def arena_process(i):
g = Game(8)
nnet = nn(g)
nnet.load_model(filename=("model_auto_" + str(i + 1)))
nmcts = MCTS(g, nnet, args)
pnet = nn(g)
if i != 0:
pnet.load_model(filename=("model_auto_" + str(i)))
pmcts = MCTS(g, pnet, args)
def player1(x):
pi = pmcts.get_action_prob(x)
# display_pi(np.array(pi[:-1]).reshape((len(x), len(x))))
return np.random.choice(len(pi), p=pi)
def player2(x):
pi = nmcts.get_action_prob(x)
return np.random.choice(len(pi), p=pi)
arena = Arena(player1=lambda x: player1(x),
player2=lambda x: player2(x),
game=g,
display=display)
return arena.play_games(8)
def mcts_test():
g = Game(15)
b = Board(15)
nnet = nn(g)
mcts = MCTS(g, nnet, args)
b.execute_move((4, 1), 1)
b.execute_move((3, 2), 1)
b.execute_move((2, 3), 1)
b.execute_move((5, 0), 1)
# b.execute_move((1, 4), 1)
b.execute_move((3, 3), -1)
b.execute_move((3, 4), -1)
b.execute_move((3, 5), -1)
b.execute_move((3, 6), -1)
# b.execute_move((3, 5), -1)
curPlayer = 1
canonicalBoard = g.get_canonical_form(np.array(b.pieces), curPlayer)
pi = mcts.get_action_prob(canonicalBoard)
display(canonicalBoard)
display_pi(np.array(pi[:-1]).reshape((len(canonicalBoard), len(canonicalBoard))))
def generate_data(l):
g = Game(args.goBang_n)
nnet = nn(g)
c = Coach(g, nnet, args)
train_example = c.execute_episode()
l.acquire()
try:
folder = args.checkpoint
if not os.path.exists(folder):
os.makedirs(folder)
filename = os.path.join(folder + "train_examples_4")
with open(filename, "ab+") as f:
pickle.dump(train_example, f)
finally:
l.release()
def generate_data(l, model_iter):
g = Game(8)
nnet = nn(g)
nnet.load_model(filename=("model_auto_" + str(model_iter + 1)))
c = Coach(g, nnet, args)
train_example = c.execute_episode()
l.acquire()
try:
folder = args.checkpoint
if not os.path.exists(folder):
os.makedirs(folder)
filename = os.path.join(folder +
("train_examples_auto_" + str(model_iter + 1)))
with open(filename, "ab+") as f:
pickle.dump(train_example, f)
finally:
l.release()
'numIters': 10,
'numEps': 10,
'tempThreshold': 100,
'updateThreshold': 0.6,
'maxlenOfQueue': 2000000,
'numMCTSSims': 1000,
'arenaCompare': 64,
'cpuct': 3,
'checkpoint': './temp/',
'load_model': False,
'load_folder_file': ('/dev/models/8x100x50', 'best.pth.tar'),
'numItersForTrainExamplesHistory': 20,
}
if __name__ == "x":
g = Game(15)
nnet = nn(g)
if args.load_model:
nnet.load_checkpoint(args.load_folder_file[0],
args.load_folder_file[1])
c = Coach(g, nnet, args)
if args.load_model:
print("Load trainExamples from file")
c.load_train_examples()
c.learn()
if __name__ == '__main__':
g = Game(15)
auto_run = AutoRun(g, args)
lock = Lock()
for iteration in range(200):
jobs = []
for _ in range(8):
p = Process(target=generate_data, args=(lock, i))
jobs.append(p)
p.start()
for job in jobs:
job.join()
trainExamples = merge_data(i)
print(len(trainExamples))
g = Game(8)
nnet = nn(g)
nnet.train(trainExamples)
nnet.save_model(filename="model_auto_" + str(i + 2))
else:
break
print(i, 'one model')
if __name__ == 'x':
with Pool(8) as p:
result = p.map(arena_process, range(8))
win_1 = sum([i[0] for i in result])
win_2 = sum([i[1] for i in result])