-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
executable file
·167 lines (158 loc) · 6.84 KB
/
train.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
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
#!/usr/bin/env python3
# pylint: disable=multiple-imports
import time, signal, argparse, logging, random
from episode import Episode
from state import State
from algo import AlgoSarsa, AlgoQLearning, AlgoPlay
from policy import Policy, PolicyExploit
DEFAULT_EPISODES = 10000
DEFAULT_STEP = 100
DEFAULT_ALPHA = .3
DEFAULT_EPSILON = .1
DEFAULT_LAYERS = 3
DEFAULT_TEMPERATURE = 0
running = True
logging.basicConfig(
level=logging.INFO,
format="[%(asctime)s] %(levelname)s %(name)s: %(message)s",
)
log = logging.getLogger('main')
def stop(_signum, _frame):
log.info('stopping...')
global running
running = False
def train(parser):
# sigterm
signal.signal(signal.SIGINT, stop)
signal.signal(signal.SIGTERM, stop)
# parse args
parser.add_argument('--game', '-g', metavar='GAME', type=str, default='piko',
help='name of game (default=piko)')
parser.add_argument('--episodes', '-e', metavar='N', type=int, default=DEFAULT_EPISODES,
help='total number of episodes (default=%d)'%(DEFAULT_EPISODES))
parser.add_argument('--step', '-s', metavar='S', type=int, default=DEFAULT_STEP,
help='number of episodes per step (default=%d)'%(DEFAULT_STEP))
parser.add_argument('--validation', '-v', metavar='V', type=int, default=0,
help='number of validation episodes run at each step (default=0)')
parser.add_argument('--offset', metavar='O', type=int, default=0,
help='offset in count of episodes (default=0)')
parser.add_argument('--hash', action='store_true',
help='use hash table instead of NN')
parser.add_argument('--base', '-b', metavar='DIR', type=str, default=None,
help='base directory for models backup')
parser.add_argument('--sarsa', metavar='DECAY', type=int, default=0,
help='use algo sarsa with decaying exploration ratio (default=off)')
parser.add_argument('--alpha', metavar='ALPHA', type=float, default=DEFAULT_ALPHA,
help='learning rate (default=%.3f)'%(DEFAULT_ALPHA))
parser.add_argument('--decay', metavar='K', type=int, default=0,
help='learning rate decay (default=off)')
parser.add_argument('--epsilon', metavar='EPSILON', type=float, default=DEFAULT_EPSILON,
help='exploration ratio (default=%.3f)'%(DEFAULT_EPSILON))
parser.add_argument('--softmax', metavar='T', type=float, default=0,
help='use a softmax exploration strategy with temperature T (default=off)')
parser.add_argument('--layers', '-l', metavar='L', type=int, default=DEFAULT_LAYERS,
help='number of hidden layers (default=%d)'%(DEFAULT_LAYERS))
parser.add_argument('--width', '-w', metavar='W', type=int, default=0,
help='width of hidden layers (default=same as input layer)')
parser.add_argument('--debug', '-d', action='store_true', default=False,
help='display debug log')
args = parser.parse_args()
if args.debug:
logging.getLogger().setLevel(logging.DEBUG)
log.debug(args)
# base directory
if not args.base:
args.base = ''
elif args.base[-1] != '/':
args.base += '/'
# params of training
EPISODES = args.episodes
STEP = args.step
# q-store
state = State.create(args.game)
if args.hash:
from q_hash import HashQ
q = HashQ(args.base+args.game+'.db', state.OUTPUTS)
else:
from q_network import NetworkQ
q = NetworkQ(args.base+args.game, state.INPUTS, state.OUTPUTS, layers=args.layers, width=args.width)
# algo
if args.sarsa > 0:
algo = AlgoSarsa()
else:
algo = AlgoQLearning(q)
# policy
if args.softmax > 0:
policy = Policy.create('softmax', q)
else:
policy = Policy.create('egreedy', q)
# learning mode
algo.set_params(args.alpha)
policy.set_params(args.epsilon, args.softmax)
alpha = args.alpha
epsilon = args.epsilon
# counters
won = episodes = tot_turns = tot_backups = 0
algo.reset_stats()
policy.reset_stats()
time0 = time.time()
while running:
# initial state
state = State.create(args.game, random.choice([True, False]))
state,turns,backups = Episode(algo, policy).run(state)
episodes += 1
if not args.validation:
# no validation: update stats during training
if state.player_wins():
won += 1
tot_turns += turns
tot_backups += backups
if not episodes % STEP:
mean_duration = (time.time() - time0) * float(1000) / STEP
if args.validation == 0:
stats_step = STEP
else:
stats_step = args.validation
# run some validation episodes
won = tot_turns = tot_backups = 0
# disable training
algo.set_params(0)
algo_play = AlgoPlay()
for _ in range(stats_step):
state = State.create(args.game, random.choice([True, False]))
state,turns,backups = Episode(algo_play, PolicyExploit(q)).run(state)
if state.player_wins():
won += 1
tot_turns += turns
tot_backups += backups
# restore training settings
algo.set_params(alpha)
# report stats
rate = 100 * float(won)/stats_step
mean_td_error = algo.get_stats()
log.info('games: %d / won: %.1f%% of %d / avg turns: %.1f / avg backups: %.1f\n'
'time: %.2fms/episode / mean abs td error: %.3f',
episodes, rate, stats_step,
float(tot_turns)/stats_step, float(tot_backups)/stats_step,
mean_duration, mean_td_error
)
won = tot_turns = tot_backups = 0
algo.reset_stats()
policy.reset_stats()
q.save(epoch=(episodes+args.offset))
if args.decay:
# adjust learning rate with decay
alpha = args.alpha * args.decay / (args.decay+episodes+args.offset)
log.info('learning rate: %.3f', alpha)
if args.sarsa > 0:
# sarsa: adjust exploration rate
epsilon = args.epsilon * args.sarsa / (args.sarsa+episodes+args.offset)
log.info('exploration rate: %.3f', epsilon)
policy.set_params(epsilon, args.softmax)
algo.set_params(alpha)
time0 = time.time()
if episodes == EPISODES:
break
q.save()
if __name__ == "__main__":
train(argparse.ArgumentParser(description='Train the strategy.'))