def main():
connection = connect()
config.set_device(args.gpu)
# while True:
name = None
while name is None:
name = connection.recv()
model = Model('model_' + name)
model.forced_restore()
# print(model)
agent = Agent(model)
simulations = None
while simulations is None:
simulations = connection.recv()
simulations = int(simulations)
config.simulations = simulations
env = GameEnv()
init = deal_init(connection)
env.load(init)
game = Game(env, args.player)
game_start(agent, simulations, game, connection)
def __call__(self):
if self.empty():
return None
file, init, player, p, v = self.data[self.iter]
env = GameEnv(False)
env.load(init)
game = Game(env, player)
self.iter += 1
return game, p, v
def unpack(init, ps, player):
env = GameEnv(False)
env.load(init)
game = Game(env, player)
histories = []
for action in ps:
state = game.to_model_input()
histories.append((histories[-1] if len(histories) else []) + [state])
game.move(action)
for i in range(len(histories)):
histories[i] = np.array(histories[i], dtype=np.int8)
return histories
def main():
model: Model
config.set_device(1)
env = GameEnv()
# model = Model('model')
model = Model('model_201912080009')
# model = Model('model_tencent0824')
model.forced_restore()
# agent = Agent(model)
agent = Agent(model)
init = [[2, 1, 2, 1, 0, 1, 3, 1, 2, 1, 0, 1, 2, 0, 0],
[2, 1, 1, 2, 1, 1, 0, 3, 1, 2, 0, 2, 0, 0, 1],
[0, 1, 1, 1, 3, 2, 0, 0, 1, 1, 3, 1, 2, 1, 0]]
actions = [
352, 352, 353, 338, 343, 347, 123, 0, 0, 20, 22, 23, 24, 26, 0, 28, 0,
29, 0, 0, 39, 0, 0, 116, 0, 0, 76, 324, 0, 0, 41, 42, 0, 0, 92, 317,
320, 0, 0, 31
] #, 42, 0, 0, 15, 18]
init = np.array(init, dtype=np.int32)
env.load(init)
print(env)
root, player = new_Root(), 2
game = Game(env, player)
for action in actions:
mct = MCT(game, root)
agent.simulates([mct], store_rnn_states='Always')
root.root(action)
game.move(action)
print(game)
print('GAUSS', game.gauss())
print(game.curr_player(), game.my_player())
print('====')
print(game.curr_player())
print(game.my_player())
print(game.lord_player())
print(game.hand_cards_num())
print(game.bottom())
print('====')
mct = MCT(game, root)
for cnt in range(2000):
agent.simulates([mct], store_rnn_states='Always')
# if cnt == 0:
# history, rnn_state = root.get_model_input()
# print(history)
# if (cnt + 1) % 10 == 0:
if (cnt + 1) % 10 == 0:
print(cnt + 1)
for action, P, son in mct.root.edges():
print('%d: %.8f %d %.3f' % (action, P, son.N(), son.Q()))
print('-------------------------')
t = 1.0
s = np.array([son.N() for action, P, son in mct.root.edges()])
p = np.array([P for action, P, son in mct.root.edges()])
print(s)
print(np.mean(s))
w = s + 0.001 - np.mean(s)
w[w < 0] = 0
w = (w**t) / (w**t).sum()
print(w)
print(s / s.sum())
print(p)
mct.json()
def main():
np.set_printoptions(precision=2, linewidth=128, suppress=True)
path = 'test'
config.device = torch.device('cuda:1')
config.set_device(-1)
model = Model('model_201912080009')
model.restore()
files = os.listdir('gen/%s/' % path)
files.sort(key=lambda k: k.split('_')[0], reverse=True)
# print(files[-1])
file = np.random.choice(files[:100])
# file = '201912101326_85d94af6fe1a588b.pkl'
print(file)
data = pickle_load('gen/%s/%s' % (path, file))
# data = pickle_load('gen/test/' + files[-2])
# np.random.shuffle(data)
player = 2
init, actions, _v = [None, [], -1.0]
init = [[2, 1, 2, 1, 0, 1, 3, 1, 2, 1, 0, 1, 2, 0, 0],
[2, 1, 1, 2, 1, 1, 0, 3, 1, 2, 0, 2, 0, 0, 1],
[0, 1, 1, 1, 3, 2, 0, 0, 1, 1, 3, 1, 2, 1, 0]]
actions = [
352, 352, 353, 338, 343, 347, 123, 0, 0, 20, 22, 23, 24, 26, 0, 28, 0,
29, 0, 0, 39, 0, 0, 116, 0, 0, 76, 324, 0, 0, 41, 42, 0, 0, 92, 317,
320, 0, 0, 31, 42, 0, 0, 15, 18
]
init = np.array(init, dtype=np.int32)
# actions = [353, 352, 352, 339, 349, 349, 15]
# init = [[2, 2, 2, 2, 2, 2, 2, 1, 1, 1, 0, 0, 0, 0, 0],
# [2, 2, 2, 2, 2, 2, 2, 1, 1, 1, 0, 0, 0, 0, 0],
# [0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 4, 4, 4, 1, 1]]
# init = np.array(init, dtype=np.int32)
# actions = [353, 352, 352, 344, 345, 346, 151]
# init, actions, _v = data[player]
print(_v, player)
print(init)
print(actions)
print('=============================================')
print('player:', player)
histories = unpack(init, actions, player)
histories, lengths = history_regularize(histories)
histories, lengths = to_cuda(histories, lengths)
vs, ps, _ = model(histories, lengths, None)
env = GameEnv(False)
env.load(init)
game = Game(env, player)
for v, p, action in zip(vs, ps, actions):
print('----------------------------------')
print('my_player: %d, curr_player: %d' % (player, game.curr_player()))
# for action, _dist in enumerate(dist):
# print(action, _dist)
idx = np.argsort(p)[::-1]
for i in range(8):
print(game.could_move(idx[i]), end=' ')
print('(%d, %.2f%%)' % (idx[i], p[idx[i]] * 100), end='\n')
print('action: %d, %.2f%%' % (action, p[action] * 100))
if idx[0] == 60 and p[idx[0]] > 0.3:
print(game)
print(game.policy())
print(game.hand_cards_num())
print(game.bottom())
print(game.curr_player(), game.lord_player())
return 0
# model_input = game.to_model_input()
# for channel in range(26, 28):
# print(channel)
# print(model_input[channel])
print('%.1f, %.3f' % (_v, v[0]))
game.move(action)
print(game)
print('Gauss:', game.gauss())
import time
from game import Game, GameEnv
import numpy as np
path = 'data/'
files = os.listdir(path)
file = np.random.choice(files)
print(file)
data = pickle_load(path + file)
init, policy, v = data[0]
env = GameEnv()
print(init)
env.load(init)
print(policy)
for action in policy:
print(env.curr_player(), env.lord_player())
print(action)
if env.period() != 3:
env.move(action)
else:
env.move(-1)
print(env)
print('---------------------')
time.sleep(0.5)
print(v)
# TODO: MCTS simulate should be same as action choice