def load_model(config):
model = CChessModel(config)
if config.internet.distributed or config.opts.new or not load_best_model_weight(
model):
model.build()
save_as_best_model(model)
return model
class ObSelfPlay:
def __init__(self, config: Config):
self.config = config
self.env = CChessEnv()
self.model = None
self.pipe = None
self.ai = None
self.chessmans = None
def load_model(self):
self.model = CChessModel(self.config)
if self.config.opts.new or not load_best_model_weight(self.model):
self.model.build()
def start(self):
self.env.reset()
self.load_model()
self.pipe = self.model.get_pipes()
self.ai = CChessPlayer(self.config,
search_tree=defaultdict(VisitState),
pipes=self.pipe,
enable_resign=True,
debugging=False)
labels = ActionLabelsRed
labels_n = len(ActionLabelsRed)
self.env.board.print_to_cl()
history = [self.env.get_state()]
while not self.env.board.is_end():
no_act = None
state = self.env.get_state()
if state in history[:-1]:
no_act = []
for i in range(len(history) - 1):
if history[i] == state:
no_act.append(history[i + 1])
action, _ = self.ai.action(state, self.env.num_halfmoves, no_act)
history.append(action)
if action is None:
print("AI投降了!")
break
move = self.env.board.make_single_record(int(action[0]),
int(action[1]),
int(action[2]),
int(action[3]))
if not self.env.red_to_move:
action = flip_move(action)
self.env.step(action)
history.append(self.env.get_state())
print(f"AI选择移动 {move}")
self.env.board.print_to_cl()
sleep(1)
self.ai.close()
print(f"胜者是 is {self.env.board.winner} !!!")
self.env.board.print_record()
class ObSelfPlay:
def __init__(self, config: Config):
self.config = config
self.env = CChessEnv()
self.model = None
self.pipe = None
self.ai = None
self.chessmans = None
def load_model(self):
self.model = CChessModel(self.config)
if self.config.opts.new or not load_best_model_weight(self.model):
self.model.build()
def start(self):
self.env.reset()
self.load_model()
self.pipe = self.model.get_pipes()
self.ai = CChessPlayer(self.config, search_tree=defaultdict(VisitState), pipes=self.pipe,
enable_resign=True, debugging=False)
labels = ActionLabelsRed
labels_n = len(ActionLabelsRed)
self.env.board.print_to_cl()
history = [self.env.get_state()]
while not self.env.board.is_end():
no_act = None
state = self.env.get_state()
if state in history[:-1]:
no_act = []
for i in range(len(history) - 1):
if history[i] == state:
no_act.append(history[i + 1])
action, _ = self.ai.action(state, self.env.num_halfmoves, no_act)
history.append(action)
if action is None:
print("AI投降了!")
break
move = self.env.board.make_single_record(int(action[0]), int(action[1]), int(action[2]), int(action[3]))
if not self.env.red_to_move:
action = flip_move(action)
self.env.step(action)
history.append(self.env.get_state())
print(f"AI选择移动 {move}")
self.env.board.print_to_cl()
sleep(1)
self.ai.close()
print(f"胜者是 is {self.env.board.winner} !!!")
self.env.board.print_record()
def plot_model():
from keras.utils import plot_model
from cchess_alphazero.agent.model import CChessModel
from cchess_alphazero.config import Config
from cchess_alphazero.lib.model_helper import save_as_best_model
config = Config('distribute')
model = CChessModel(config)
model.build()
save_as_best_model(model)
plot_model(model.model,
to_file='model.png',
show_shapes=True,
show_layer_names=True)
def load_model(self, config_file=None):
use_history = False
model = CChessModel(self.config)
weight_path = self.config.resource.model_best_weight_path
if not config_file:
config_path = self.config.resource.model_best_config_path
use_history = False
else:
config_path = os.path.join(self.config.resource.model_dir, config_file)
try:
if not load_model_weight(model, config_path, weight_path):
model.build()
save_as_best_model(model)
use_history = True
except Exception as e:
logger.info(f"Exception {e}, 重新加载权重")
return self.load_model(config_file='model_192x10_config.json')
return model, use_history
class PlayWithHuman:
def __init__(self, config: Config):
self.config = config
self.env = CChessEnv()
self.model = None
self.pipe = None
self.ai = None
self.winstyle = 0
self.chessmans = None
self.human_move_first = True
self.screen_width = 720
self.height = 577
self.width = 521
self.chessman_w = 57
self.chessman_h = 57
self.disp_record_num = 15
self.rec_labels = [None] * self.disp_record_num
self.nn_value = 0
self.mcts_moves = {}
self.history = []
if self.config.opts.bg_style == 'WOOD':
self.chessman_w += 1
self.chessman_h += 1
def load_model(self):
self.model = CChessModel(self.config)
if self.config.opts.new or not load_best_model_weight(self.model):
self.model.build()
def init_screen(self):
bestdepth = pygame.display.mode_ok([self.screen_width, self.height],
self.winstyle, 32)
screen = pygame.display.set_mode([self.screen_width, self.height],
self.winstyle, bestdepth)
pygame.display.set_caption("中国象棋Zero")
# create the background, tile the bgd image
bgdtile = load_image(f'{self.config.opts.bg_style}.GIF')
bgdtile = pygame.transform.scale(bgdtile, (self.width, self.height))
board_background = pygame.Surface([self.width, self.height])
board_background.blit(bgdtile, (0, 0))
widget_background = pygame.Surface(
[self.screen_width - self.width, self.height])
white_rect = Rect(0, 0, self.screen_width - self.width, self.height)
widget_background.fill((255, 255, 255), white_rect)
#create text label
font_file = self.config.resource.font_path
font = pygame.font.Font(font_file, 16)
font_color = (0, 0, 0)
font_background = (255, 255, 255)
t = font.render("着法记录", True, font_color, font_background)
t_rect = t.get_rect()
t_rect.x = 10
t_rect.y = 10
widget_background.blit(t, t_rect)
screen.blit(board_background, (0, 0))
screen.blit(widget_background, (self.width, 0))
pygame.display.flip()
self.chessmans = pygame.sprite.Group()
creat_sprite_group(self.chessmans, self.env.board.chessmans_hash,
self.chessman_w, self.chessman_h)
return screen, board_background, widget_background
def start(self, human_first=True):
self.env.reset()
self.load_model()
self.pipe = self.model.get_pipes()
self.ai = CChessPlayer(self.config,
search_tree=defaultdict(VisitState),
pipes=self.pipe,
enable_resign=True,
debugging=True)
self.human_move_first = human_first
pygame.init()
screen, board_background, widget_background = self.init_screen()
framerate = pygame.time.Clock()
labels = ActionLabelsRed
labels_n = len(ActionLabelsRed)
current_chessman = None
if human_first:
self.env.board.calc_chessmans_moving_list()
ai_worker = Thread(target=self.ai_move, name="ai_worker")
ai_worker.daemon = True
ai_worker.start()
while not self.env.board.is_end():
for event in pygame.event.get():
if event.type == pygame.QUIT:
self.env.board.print_record()
self.ai.close(wait=False)
game_id = datetime.now().strftime("%Y%m%d-%H%M%S")
path = os.path.join(
self.config.resource.play_record_dir,
self.config.resource.play_record_filename_tmpl %
game_id)
self.env.board.save_record(path)
sys.exit()
elif event.type == VIDEORESIZE:
pass
elif event.type == MOUSEBUTTONDOWN:
if human_first == self.env.red_to_move:
pressed_array = pygame.mouse.get_pressed()
for index in range(len(pressed_array)):
if index == 0 and pressed_array[index]:
mouse_x, mouse_y = pygame.mouse.get_pos()
col_num, row_num = translate_hit_area(
mouse_x, mouse_y, self.chessman_w,
self.chessman_h)
chessman_sprite = select_sprite_from_group(
self.chessmans, col_num, row_num)
if current_chessman is None and chessman_sprite != None:
if chessman_sprite.chessman.is_red == self.env.red_to_move:
current_chessman = chessman_sprite
chessman_sprite.is_selected = True
elif current_chessman != None and chessman_sprite != None:
if chessman_sprite.chessman.is_red == self.env.red_to_move:
current_chessman.is_selected = False
current_chessman = chessman_sprite
chessman_sprite.is_selected = True
else:
move = str(current_chessman.chessman.col_num) + str(current_chessman.chessman.row_num) +\
str(col_num) + str(row_num)
success = current_chessman.move(
col_num, row_num, self.chessman_w,
self.chessman_h)
self.history.append(move)
if success:
self.chessmans.remove(
chessman_sprite)
chessman_sprite.kill()
current_chessman.is_selected = False
current_chessman = None
self.history.append(
self.env.get_state())
elif current_chessman != None and chessman_sprite is None:
move = str(current_chessman.chessman.col_num) + str(current_chessman.chessman.row_num) +\
str(col_num) + str(row_num)
success = current_chessman.move(
col_num, row_num, self.chessman_w,
self.chessman_h)
self.history.append(move)
if success:
current_chessman.is_selected = False
current_chessman = None
self.history.append(
self.env.get_state())
self.draw_widget(screen, widget_background)
framerate.tick(20)
# clear/erase the last drawn sprites
self.chessmans.clear(screen, board_background)
# update all the sprites
self.chessmans.update()
self.chessmans.draw(screen)
pygame.display.update()
self.ai.close(wait=False)
logger.info(f"Winner is {self.env.board.winner} !!!")
self.env.board.print_record()
game_id = datetime.now().strftime("%Y%m%d-%H%M%S")
path = os.path.join(
self.config.resource.play_record_dir,
self.config.resource.play_record_filename_tmpl % game_id)
self.env.board.save_record(path)
sleep(3)
def ai_move(self):
ai_move_first = not self.human_move_first
self.history = [self.env.get_state()]
no_act = None
while not self.env.done:
if ai_move_first == self.env.red_to_move:
labels = ActionLabelsRed
labels_n = len(ActionLabelsRed)
self.ai.search_results = {}
state = self.env.get_state()
logger.info(f"state = {state}")
_, _, _, check = senv.done(state, need_check=True)
if not check and state in self.history[:-1]:
no_act = []
free_move = defaultdict(int)
for i in range(len(self.history) - 1):
if self.history[i] == state:
# 如果走了下一步是将军或捉：禁止走那步
if senv.will_check_or_catch(
state, self.history[i + 1]):
no_act.append(self.history[i + 1])
# 否则当作闲着处理
else:
free_move[state] += 1
if free_move[state] >= 2:
# 作和棋处理
self.env.winner = Winner.draw
self.env.board.winner = Winner.draw
break
if no_act:
logger.debug(f"no_act = {no_act}")
action, policy = self.ai.action(state, self.env.num_halfmoves,
no_act)
if action is None:
logger.info("AI has resigned!")
return
self.history.append(action)
if not self.env.red_to_move:
action = flip_move(action)
key = self.env.get_state()
p, v = self.ai.debug[key]
logger.info(f"check = {check}, NN value = {v:.3f}")
self.nn_value = v
logger.info("MCTS results:")
self.mcts_moves = {}
for move, action_state in self.ai.search_results.items():
move_cn = self.env.board.make_single_record(
int(move[0]), int(move[1]), int(move[2]), int(move[3]))
logger.info(
f"move: {move_cn}-{move}, visit count: {action_state[0]}, Q_value: {action_state[1]:.3f}, Prior: {action_state[2]:.3f}"
)
self.mcts_moves[move_cn] = action_state
x0, y0, x1, y1 = int(action[0]), int(action[1]), int(
action[2]), int(action[3])
chessman_sprite = select_sprite_from_group(
self.chessmans, x0, y0)
sprite_dest = select_sprite_from_group(self.chessmans, x1, y1)
if sprite_dest:
self.chessmans.remove(sprite_dest)
sprite_dest.kill()
chessman_sprite.move(x1, y1, self.chessman_w, self.chessman_h)
self.history.append(self.env.get_state())
def draw_widget(self, screen, widget_background):
white_rect = Rect(0, 0, self.screen_width - self.width, self.height)
widget_background.fill((255, 255, 255), white_rect)
pygame.draw.line(widget_background, (255, 0, 0), (10, 285),
(self.screen_width - self.width - 10, 285))
screen.blit(widget_background, (self.width, 0))
self.draw_records(screen, widget_background)
self.draw_evaluation(screen, widget_background)
def draw_records(self, screen, widget_background):
text = '着法记录'
self.draw_label(screen, widget_background, text, 10, 16, 10)
records = self.env.board.record.split('\n')
font_file = self.config.resource.font_path
font = pygame.font.Font(font_file, 12)
i = 0
for record in records[-self.disp_record_num:]:
self.rec_labels[i] = font.render(record, True, (0, 0, 0),
(255, 255, 255))
t_rect = self.rec_labels[i].get_rect()
# t_rect.centerx = (self.screen_width - self.width) / 2
t_rect.y = 35 + i * 15
t_rect.x = 10
t_rect.width = self.screen_width - self.width
widget_background.blit(self.rec_labels[i], t_rect)
i += 1
screen.blit(widget_background, (self.width, 0))
def draw_evaluation(self, screen, widget_background):
title_label = 'CC-Zero信息'
self.draw_label(screen, widget_background, title_label, 300, 16, 10)
info_label = f'MCTS搜索次数：{self.config.play.simulation_num_per_move}'
self.draw_label(screen, widget_background, info_label, 335, 14, 10)
eval_label = f"当前局势评估: {self.nn_value:.3f}"
self.draw_label(screen, widget_background, eval_label, 360, 14, 10)
label = f"MCTS搜索结果:"
self.draw_label(screen, widget_background, label, 395, 14, 10)
label = f"着法 访问计数 动作价值 先验概率"
self.draw_label(screen, widget_background, label, 415, 12, 10)
i = 0
tmp = copy.deepcopy(self.mcts_moves)
for mov, action_state in tmp.items():
label = f"{mov}"
self.draw_label(screen, widget_background, label, 435 + i * 20, 12,
10)
label = f"{action_state[0]}"
self.draw_label(screen, widget_background, label, 435 + i * 20, 12,
70)
label = f"{action_state[1]:.2f}"
self.draw_label(screen, widget_background, label, 435 + i * 20, 12,
100)
label = f"{action_state[2]:.3f}"
self.draw_label(screen, widget_background, label, 435 + i * 20, 12,
150)
i += 1
def draw_label(self,
screen,
widget_background,
text,
y,
font_size,
x=None):
font_file = self.config.resource.font_path
font = pygame.font.Font(font_file, font_size)
label = font.render(text, True, (0, 0, 0), (255, 255, 255))
t_rect = label.get_rect()
t_rect.y = y
if x != None:
t_rect.x = x
else:
t_rect.centerx = (self.screen_width - self.width) / 2
widget_background.blit(label, t_rect)
screen.blit(widget_background, (self.width, 0))
def load_model(self):
model = CChessModel(self.config)
if self.config.opts.new or not load_sl_best_model_weight(model):
model.build()
save_as_sl_best_model(model)
return model
class PlayWithHuman:
def __init__(self, config: Config):
self.config = config
self.env = CChessEnv()
self.model = None
self.pipe = None
self.ai = None
self.chessmans = None
self.human_move_first = True
def load_model(self):
self.model = CChessModel(self.config)
if self.config.opts.new or not load_best_model_weight(self.model):
self.model.build()
def start(self, human_first=True):
self.env.reset()
self.load_model()
self.pipe = self.model.get_pipes()
self.ai = CChessPlayer(self.config,
search_tree=defaultdict(VisitState),
pipes=self.pipe,
enable_resign=True,
debugging=False)
self.human_move_first = human_first
labels = ActionLabelsRed
labels_n = len(ActionLabelsRed)
self.env.board.print_to_cl()
while not self.env.board.is_end():
if human_first == self.env.red_to_move:
self.env.board.calc_chessmans_moving_list()
is_correct_chessman = False
is_correct_position = False
chessman = None
while not is_correct_chessman:
title = "请输入棋子位置: "
input_chessman_pos = input(title)
x, y = int(input_chessman_pos[0]), int(
input_chessman_pos[1])
chessman = self.env.board.chessmans[x][y]
if chessman != None and chessman.is_red == self.env.board.is_red_turn:
is_correct_chessman = True
print(f"当前棋子为{chessman.name_cn}，可以落子的位置有：")
for point in chessman.moving_list:
print(point.x, point.y)
else:
print("没有找到此名字的棋子或未轮到此方走子")
while not is_correct_position:
title = "请输入落子的位置: "
input_chessman_pos = input(title)
x, y = int(input_chessman_pos[0]), int(
input_chessman_pos[1])
is_correct_position = chessman.move(x, y)
if is_correct_position:
self.env.board.print_to_cl()
self.env.board.clear_chessmans_moving_list()
else:
action, policy = self.ai.action(self.env.get_state(),
self.env.num_halfmoves)
if not self.env.red_to_move:
action = flip_move(action)
if action is None:
print("AI投降了!")
break
self.env.step(action)
print(f"AI选择移动 {action}")
self.env.board.print_to_cl()
self.ai.close()
print(f"胜者是 is {self.env.board.winner} !!!")
self.env.board.print_record()
class PlayWithHuman:
def __init__(self, config: Config):
self.config = config
self.env = CChessEnv()
self.model = None
self.pipe = None
self.ai = None
self.chessmans = None
self.human_move_first = True
def load_model(self):
self.model = CChessModel(self.config)
if self.config.opts.new or not load_best_model_weight(self.model):
self.model.build()
def start(self, human_first=True):
self.env.reset()
self.load_model()
self.pipe = self.model.get_pipes()
self.ai = CChessPlayer(self.config, search_tree=defaultdict(VisitState), pipes=self.pipe,
enable_resign=True, debugging=False)
self.human_move_first = human_first
labels = ActionLabelsRed
labels_n = len(ActionLabelsRed)
self.env.board.print_to_cl()
while not self.env.board.is_end():
if human_first == self.env.red_to_move:
self.env.board.calc_chessmans_moving_list()
is_correct_chessman = False
is_correct_position = False
chessman = None
while not is_correct_chessman:
title = "请输入棋子位置: "
input_chessman_pos = input(title)
x, y = int(input_chessman_pos[0]), int(input_chessman_pos[1])
chessman = self.env.board.chessmans[x][y]
if chessman != None and chessman.is_red == self.env.board.is_red_turn:
is_correct_chessman = True
print(f"当前棋子为{chessman.name_cn}，可以落子的位置有：")
for point in chessman.moving_list:
print(point.x, point.y)
else:
print("没有找到此名字的棋子或未轮到此方走子")
while not is_correct_position:
title = "请输入落子的位置: "
input_chessman_pos = input(title)
x, y = int(input_chessman_pos[0]), int(input_chessman_pos[1])
is_correct_position = chessman.move(x, y)
if is_correct_position:
self.env.board.print_to_cl()
self.env.board.clear_chessmans_moving_list()
else:
action, policy = self.ai.action(self.env.get_state(), self.env.num_halfmoves)
if not self.env.red_to_move:
action = flip_move(action)
if action is None:
print("AI投降了!")
break
self.env.step(action)
print(f"AI选择移动 {action}")
self.env.board.print_to_cl()
self.ai.close()
print(f"胜者是 is {self.env.board.winner} !!!")
self.env.board.print_record()
class PlayWithHuman:
def __init__(self, config: Config):
self.config = config
self.env = CChessEnv()
self.model = None
self.pipe = None
self.ai = None
self.chessmans = None
self.human_move_first = True
def load_model(self):
self.model = CChessModel(self.config)
if self.config.opts.new or not load_best_model_weight(self.model):
self.model.build()
def start(self, human_first=True):
self.env.reset()
self.load_model()
self.pipe = self.model.get_pipes()
self.ai = CChessPlayer(self.config,
search_tree=defaultdict(VisitState),
pipes=self.pipe,
enable_resign=True,
debugging=False)
self.human_move_first = human_first
labels = ActionLabelsRed
labels_n = len(ActionLabelsRed)
self.env.board.print_to_cl()
while not self.env.board.is_end():
if human_first == self.env.red_to_move:
self.env.board.calc_chessmans_moving_list()
is_correct_chessman = False
is_correct_position = False
chessman = None
while not is_correct_chessman:
title = "Please enter the chess piece position: "
input_chessman_pos = input(title)
print(input_chessman_pos)
x, y = int(input_chessman_pos[0]), int(
input_chessman_pos[1])
chessman = self.env.board.chessmans[x][y]
if chessman != None and chessman.is_red == self.env.board.is_red_turn:
is_correct_chessman = True
print(
f"The current chess piece is {chessman.name}，places where you can play："
)
for point in chessman.moving_list:
print(point.x, point.y)
else:
print(
"No chess piece with this name was found or it was not his turn to walk"
)
while not is_correct_position:
title = "Please enter the location of the child: "
input_chessman_pos = input(title)
x, y = int(input_chessman_pos[0]), int(
input_chessman_pos[1])
is_correct_position = chessman.move(x, y)
if is_correct_position:
self.env.board.print_to_cl()
self.env.board.clear_chessmans_moving_list()
else:
action, policy = self.ai.action(self.env.get_state(),
self.env.num_halfmoves)
if not self.env.red_to_move:
action = flip_move(action)
if action is None:
print("AI surrendered!")
break
self.env.step(action)
print(f"AI chooses to move {action}")
self.env.board.print_to_cl()
self.ai.close()
print(f"The winner is is {self.env.board.winner} !!!")
self.env.board.print_record()
class ObSelfPlayUCCI:
def __init__(self, config: Config, ai_move_first=True):
self.config = config
self.env = CChessEnv()
self.model = None
self.pipe = None
self.ai = None
self.chessmans = None
self.ai_move_first = ai_move_first
def load_model(self):
self.model = CChessModel(self.config)
if self.config.opts.new or not load_best_model_weight(self.model):
self.model.build()
def start(self):
self.env.reset()
self.load_model()
self.pipe = self.model.get_pipes()
self.ai = CChessPlayer(self.config,
search_tree=defaultdict(VisitState),
pipes=self.pipe,
enable_resign=True,
debugging=False)
labels = ActionLabelsRed
labels_n = len(ActionLabelsRed)
self.env.board.print_to_cl()
history = [self.env.get_state()]
turns = 0
game_over = False
final_move = None
while not game_over:
if (self.ai_move_first
and turns % 2 == 0) or (not self.ai_move_first
and turns % 2 == 1):
start_time = time()
no_act = None
state = self.env.get_state()
if state in history[:-1]:
no_act = []
for i in range(len(history) - 1):
if history[i] == state:
act = history[i + 1]
if not self.env.red_to_move:
act = flip_move(act)
no_act.append(act)
action, _ = self.ai.action(state, self.env.num_halfmoves,
no_act)
end_time = time()
if action is None:
print("AlphaZero 投降了!")
break
move = self.env.board.make_single_record(
int(action[0]), int(action[1]), int(action[2]),
int(action[3]))
print(
f"AlphaZero 选择移动 {move}, 消耗时间 {(end_time - start_time):.2f}s"
)
if not self.env.red_to_move:
action = flip_move(action)
else:
state = self.env.get_state()
print(state)
fen = senv.state_to_fen(state, turns)
action = self.get_ucci_move(fen)
if action is None:
print("Eleeye 投降了!")
break
print(action)
if not self.env.red_to_move:
rec_action = flip_move(action)
else:
rec_action = action
move = self.env.board.make_single_record(
int(rec_action[0]), int(rec_action[1]), int(rec_action[2]),
int(rec_action[3]))
print(f"Eleeye 选择移动 {move}")
history.append(action)
self.env.step(action)
history.append(self.env.get_state())
self.env.board.print_to_cl()
turns += 1
sleep(1)
game_over, final_move = self.env.board.is_end_final_move()
print(game_over, final_move)
if final_move:
move = self.env.board.make_single_record(int(final_move[0]),
int(final_move[1]),
int(final_move[2]),
int(final_move[3]))
print(f"Final Move {move}")
if not self.env.red_to_move:
final_move = flip_move(final_move)
self.env.step(final_move)
self.env.board.print_to_cl()
self.ai.close()
print(f"胜者是 is {self.env.board.winner} !!!")
self.env.board.print_record()
def get_ucci_move(self, fen, time=3):
p = subprocess.Popen(self.config.resource.eleeye_path,
stdin=subprocess.PIPE,
stdout=subprocess.PIPE,
stderr=subprocess.PIPE,
universal_newlines=True)
setfen = f'position fen {fen}\n'
setrandom = 'setoption randomness small\n'
cmd = 'ucci\n' + setrandom + setfen + f'go time {time * 1000}\n'
try:
out, err = p.communicate(cmd, timeout=time + 0.5)
except:
p.kill()
try:
out, err = p.communicate()
except Exception as e:
logger.error(f"{e}, cmd = {cmd}")
return self.get_ucci_move(fen, time + 1)
print(out)
lines = out.split('\n')
if lines[-2] == 'nobestmove':
return None
move = lines[-2].split(' ')[1]
if move == 'depth':
move = lines[-1].split(' ')[6]
return senv.parse_ucci_move(move)
def load_model(config):
model = CChessModel(config)
if config.internet.distributed or config.opts.new or not load_best_model_weight(model):
model.build()
save_as_best_model(model)
return model
class ObSelfPlayUCCI:
def __init__(self, config: Config, ai_move_first=True):
self.config = config
self.env = CChessEnv()
self.model = None
self.pipe = None
self.ai = None
self.chessmans = None
self.ai_move_first = ai_move_first
def load_model(self):
self.model = CChessModel(self.config)
if self.config.opts.new or not load_best_model_weight(self.model):
self.model.build()
def start(self):
self.env.reset()
self.load_model()
self.pipe = self.model.get_pipes()
self.ai = CChessPlayer(self.config, search_tree=defaultdict(VisitState), pipes=self.pipe,
enable_resign=True, debugging=False)
labels = ActionLabelsRed
labels_n = len(ActionLabelsRed)
self.env.board.print_to_cl()
history = [self.env.get_state()]
turns = 0
game_over = False
final_move = None
while not game_over:
if (self.ai_move_first and turns % 2 == 0) or (not self.ai_move_first and turns % 2 == 1):
start_time = time()
no_act = None
state = self.env.get_state()
if state in history[:-1]:
no_act = []
for i in range(len(history) - 1):
if history[i] == state:
act = history[i + 1]
if not self.env.red_to_move:
act = flip_move(act)
no_act.append(act)
action, _ = self.ai.action(state, self.env.num_halfmoves, no_act)
end_time = time()
if action is None:
print("AlphaZero 投降了!")
break
move = self.env.board.make_single_record(int(action[0]), int(action[1]), int(action[2]), int(action[3]))
print(f"AlphaZero 选择移动 {move}, 消耗时间 {(end_time - start_time):.2f}s")
if not self.env.red_to_move:
action = flip_move(action)
else:
state = self.env.get_state()
print(state)
fen = senv.state_to_fen(state, turns)
action = self.get_ucci_move(fen)
if action is None:
print("Eleeye 投降了!")
break
print(action)
if not self.env.red_to_move:
rec_action = flip_move(action)
else:
rec_action = action
move = self.env.board.make_single_record(int(rec_action[0]), int(rec_action[1]), int(rec_action[2]), int(rec_action[3]))
print(f"Eleeye 选择移动 {move}")
history.append(action)
self.env.step(action)
history.append(self.env.get_state())
self.env.board.print_to_cl()
turns += 1
sleep(1)
game_over, final_move = self.env.board.is_end_final_move()
print(game_over, final_move)
if final_move:
move = self.env.board.make_single_record(int(final_move[0]), int(final_move[1]), int(final_move[2]), int(final_move[3]))
print(f"Final Move {move}")
if not self.env.red_to_move:
final_move = flip_move(final_move)
self.env.step(final_move)
self.env.board.print_to_cl()
self.ai.close()
print(f"胜者是 is {self.env.board.winner} !!!")
self.env.board.print_record()
def get_ucci_move(self, fen, time=3):
p = subprocess.Popen(self.config.resource.eleeye_path,
stdin=subprocess.PIPE,
stdout=subprocess.PIPE,
stderr=subprocess.PIPE,
universal_newlines=True)
setfen = f'position fen {fen}\n'
setrandom = 'setoption randomness small\n'
cmd = 'ucci\n' + setrandom + setfen + f'go time {time * 1000}\n'
try:
out, err = p.communicate(cmd, timeout=time+0.5)
except:
p.kill()
try:
out, err = p.communicate()
except Exception as e:
logger.error(f"{e}, cmd = {cmd}")
return self.get_ucci_move(fen, time+1)
print(out)
lines = out.split('\n')
if lines[-2] == 'nobestmove':
return None
move = lines[-2].split(' ')[1]
if move == 'depth':
move = lines[-1].split(' ')[6]
return senv.parse_ucci_move(move)
class UCI:
def __init__(self, config: Config):
self.config = config
self.args = None
self.state = None
self.is_red_turn = None
self.player = None
self.model = None
self.pipe = None
self.is_ready = False
self.search_tree = defaultdict(VisitState)
self.remain_time = None
self.history = None
self.turns = 0
self.start_time = None
self.end_time = None
self.t = None
self.use_history = False
def main(self):
while True:
cmd = input()
logger.debug(f"CMD: {cmd}")
cmds = cmd.split(' ')
self.args = cmds[1:]
method = getattr(self, 'cmd_' + cmds[0], None)
if method != None:
method()
else:
logger.error(f"Error command: {cmd}")
def cmd_uci(self):
print('id name CCZero')
print('id author https://cczero.org')
print('id version 2.4')
print('option name gpu spin default 0 min 0 max 7')
print('option name Threads spin default 10 min 0 max 1024')
print('uciok')
sys.stdout.flush()
set_session_config(per_process_gpu_memory_fraction=1,
allow_growth=True,
device_list=self.config.opts.device_list)
self.use_history = self.load_model()
self.is_ready = True
self.turns = 0
self.remain_time = None
self.state = senv.INIT_STATE
self.history = [self.state]
self.is_red_turn = True
def cmd_ucinewgame(self):
self.state = senv.INIT_STATE
self.history = [self.state]
self.is_ready = True
self.is_red_turn = True
self.search_tree = defaultdict(VisitState)
def cmd_setoption(self):
'''
setoption name <id> [value <x>]
'''
if len(self.args) > 3:
id = self.args[1]
if id == 'gpu':
value = int(self.args[3])
self.config.opts.device_list = value
set_session_config(per_process_gpu_memory_fraction=1,
allow_growth=True,
device_list=self.config.opts.device_list)
if id == 'Threads':
value = int(self.args[3])
self.config.play.search_threads = value
def cmd_isready(self):
if self.is_ready == True:
print('readyok')
sys.stdout.flush()
logger.debug(f"is_ready = {self.is_ready}")
def cmd_position(self):
'''
position {fen <fenstring> | startpos } [moves <move1> .... <moven>]
'''
if not self.is_ready:
return
move_idx = -1
if len(self.args) > 0:
if self.args[0] == 'fen':
# init with fen string
fen = self.args[1]
try:
self.state = senv.fen_to_state(fen)
except Exception as e:
logger.error(f"cmd position error! cmd = {self.args}, {e}")
return
self.history = [self.state]
turn = self.args[2]
if turn == 'b':
self.state = senv.fliped_state(self.state)
self.is_red_turn = False
self.turns = (int(self.args[6]) - 1) * 2 + 1
else:
self.is_red_turn = True
self.turns = (int(self.args[6]) - 1) * 2
if len(self.args) > 7 and self.args[7] == 'moves':
move_idx = 8
elif self.args[0] == 'startpos':
self.state = senv.INIT_STATE
self.is_red_turn = True
self.history = [self.state]
self.turns = 0
if len(self.args) > 1 and self.args[1] == 'moves':
move_idx = 2
elif self.args[0] == 'moves':
move_idx = 1
else:
self.state = senv.INIT_STATE
self.is_red_turn = True
self.history = [self.state]
self.turns = 0
logger.debug(f"state = {self.state}")
# senv.render(self.state)
# execute moves
if move_idx != -1:
for i in range(move_idx, len(self.args)):
action = senv.parse_ucci_move(self.args[i])
if not self.is_red_turn:
action = flip_move(action)
self.history.append(action)
self.state = senv.step(self.state, action)
self.is_red_turn = not self.is_red_turn
self.turns += 1
self.history.append(self.state)
logger.debug(f"state = {self.state}")
# senv.render(self.state)
def cmd_fen(self):
self.args.insert(0, 'fen')
self.cmd_position()
def cmd_go(self):
'''
go ...
让引擎根据内置棋盘的设置和设定的搜索方式来思考，有以下搜索方式可供选择(可以多选，直接跟在go后面)：
❌(1) searchmoves <move1> .... <moven>，只让引擎在这几步中选择一步；
✅(2) wtime <x>，白方剩余时间(单位是毫秒)；
btime <x>，黑方剩余时间；
❌winc <x>，白方每步增加的时间(适用于Fischer制)；
❌binc <x>，黑方每步增加的时间；
❌movestogo <x>，还有多少回合进入下一时段(适用于时段制)；
这些选项用来设定时钟，它决定了引擎的思考时间；
❌(3) ponder，让引擎进行后台思考(即对手在用时，引擎的时钟不起作用)；
✅(4) depth <x>，指定搜索深度；
❌(5) nodes <x>，指定搜索的节点数(即分析的局面数，一般它和时间成正比)；
❌(6) mate <x>，在指定步数内只搜索杀棋；
✅(7) movetime <x>，只花规定的时间搜索；
✅(8) infinite，无限制搜索，直到杀棋。
'''
if not self.is_ready:
return
self.start_time = time()
self.t = None
depth = None
infinite = True
self.remain_time = None
self.model.close_pipes()
self.pipe = self.model.get_pipes(need_reload=False)
self.search_tree = defaultdict(VisitState)
self.player = CChessPlayer(self.config,
search_tree=self.search_tree,
pipes=self.pipe,
enable_resign=False,
debugging=True,
uci=True,
use_history=self.use_history,
side=self.turns % 2)
for i in range(len(self.args)):
if self.args[i] == 'depth':
depth = int(self.args[i + 1]) * 100
infinite = False
if self.args[i] == 'movetime' or self.args[i] == 'time':
self.remain_time = int(self.args[i + 1]) / 1000
if self.args[i] == 'infinite':
infinite = True
if self.args[i] == 'wtime':
if self.is_red_turn:
self.remain_time = int(self.args[i + 1]) / 1000
depth = 3000
infinite = False
if self.args[i] == 'btime':
if not self.is_red_turn:
self.remain_time = int(self.args[i + 1]) / 1000
depth = 3000
infinite = False
logger.debug(
f"depth = {depth}, infinite = {infinite}, remain_time = {self.remain_time}"
)
search_worker = Thread(target=self.search_action,
args=(depth, infinite))
search_worker.daemon = True
search_worker.start()
if self.remain_time:
self.t = Timer(self.remain_time - 0.01, self.cmd_stop)
self.t.start()
def cmd_stop(self):
if not self.is_ready:
return
if self.player:
no_act = None
if self.state in self.history[:-1]:
no_act = []
for i in range(len(self.history) - 1):
if self.history[i] == self.state:
no_act.append(self.history[i + 1])
action, value, depth = self.player.close_and_return_action(
self.state, self.turns, no_act)
self.player = None
self.model.close_pipes()
self.info_best_move(action, value, depth)
else:
logger.error(f"bestmove none")
def cmd_quit(self):
sys.exit()
def load_model(self, config_file=None):
use_history = True
self.model = CChessModel(self.config)
weight_path = self.config.resource.model_best_weight_path
if not config_file:
config_path = config.resource.model_best_path
use_history = False
else:
config_path = os.path.join(config.resource.model_dir, config_file)
try:
if not load_model_weight(self.model, config_path, weight_path):
self.model.build()
use_history = True
except Exception as e:
logger.info(f"Exception {e}, 重新加载权重")
return self.load_model(config_file='model_128_l1_config.json')
logger.info(f"use_history = {use_history}")
return use_history
def search_action(self, depth, infinite):
no_act = None
_, _, _, check = senv.done(self.state, need_check=True)
logger.debug(f"Check = {check}, state = {self.state}")
if not check and self.state in self.history[:-1]:
no_act = []
for i in range(len(self.history) - 1):
if self.history[i] == self.state:
if senv.will_check_or_catch(self.state,
self.history[i + 1]):
no_act.append(self.history[i + 1])
logger.debug(f"Foul: no act = {no_act}")
action, _ = self.player.action(self.state,
self.turns,
no_act=no_act,
depth=depth,
infinite=infinite,
hist=self.history)
if self.t:
self.t.cancel()
_, value = self.player.debug[self.state]
depth = self.player.done_tasks // 100
self.player.close(wait=False)
self.player = None
self.model.close_pipes()
self.info_best_move(action, value, depth)
def info_best_move(self, action, value, depth):
self.end_time = time()
if not self.is_red_turn:
value = -value
score = int(value * 1000)
duration = self.end_time - self.start_time
nps = int(depth * 100 / duration) * 1000
print(
f"info depth {depth} score {score} time {int(duration * 1000)} nps {nps}"
)
logger.debug(
f"info depth {depth} score {score} time {int((self.end_time - self.start_time) * 1000)}"
)
sys.stdout.flush()
# get ponder
state = senv.step(self.state, action)
ponder = None
if state in self.search_tree:
node = self.search_tree[state]
cnt = 0
for mov, action_state in node.a.items():
if action_state.n > cnt:
ponder = mov
cnt = action_state.n
if not self.is_red_turn:
action = flip_move(action)
action = senv.to_uci_move(action)
output = f"bestmove {action}"
if ponder:
if self.is_red_turn:
ponder = flip_move(ponder)
ponder = senv.to_uci_move(ponder)
output += f" ponder {ponder}"
print(output)
logger.debug(output)
sys.stdout.flush()