if the result would not change any more
• unstable_overtime: think timeout_ms_per_step * time_factor more if the result still unstable
• behind_overtime: think timeout_ms_per_step * time_factor more if winrate less than act_threshold

Options for distribute mode:

• enable_dist: enable distribute mode
• dist_svr_addrs: ip:port of distributed workers, multiple lines, one ip:port in each line
• dist_config -> timeout_ms: RPC timeout

Options for async distribute mode:

Async mode is used when there are huge number of distributed workers (more than 200), which need too many eval threads and search threads in sync mode. etc/mcts_async_dist.conf is an example config for 256 workers.

• enable_async: enable async mode
• enable_dist: enable distribute mode
• dist_svr_addrs: multiple lines, comma sperated lists of ip:port for each line
• num_eval_threads: should equal to number of dist_svr_addrs lines
• eval_task_queue_size: tunning depend on number of distribute workers
• num_search_threads: tunning depend on number of distribute workers

Read mcts/mcts_config.proto for more config options.

mcts_main accept options from command line:

• --config_path: path of config file
• --gtp: run as a GTP engine, if disable, gen next move only
• --init_moves: initial moves on the go board
• --gpu_list: override gpu_list in config file
• --listen_port: work with --gtp, run gtp engine on port in TCP protocol
• --allow_ip: work with --listen_port, list of client ip allowed to connect
• --fork_per_request: work with --listen_port, fork for each request or not

Glog options are also supported:

• --logtostderr: log message to stderr
• --log_dir: log to files in this directory
• --minloglevel: log level, 0 - INFO, 1 - WARNING, 2 - ERROR
• --v: verbose log, --v=1 for turning on some debug log, --v=0 to turning off

mcts_main --help for more command line options.