-
Notifications
You must be signed in to change notification settings - Fork 6
/
Renju.py
127 lines (108 loc) · 7.86 KB
/
Renju.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
#!/usr/bin/python
# -*- coding: utf-8 -*-
# author: <[email protected]>
import argparse
from AI_train import *
from util.AI_tools import Utility
from AI_mcts import Node, MCTSThreadPool
def import_corpus_to_db(file_path):
if os.path.exists(file_path):
corpus.import_RenjuNet(file_path)
def parser_argument():
str2bool = lambda v: v.lower() in ("yes", "true", "t", "1")
parse = argparse.ArgumentParser()
parse.add_argument("--handle", choices=['train_policy_dl', 'train_policy_rl', 'train_policy_rollout',
'train_value_network', 'play', 'import_corpus'], help="handle type")
# common parameters
parse.add_argument("--board_size", type=int, default=15, help="board size")
parse.add_argument("--policy_planes", type=int, default=3, help="number of policy planes")
parse.add_argument("--value_planes", type=int, default=4, help="number of value planes")
parse.add_argument("--pattern_features", type=int, default=64, help="number of pattern features")
# deploy parameters
parse.add_argument("--ps_hosts", type=str, default="localhost:2222",
help="Comma-separated list of hostname:port pairs")
parse.add_argument("--worker_hosts", type=str,
default="localhost:2223,localhost:2224,localhost:2225,localhost:2226",
help="Comma-separated list of hostname:port pairs")
parse.add_argument("--job_name", type=str, default="worker", help="[ps, worker]")
parse.add_argument("--task_index", type=int, default=0, help="distributed task index")
parse.add_argument("--gpu_id", type=int, default=0, help="gpu card id")
# policy network parameters of supervised learning(DCNN)
parse.add_argument("--policy_dl_models_dir", type=str, default="./policy_dl_multi_gpu_models", help="policy dl models")
parse.add_argument("--policy_dl_filters", type=int, default=192, help="policy dl filters")
parse.add_argument("--policy_dl_optimizer", choices=['rmsprop', 'adam', 'sgd'], default='adam', help='network optimization')
parse.add_argument("--policy_dl_learn_rate", type=float, default=0.25, help="policy dl learn rate")
parse.add_argument("--policy_dl_decay_rate", type=float, default=0.5, help="policy dl decay rate")
parse.add_argument("--policy_dl_epochs_per_decay", type=float, default=20, help="policy dl num epochs per decay")
parse.add_argument("--policy_dl_gpu", type=int, default=1, help="policy dl train gpu")
parse.add_argument("--policy_dl_epochs", type=int, default=100, help="policy dl train epochs")
parse.add_argument("--policy_dl_batch_size", type=int, default=128, help="policy dl train batch size")
# policy network parameters of supervised learning(rollout)
parse.add_argument("--policy_rollout_models_dir", type=str, default="./policy_rollout_multi_gpu_models", help="policy rollout models")
parse.add_argument("--policy_rollout_optimizer", choices=['rmsprop', 'adam', 'sgd'], default='sgd', help='network optimization')
parse.add_argument("--policy_rollout_learn_rate", type=float, default=0.25, help="policy dl learn rate")
parse.add_argument("--policy_rollout_decay_rate", type=float, default=0.5, help="policy dl decay rate")
parse.add_argument("--policy_rollout_epochs_per_decay", type=float, default=50, help="policy dl num epochs per decay")
parse.add_argument("--policy_rollout_gpu", type=int, default=2, help="policy rollout train gpu")
parse.add_argument("--policy_rollout_epochs", type=int, default=100, help="policy rollout train epochs")
parse.add_argument("--policy_rollout_batch_size", type=int, default=128, help="policy rollout train batch size")
# policy network parameters of reinforcement learning
parse.add_argument("--policy_rl_models_dir", type=str, default="./policy_rl_multi_gpu_models", help="policy rl models")
parse.add_argument("--policy_rl_filters", type=int, default=192, help="policy rl filters")
parse.add_argument("--policy_rl_optimizer", choices=['rmsprop', 'adam', 'sgd'], default='adam', help='network optimization')
parse.add_argument("--policy_rl_learn_rate", type=float, default=0.00025, help="policy rl learn rate")
parse.add_argument("--policy_rl_gpu", type=int, default=1, help="policy rl gpu")
parse.add_argument("--policy_rl_reset", type=str2bool, default=False, help="policy rl model reset")
parse.add_argument("--policy_rl_phase", type=int, default=1, help="policy rl model phase")
parse.add_argument("--policy_rl_batch_games", type=int, default=10, help="policy rl train batch games")
parse.add_argument("--policy_rl_save_step", type=int, default=200, help="policy rl save step")
# value network parameters
parse.add_argument("--values_net_models_dir", type=str, default="./value_sl_models", help="values network models")
parse.add_argument("--values_net_filters", type=int, default=192, help="values network filters")
parse.add_argument("--values_net_optimizer", choices=['rmsprop', 'adam', 'sgd'], default='sgd', help='network optimization')
parse.add_argument("--values_net_learn_rate", type=float, default=0.025, help="values network learn rate")
parse.add_argument("--values_net_gpu", type=int, default=1, help="values network gpu")
parse.add_argument("--values_net_phase", type=int, default=1, help="values network phase")
parse.add_argument("--values_net_sample_num", type=int, default=10000, help="values network sampling number")
parse.add_argument("--values_net_max_time_steps", type=int, default=225, help="values network max time steps")
parse.add_argument("--values_net_batch_size", type=int, default=128, help="values network batch size")
parse.add_argument("--values_net_epochs", type=int, default=1, help="values network epochs")
# MCTS parameters
parse.add_argument("--mcts_visit_threshold", type=float, default=5, help="mcst visit threshold")
parse.add_argument("--mcts_virtual_loss", type=float, default=0.2, help="mcst virtual loss")
parse.add_argument("--mcts_explore_rate", type=float, default=5.0, help="mcst explore rate")
parse.add_argument("--mcts_mix_lambda", type=float, default=0.75, help="mcst mix lambda")
# play games
parse.add_argument("--player", choices=["black", "white"], default="black", help="player role")
parse.add_argument("--board", type=str, default="", help="board status, format is '010110110100011|0110...'")
# server parameters
parse.add_argument("--model_type", choices=["policy_dl", "policy_rl", "policy_rollout", "value_net"],
help="ip port for server")
parse.add_argument("--main_ip_port", type=str, default="localhost:2220", help="ip port for main server")
parse.add_argument("--policy_dl_ip_port", type=str, default="localhost:2221", help="ip port for policy dl server")
parse.add_argument("--policy_rl_ip_port", type=str, default="localhost:2222", help="ip port for policy rl server")
parse.add_argument("--policy_rollout_ip_port", type=str, default="localhost:2223",
help="ip port for policy rollout server")
parse.add_argument("--value_net_ip_port", type=str, default="localhost:2224", help="ip port for value net server")
# other parameters
parse.add_argument("--corpus", type=str, default="", help="corpus to import")
return parse
def run(parse):
args = parse.parse_args()
if args.handle == "import_corpus":
import_corpus_to_db(args.corpus)
elif args.handle == "train_policy_dl":
train_policy_network_dl(args)
elif args.handle == "train_policy_rl":
train_policy_network_rl(args)
elif args.handle == "train_policy_rollout":
train_policy_rollout(args)
elif args.handle == "train_value_network":
train_value_network_sl(args)
elif args.handle == "play":
play_games(args)
else:
parse.print_help()
if __name__ == "__main__":
arg_parser = parser_argument()
run(arg_parser)