def supervised_buffer(config, game) -> (ChessEnv, list):
env = ChessEnv(config).reset()
white = ChessPlayer(config, dummy=True)
black = ChessPlayer(config, dummy=True)
result = game.headers["Result"]
env.board = game.board()
for move in game.main_line():
ai = white if env.board.turn == chess.WHITE else black
ai.sl_action(env, move)
env.step(move)
if not env.board.is_game_over() and result != '1/2-1/2':
env.resigned = True
if result == '1-0':
env.winner = Winner.WHITE
white_win = 1
elif result == '0-1':
env.winner = Winner.BLACK
white_win = -1
else:
env.winner = Winner.DRAW
white_win = 0
white.finish_game(white_win)
black.finish_game(-white_win)
return env, merge_data(white, black)
def self_play_buffer(config, cur) -> (ChessEnv, list):
pipes = cur.pop() # borrow
env = ChessEnv().reset()
white = ChessPlayer(config, pipes=pipes)
black = ChessPlayer(config, pipes=pipes)
while not env.done:
if env.white_to_move:
action = white.action(env)
else:
action = black.action(env)
env.step(action)
if env.num_halfmoves >= config.play.max_game_length:
env.adjudicate()
if env.winner == Winner.white:
black_win = -1
elif env.winner == Winner.black:
black_win = 1
else:
black_win = 0
black.finish_game(black_win)
white.finish_game(-black_win)
data = []
for i in range(len(white.moves)):
data.append(white.moves[i])
if i < len(black.moves):
data.append(black.moves[i])
cur.append(pipes)
return env, data
def get_buffer(config, game) -> (ChessEnv, list):
"""
Gets data to load into the buffer by playing a game using PGN data.
:param Config config: config to use to play the game
:param pgn.Game game: game to play
:return list(str,list(float)): data from this game for the SupervisedLearningWorker.buffer
"""
env = ChessEnv().reset()
white = ChessPlayer(config, dummy=True)
black = ChessPlayer(config, dummy=True)
result = game.headers["Result"]
white_elo, black_elo = int(game.headers["WhiteElo"]), int(
game.headers["BlackElo"])
white_weight = clip_elo_policy(config, white_elo)
black_weight = clip_elo_policy(config, black_elo)
actions = []
while not game.is_end():
game = game.variation(0)
actions.append(game.move.uci())
k = 0
while not env.done and k < len(actions):
if env.white_to_move:
action = white.sl_action(env.observation,
actions[k],
weight=white_weight) #ignore=True
else:
action = black.sl_action(env.observation,
actions[k],
weight=black_weight) #ignore=True
env.step(action, False)
k += 1
if not env.board.is_game_over() and result != '1/2-1/2':
env.resigned = True
if result == '1-0':
env.winner = Winner.white
black_win = -1
elif result == '0-1':
env.winner = Winner.black
black_win = 1
else:
env.winner = Winner.draw
black_win = 0
black.finish_game(black_win)
white.finish_game(-black_win)
data = []
for i in range(len(white.moves)):
data.append(white.moves[i])
if i < len(black.moves):
data.append(black.moves[i])
return env, data
def get_buffer(config, game) -> (ChessEnv, list):
"""
Gets data to load into the buffer by playing a game using PGN data.
:param Config config: config to use to play the game
:param pgn.Game game: game to play
:return list(str,list(float)): data from this game for the SupervisedLearningWorker.buffer
"""
env = ChessEnv().reset()
white = ChessPlayer(config, dummy=True)
black = ChessPlayer(config, dummy=True)
result = game.headers["Result"]
white_elo, black_elo = int(game.headers["WhiteElo"]), int(game.headers["BlackElo"])
white_weight = clip_elo_policy(config, white_elo)
black_weight = clip_elo_policy(config, black_elo)
actions = []
while not game.is_end():
game = game.variation(0)
actions.append(game.move.uci())
k = 0
while not env.done and k < len(actions):
if env.white_to_move:
action = white.sl_action(env.observation, actions[k], weight=white_weight) #ignore=True
else:
action = black.sl_action(env.observation, actions[k], weight=black_weight) #ignore=True
env.step(action, False)
k += 1
if not env.board.is_game_over() and result != '1/2-1/2':
env.resigned = True
if result == '1-0':
env.winner = Winner.white
black_win = -1
elif result == '0-1':
env.winner = Winner.black
black_win = 1
else:
env.winner = Winner.draw
black_win = 0
black.finish_game(black_win)
white.finish_game(-black_win)
data = []
for i in range(len(white.moves)):
data.append(white.moves[i])
if i < len(black.moves):
data.append(black.moves[i])
return env, data
def self_play_buffer(config, cur) -> (GoBangEnv, list):
"""
Play one game and add the play data to the buffer
:param Config config: config for how to play
:param list(Connection) cur: list of pipes to use to get a pipe to send observations to for getting
predictions. One will be removed from this list during the game, then added back
:return (GoBangEnv,list((str,list(float)): a tuple containing the final GoBangEnv state and then a list
of data to be appended to the SelfPlayWorker.buffer
"""
pipes = cur.pop() # borrow
env = GoBangEnv().reset()
white = ChessPlayer(config, pipes=pipes)
black = ChessPlayer(config, pipes=pipes)
while not env.done:
if env.white_to_move:
action = white.action(env)
else:
action = black.action(env)
# pretty_print_panel(env.board.panel)
# print(f'After action:{action}')
env.step(action)
# pretty_print_panel(env.board.panel)
# print()
# if env.num_halfmoves >= config.play.max_game_length:
# env.adjudicate()
if env.winner == Winner.white:
black_score, white_score = -1, 1
elif env.winner == Winner.black:
black_score, white_score = 1, -1
else:
black_score, white_score = -0.5, -0.5
black.finish_game(black_score)
white.finish_game(white_score)
data = []
for i in range(len(black.moves)):
data.append(black.moves[i])
if i < len(white.moves):
data.append(white.moves[i])
pretty_print_panel(env.board.panel)
print()
#print(data)
cur.append(pipes)
return env, data
def get_buffer(config, game) -> (ChessEnv, list):
env = ChessEnv().reset()
white = ChessPlayer(config, dummy=True)
black = ChessPlayer(config, dummy=True)
result = game.headers["Result"]
white_elo, black_elo = int(game.headers["WhiteElo"]), int(game.headers["BlackElo"])
white_weight = clip_elo_policy(config, white_elo)
black_weight = clip_elo_policy(config, black_elo)
actions = []
while not game.is_end():
game = game.variation(0)
actions.append(game.move.uci())
k = 0
while not env.done and k < len(actions):
if env.white_to_move:
action = white.sl_action(env.observation, actions[k], weight=white_weight) #ignore=True
else:
action = black.sl_action(env.observation, actions[k], weight=black_weight) #ignore=True
env.step(action, False)
k += 1
if not env.board.is_game_over() and result != '1/2-1/2':
env.resigned = True
if result == '1-0':
env.winner = Winner.white
black_win = -1
elif result == '0-1':
env.winner = Winner.black
black_win = 1
else:
env.winner = Winner.draw
black_win = 0
black.finish_game(black_win)
white.finish_game(-black_win)
data = []
for i in range(len(white.moves)):
data.append(white.moves[i])
if i < len(black.moves):
data.append(black.moves[i])
return env, data
def get_buffer(game, config) -> (ChessEnv, list):
env = ChessEnv().reset()
black = ChessPlayer(config, dummy=True)
white = ChessPlayer(config, dummy=True)
result = game.headers["Result"]
actions = []
while not game.is_end():
game = game.variation(0)
actions.append(game.move.uci())
k = 0
observation = env.observation
while not env.done and k < len(actions):
if env.board.turn == chess.WHITE:
action = white.sl_action(observation, actions[k]) #ignore=True
else:
action = black.sl_action(observation, actions[k]) #ignore=True
board, info = env.step(action, False)
observation = board.fen()
k += 1
env.done = True
if not env.board.is_game_over() and result != '1/2-1/2':
env.resigned = True
if result == '1-0':
env.winner = Winner.white
black_win = -1
elif result == '0-1':
env.winner = Winner.black
black_win = 1
else:
env.winner = Winner.draw
black_win = 0
black.finish_game(black_win)
white.finish_game(-black_win)
data = []
for i in range(len(white.moves)):
data.append(white.moves[i])
if i < len(black.moves):
data.append(black.moves[i])
return env, data
def self_play_buffer(config, cur) -> (ChessEnv, list):
pipes = cur.pop() # borrow
env = ChessEnv().reset()
search_tree = defaultdict(VisitStats)
white = ChessPlayer(config, search_tree=search_tree, pipes=pipes)
black = ChessPlayer(config, search_tree=search_tree, pipes=pipes)
history = []
cc = 0
while not env.done:
if env.white_to_move:
action = white.action(env)
else:
action = black.action(env)
env.step(action)
history.append(action)
if len(history) > 6 and history[-1] == history[-5]:
cc = cc + 1
else:
cc = 0
if env.num_halfmoves >= config.play.max_game_length or cc >= 4:
env.adjudicate()
if env.winner == Winner.white:
black_win = -1
elif env.winner == Winner.black:
black_win = 1
else:
black_win = 0
black.finish_game(black_win)
white.finish_game(-black_win)
data = []
for i in range(len(white.moves)):
data.append(white.moves[i])
if i < len(black.moves):
data.append(black.moves[i])
cur.append(pipes)
return env, data
class SelfPlayWorker:
def __init__(self, config: Config, env=None, model=None):
"""
:param config:
:param ChessEnv|None env:
:param chess_zero.agent.model_chess.ChessModel|None model:
"""
self.config = config
self.model = model
self.env = env # type: ChessEnv
self.black = None # type: ChessPlayer
self.white = None # type: ChessPlayer
self.buffer = []
def start(self):
if self.model is None:
self.model = self.load_model()
self.buffer = []
idx = 1
while True:
start_time = time()
env = self.start_game(idx)
end_time = time()
logger.debug(f"game {idx} time={end_time - start_time} sec, "
f"turn={env.turn}:{env.observation} - Winner:{env.winner} - by resignation?:{env.resigned}")
if (idx % self.config.play_data.nb_game_in_file) == 0:
reload_best_model_weight_if_changed(self.model)
idx += 1
def start_game(self, idx):
self.env.reset()
self.black = ChessPlayer(self.config, self.model)
self.white = ChessPlayer(self.config, self.model)
observation = self.env.observation
while not self.env.done:
if self.env.board.turn == chess.BLACK:
action = self.black.action(observation)
else:
action = self.white.action(observation)
board, info = self.env.step(action)
observation = board.fen()
self.finish_game()
self.save_play_data(write=idx % self.config.play_data.nb_game_in_file == 0)
self.remove_play_data()
return self.env
def save_play_data(self, write=True):
data = self.black.moves + self.white.moves
self.buffer += data
if not write:
return
rc = self.config.resource
game_id = datetime.now().strftime("%Y%m%d-%H%M%S.%f")
path = os.path.join(rc.play_data_dir, rc.play_data_filename_tmpl % game_id)
logger.info(f"save play data to {path}")
write_game_data_to_file(path, self.buffer)
self.buffer = []
def remove_play_data(self):
files = get_game_data_filenames(self.config.resource)
if len(files) < self.config.play_data.max_file_num:
return
for i in range(len(files) - self.config.play_data.max_file_num):
os.remove(files[i])
def finish_game(self):
if self.env.winner == Winner.black:
black_win = 1
elif self.env.winner == Winner.white:
black_win = -1
else:
black_win = 0
self.black.finish_game(black_win)
self.white.finish_game(-black_win)
def load_model(self):
from chess_zero.agent.model_chess import ChessModel
model = ChessModel(self.config)
if self.config.opts.new or not load_best_model_weight(model):
model.build()
save_as_best_model(model)
return model
class SupervisedLearningWorker:
def __init__(self, config: Config, env=None, model=None):
"""
:param config:
:param ChessEnv|None env:
:param chess_zero.agent.model_chess.ChessModel|None model:
"""
self.config = config
self.model = model
self.env = env # type: ChessEnv
self.black = None # type: ChessPlayer
self.white = None # type: ChessPlayer
self.buffer = []
def start(self):
if self.model is None:
self.model = self.load_model()
self.buffer = []
idx = 1
while True:
start_time = time()
env = self.read_game(idx)
end_time = time()
logger.debug(f"game {idx} time={end_time - start_time} sec, "
f"turn={int(env.turn/2)}:{env.observation} - Winner:{env.winner} - by resignation?:{env.resigned}")
if (idx % self.config.play_data.nb_game_in_file) == 0:
reload_best_model_weight_if_changed(self.model)
idx += 1
def read_game(self, idx):
self.env.reset()
self.black = ChessPlayer(self.config, self.model)
self.white = ChessPlayer(self.config, self.model)
files = find_pgn_files(self.config.resource.play_data_dir)
if len(files) > 0:
random.shuffle(files)
filename = files[0]
pgn = open(filename, errors='ignore')
size = os.path.getsize(filename)
pos = random.randint(0, size)
pgn.seek(pos)
line = pgn.readline()
offset = 0
# Parse game headers.
while line:
if line.isspace() or line.startswith("%"):
line = pgn.readline()
continue
# Read header tags.
tag_match = TAG_REGEX.match(line)
if tag_match:
offset = pgn.tell()
break
line = pgn.readline()
pgn.seek(offset)
game = chess.pgn.read_game(pgn)
node = game
result = game.headers["Result"]
actions = []
while not node.is_end():
next_node = node.variation(0)
actions.append(node.board().uci(next_node.move))
node = next_node
pgn.close()
k = 0
observation = self.env.observation
while not self.env.done and k < len(actions):
if self.env.board.turn == chess.BLACK:
action = self.black.sl_action(observation, actions[k])
else:
action = self.white.sl_action(observation, actions[k])
board, info = self.env.step(action)
observation = board.fen()
k += 1
self.env.done = True
if not self.env.board.is_game_over() and result != '1/2-1/2':
self.env.resigned = True
if result == '1-0':
self.env.winner = Winner.white
elif result == '0-1':
self.env.winner = Winner.black
else:
self.env.winner = Winner.draw
self.finish_game()
self.save_play_data(write=idx % self.config.play_data.nb_game_in_file == 0)
self.remove_play_data()
return self.env
def save_play_data(self, write=True):
data = self.black.moves + self.white.moves
self.buffer += data
if not write:
return
rc = self.config.resource
game_id = datetime.now().strftime("%Y%m%d-%H%M%S.%f")
path = os.path.join(rc.play_data_dir, rc.play_data_filename_tmpl % game_id)
logger.info(f"save play data to {path}")
write_game_data_to_file(path, self.buffer)
self.buffer = []
def remove_play_data(self):
files = get_game_data_filenames(self.config.resource)
if len(files) < self.config.play_data.max_file_num:
return
for i in range(len(files) - self.config.play_data.max_file_num):
os.remove(files[i])
def finish_game(self):
if self.env.winner == Winner.black:
black_win = 1
elif self.env.winner == Winner.white:
black_win = -1
else:
black_win = 0
self.black.finish_game(black_win)
self.white.finish_game(-black_win)
def load_model(self):
from chess_zero.agent.model_chess import ChessModel
model = ChessModel(self.config)
if self.config.opts.new or not load_best_model_weight(model):
model.build()
save_as_best_model(model)
return model
def self_play_buffer(config, cur) -> (ChessEnv, list):
"""
Play one game and add the play data to the buffer
:param Config config: config for how to play
:param list(Connection) cur: list of pipes to use to get a pipe to send observations to for getting
predictions. One will be removed from this list during the game, then added back
:return (ChessEnv,list((str,list(float)): a tuple containing the final ChessEnv state and then a list
of data to be appended to the SelfPlayWorker.buffer
"""
pipes = cur.pop() # borrow
env = ChessEnv().reset()
# EDIT CODE HERE TO CHANGE THE ENVIRONMENT
white = ChessPlayer(config, pipes=pipes)
black = ChessPlayer(config, pipes=pipes)
move = 0
failed_play = 0
total_failed_plays = 0
print("Match Started")
moves_list = ""
while not env.done:
# CHANGES_MADE_HERE
temp = deepcopy(env)
black_pieces = set("prnbqk")
white_pieces = set("PRNBQK")
if env.white_to_move:
x = temp.board.piece_map()
for i in x:
if str(x[i]) in black_pieces:
temp.board.remove_piece_at(i)
action = white.action(temp)
else:
x = temp.board.piece_map()
for i in x:
if str(x[i]) in white_pieces:
temp.board.remove_piece_at(i)
action = black.action(temp)
print("Match in Progress: ",
move,
"Moves made in the game, Failed Plays: ",
total_failed_plays,
end='\r')
try:
env.step(action)
moves_list += action + ', '
failed_play = 0
move += 1
if env.num_halfmoves >= config.play.max_game_length:
env.adjudicate()
except ValueError:
failed_play += 1
total_failed_plays += 1
if failed_play == 50:
logger.warning("\nEnding the Game due to lack of development")
env.adjudicate()
continue
# END_OF_CHANGES
with open("result.csv", "a+") as fp:
result = str(move) + ", " + str(total_failed_plays) + ", " + str(
env.winner) + ", <" + env.board.fen()
result += ">, Adjudicated\n" if failed_play == 50 else ">, Game End\n"
fp.write(result)
fp.close()
with open("moves_list.csv", "a+") as fp:
fp.write(moves_list)
fp.write("\n")
fp.close()
if env.winner == Winner.white:
black_win = -1
logger.info("White wins")
elif env.winner == Winner.black:
black_win = 1
logger.info("Black wins")
else:
black_win = 0
logger.info("Draw Match")
black.finish_game(black_win)
white.finish_game(-black_win)
data = []
for i in range(len(white.moves)):
data.append(white.moves[i])
if i < len(black.moves):
data.append(black.moves[i])
cur.append(pipes)
return env, data