-
Notifications
You must be signed in to change notification settings - Fork 169
/
optimize.py
182 lines (154 loc) · 6.53 KB
/
optimize.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
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
import os
from datetime import datetime
from logging import getLogger
from time import sleep
import keras.backend as K
import numpy as np
from keras.optimizers import SGD
from reversi_zero.agent.model import ReversiModel, objective_function_for_policy, \
objective_function_for_value
from reversi_zero.config import Config
from reversi_zero.lib import tf_util
from reversi_zero.lib.bitboard import bit_to_array
from reversi_zero.lib.data_helper import get_game_data_filenames, read_game_data_from_file, \
get_next_generation_model_dirs
from reversi_zero.lib.model_helpler import load_best_model_weight
logger = getLogger(__name__)
def start(config: Config):
tf_util.set_session_config(per_process_gpu_memory_fraction=0.59)
return OptimizeWorker(config).start()
class OptimizeWorker:
def __init__(self, config: Config):
self.config = config
self.model = None # type: ReversiModel
self.loaded_filenames = set()
self.loaded_data = {}
self.dataset = None
self.optimizer = None
def start(self):
self.model = self.load_model()
self.training()
def training(self):
self.compile_model()
last_save_step = total_steps = self.config.trainer.start_total_steps
min_data_size_to_learn = 100000
self.load_play_data()
while True:
if self.dataset_size < min_data_size_to_learn:
logger.info(f"dataset_size={self.dataset_size} is less than {min_data_size_to_learn}")
sleep(60)
continue
self.update_learning_rate(total_steps)
steps = self.train_epoch(self.config.trainer.epoch_to_checkpoint)
total_steps += steps
if last_save_step + self.config.trainer.save_model_steps < total_steps:
self.save_current_model()
last_save_step = total_steps
self.load_play_data()
def train_epoch(self, epochs):
tc = self.config.trainer
state_ary, policy_ary, z_ary = self.dataset
self.model.model.fit(state_ary, [policy_ary, z_ary],
batch_size=tc.batch_size,
epochs=epochs)
steps = (state_ary.shape[0] // tc.batch_size) * epochs
return steps
def compile_model(self):
self.optimizer = SGD(lr=1e-2, momentum=0.9)
losses = [objective_function_for_policy, objective_function_for_value]
self.model.model.compile(optimizer=self.optimizer, loss=losses)
def update_learning_rate(self, total_steps):
# The deepmind paper says
# ~400k: 1e-2
# 400k~600k: 1e-3
# 600k~: 1e-4
if total_steps < 100000:
lr = 1e-2
elif total_steps < 200000:
lr = 1e-3
else:
lr = 1e-4
K.set_value(self.optimizer.lr, lr)
logger.debug(f"total step={total_steps}, set learning rate to {lr}")
def save_current_model(self):
rc = self.config.resource
model_id = datetime.now().strftime("%Y%m%d-%H%M%S.%f")
model_dir = os.path.join(rc.next_generation_model_dir, rc.next_generation_model_dirname_tmpl % model_id)
os.makedirs(model_dir, exist_ok=True)
config_path = os.path.join(model_dir, rc.next_generation_model_config_filename)
weight_path = os.path.join(model_dir, rc.next_generation_model_weight_filename)
self.model.save(config_path, weight_path)
def collect_all_loaded_data(self):
state_ary_list, policy_ary_list, z_ary_list = [], [], []
for s_ary, p_ary, z_ary_ in self.loaded_data.values():
state_ary_list.append(s_ary)
policy_ary_list.append(p_ary)
z_ary_list.append(z_ary_)
state_ary = np.concatenate(state_ary_list)
policy_ary = np.concatenate(policy_ary_list)
z_ary = np.concatenate(z_ary_list)
return state_ary, policy_ary, z_ary
@property
def dataset_size(self):
if self.dataset is None:
return 0
return len(self.dataset[0])
def load_model(self):
from reversi_zero.agent.model import ReversiModel
model = ReversiModel(self.config)
rc = self.config.resource
dirs = get_next_generation_model_dirs(rc)
if not dirs:
logger.debug(f"loading best model")
if not load_best_model_weight(model):
raise RuntimeError(f"Best model can not loaded!")
else:
latest_dir = dirs[-1]
logger.debug(f"loading latest model")
config_path = os.path.join(latest_dir, rc.next_generation_model_config_filename)
weight_path = os.path.join(latest_dir, rc.next_generation_model_weight_filename)
model.load(config_path, weight_path)
return model
def load_play_data(self):
filenames = get_game_data_filenames(self.config.resource)
updated = False
for filename in filenames:
if filename in self.loaded_filenames:
continue
self.load_data_from_file(filename)
updated = True
for filename in (self.loaded_filenames - set(filenames)):
self.unload_data_of_file(filename)
updated = True
if updated:
logger.debug("updating training dataset")
self.dataset = self.collect_all_loaded_data()
def load_data_from_file(self, filename):
try:
logger.debug(f"loading data from {filename}")
data = read_game_data_from_file(filename)
self.loaded_data[filename] = self.convert_to_training_data(data)
self.loaded_filenames.add(filename)
except Exception as e:
logger.warning(str(e))
def unload_data_of_file(self, filename):
logger.debug(f"removing data about {filename} from training set")
self.loaded_filenames.remove(filename)
if filename in self.loaded_data:
del self.loaded_data[filename]
@staticmethod
def convert_to_training_data(data):
"""
:param data: format is SelfPlayWorker.buffer
list of [(own: bitboard, enemy: bitboard), [policy: float 64 items], z: number]
:return:
"""
state_list = []
policy_list = []
z_list = []
for state, policy, z in data:
own, enemy = bit_to_array(state[0], 64).reshape((8, 8)), bit_to_array(state[1], 64).reshape((8, 8))
state_list.append([own, enemy])
policy_list.append(policy)
z_list.append(z)
return np.array(state_list), np.array(policy_list), np.array(z_list)