-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
144 lines (105 loc) · 5.7 KB
/
main.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
# -*- coding: utf-8 -*-
# %matplotlib inline
import numpy as np
np.set_printoptions(suppress=True)
from shutil import copyfile
import random
from importlib import reload
from keras.utils import plot_model
from game import Game, GameState
from agent import Agent
from memory import Memory
from model import Residual_CNN
from funcs import playMatches, playMatchesBetweenVersions
import loggers as lg
from settings import run_folder, run_archive_folder
import initialise
import pickle
lg.logger_main.info('=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*')
lg.logger_main.info('=*=*=*=*=*=. NEW LOG =*=*=*=*=*')
lg.logger_main.info('=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*=*')
env = Game()
# If loading an existing neural network, copy the config file to root
if initialise.INITIAL_RUN_NUMBER != None:
copyfile(run_archive_folder + env.name + '/run' + str(initialise.INITIAL_RUN_NUMBER).zfill(4) + '/config.py', './config.py')
import config
######## LOAD MEMORIES IF NECESSARY ########
if initialise.INITIAL_MEMORY_VERSION == None:
memory = Memory(config.MEMORY_SIZE)
else:
print('LOADING MEMORY VERSION ' + str(initialise.INITIAL_MEMORY_VERSION) + '...')
memory = pickle.load( open( run_archive_folder + env.name + '/run' + str(initialise.INITIAL_RUN_NUMBER).zfill(4) + "/memory/memory" + str(initialise.INITIAL_MEMORY_VERSION).zfill(4) + ".p", "rb" ) )
######## LOAD MODEL IF NECESSARY ########
# create an untrained neural network objects from the config file
current_NN = Residual_CNN(config.REG_CONST, config.LEARNING_RATE, (2,) + env.grid_shape, env.action_size, config.HIDDEN_CNN_LAYERS)
best_NN = Residual_CNN(config.REG_CONST, config.LEARNING_RATE, (2,) + env.grid_shape, env.action_size, config.HIDDEN_CNN_LAYERS)
#If loading an existing neural netwrok, set the weights from that model
if initialise.INITIAL_MODEL_VERSION != None:
best_player_version = initialise.INITIAL_MODEL_VERSION
print('LOADING MODEL VERSION ' + str(initialise.INITIAL_MODEL_VERSION) + '...')
m_tmp = best_NN.read(env.name, initialise.INITIAL_RUN_NUMBER, best_player_version)
current_NN.model.set_weights(m_tmp.get_weights())
best_NN.model.set_weights(m_tmp.get_weights())
#otherwise just ensure the weights on the two players are the same
else:
best_player_version = 0
best_NN.model.set_weights(current_NN.model.get_weights())
#copy the config file to the run folder
copyfile('./config.py', run_folder + 'config.py')
plot_model(current_NN.model, to_file=run_folder + 'models/model.png', show_shapes = True)
print('\n')
######## CREATE THE PLAYERS ########
current_player = Agent('current_player', env.state_size, env.action_size, config.MCTS_SIMS, config.CPUCT, current_NN)
best_player = Agent('best_player', env.state_size, env.action_size, config.MCTS_SIMS, config.CPUCT, best_NN)
#user_player = User('player1', env.state_size, env.action_size)
iteration = 0
while 1:
iteration += 1
reload(lg)
reload(config)
print('ITERATION NUMBER ' + str(iteration))
lg.logger_main.info('BEST PLAYER VERSION: %d', best_player_version)
print('BEST PLAYER VERSION ' + str(best_player_version))
######## SELF PLAY ########
print('SELF PLAYING ' + str(config.EPISODES) + ' EPISODES...')
_, memory, _, _ = playMatches(best_player, best_player, config.EPISODES, lg.logger_main, turns_until_tau0 = config.TURNS_UNTIL_TAU0, memory = memory)
print('\n')
memory.clear_stmemory()
if len(memory.ltmemory) >= config.MEMORY_SIZE:
######## RETRAINING ########
print('RETRAINING...')
current_player.replay(memory.ltmemory)
print('')
if iteration % 5 == 0:
pickle.dump( memory, open( run_folder + "memory/memory" + str(iteration).zfill(4) + ".p", "wb" ) )
lg.logger_memory.info('====================')
lg.logger_memory.info('NEW MEMORIES')
lg.logger_memory.info('====================')
memory_samp = random.sample(memory.ltmemory, min(1000, len(memory.ltmemory)))
for s in memory_samp:
current_value, current_probs, _ = current_player.get_preds(s['state'])
best_value, best_probs, _ = best_player.get_preds(s['state'])
lg.logger_memory.info('MCTS VALUE FOR %s: %f', s['playerTurn'], s['value'])
lg.logger_memory.info('CUR PRED VALUE FOR %s: %f', s['playerTurn'], current_value)
lg.logger_memory.info('BES PRED VALUE FOR %s: %f', s['playerTurn'], best_value)
lg.logger_memory.info('THE MCTS ACTION VALUES: %s', ['%.2f' % elem for elem in s['AV']] )
lg.logger_memory.info('CUR PRED ACTION VALUES: %s', ['%.2f' % elem for elem in current_probs])
lg.logger_memory.info('BES PRED ACTION VALUES: %s', ['%.2f' % elem for elem in best_probs])
lg.logger_memory.info('ID: %s', s['state'].id)
lg.logger_memory.info('INPUT TO MODEL: %s', current_player.model.convertToModelInput(s['state']))
s['state'].render(lg.logger_memory)
######## TOURNAMENT ########
print('TOURNAMENT...')
scores, _, points, sp_scores = playMatches(best_player, current_player, config.EVAL_EPISODES, lg.logger_tourney, turns_until_tau0 = 0, memory = None)
print('\nSCORES')
print(scores)
print('\nSTARTING PLAYER / NON-STARTING PLAYER SCORES')
print(sp_scores)
#print(points)
print('\n\n')
if scores['current_player'] > scores['best_player'] * config.SCORING_THRESHOLD:
best_player_version = best_player_version + 1
best_NN.model.set_weights(current_NN.model.get_weights())
best_NN.write(env.name, best_player_version)
else:
print('MEMORY SIZE: ' + str(len(memory.ltmemory)))