def replay(self, pos, data=None, times=80):
final_score = strategies.fast_score(pos)
mcts = strategies.RandomPlayer()
rboard = pos.copy()
name = mp.current_process().name
for i in range(times):
if not config.running:
break
rboard.undo()
curboard = rboard.copy()
c=curboard.to_move
while config.running and not curboard.is_gameover:
move, move_values, points = mcts.suggest_move(curboard)
curboard.win_rate = points if c == go.BLACK else -points
data.add_from_node(curboard, move_values, points)
if data.data_size%1024==5:
print("name:", name, ", data size:", data.data_size, file=sys.stderr)
if data.data_size>25600:
print("name:", name, "保存训练数据……", end="")
data.save()
data.clear()
print("name:", name, "保存完毕！")
curboard.play_move(move)
score = curboard.win_rate if c==go.BLACK else -curboard.win_rate
cscore = score-final_score
cscore = cscore if c == go.BLACK else -cscore
if i%20==15:
print("name:", name, "倒退{}步，全探查结果：{}，{}{}了{:.1f}子".format(i+1,
go.result_str(score), go.get_color_str(c),
"增加" if cscore>0 else "减少", abs(cscore)), file=sys.stderr)
def play(self, pos=None, forbid_pass=True, sgfdir='sgf'):
if not pos:
pos = board.Board(komi=7.5)
if self.qp:
self.qp.start(pos)
passbw = 0
mcts = strategies.MCTSSercher(self.net, qp=self.qp)
caps = None
while config.running:
if pos.step>go.N*go.N*2:
pos.result=0
break
c = pos.to_move
move, values, win_rate = mcts.suggest_move(pos, forbid_pass=forbid_pass)
self.datas.add_from_node(pos, values, win_rate)
if self.datas.data_size>12800:
print("保存训练数据……", end="", file=sys.stderr)
self.datas.save()
self.datas.clear()
print("完毕！", file=sys.stderr)
pos.win_rate = win_rate
if (move is None or move == go.PASS):
passbw |= (1 if c == go.BLACK else 2)
if passbw == 3 or (not forbid_pass):
pos.play_move(go.PASS)
score = strategies.fast_score(pos)
pos.result = score / 2
'''msg = '%s方第%d手PASS，对局结束, %s。' % (go.get_color_str(c),
pos.step, go.result_str(score / 2))
if self.qp:
self.qp.show_message(msg=msg)'''
break
else:
pos.play_move(go.PASS)
continue
elif move == go.RESIGN:
pos.play_move(move)
'''msg = '%s方第%d手投子认负，对局结束, %s。' % (go.get_color_str(c),
pos.step, go.result_str(pos.result))
if self.qp:
self.qp.show_message(msg=msg)'''
break
illegal, caps = pos.play_move(move)
if illegal == 0:
passbw = 0
if self.qp:
self.qp.update(pos)
if sgfdir:
dt = sgfdir + '/self_' + time.strftime('%Y-%m-%d_%H_%M_%S') + '.sgf'
'''msg = '%.1f：\t保存sgf棋谱文件到%s' % (time.time() - policy.start_time, dt)
if self.qp:
self.qp.show_message(msg=msg)'''
if not os.path.exists(sgfdir):
os.makedirs(sgfdir)
sgfparser.save_board(pos, dt)
return pos
def play(board=None, datas=None, queue=None, net=None, qp=None, sgfdir='sgf'):
if not board:
board = board.Board(komi=7.5)
if qp:
qp.start(board)
passbw = 0
mcts = strategies.MTCSSercher(net=net, qp=qp)
caps = None
while config.running:
if board.step > go.N * go.N * 2:
board.result = 0
break
c = board.to_move
move, values, points = mcts.suggest_move(board)
if board.step > go.N:
datas.add_from_node(board, values, points)
if datas.data_size % 1024 == 5:
queue.put("{}-- data size: {}".format(name, data.data_size))
if datas.data_size > 12800:
print("保存训练数据……", end="", file=sys.stderr)
datas.save()
datas.clear()
print("完毕！", file=sys.stderr)
board.win_rate = points if c == go.BLACK else -points
if (move is None or move == go.PASS):
passbw |= (1 if c == go.BLACK else 2)
if passbw == 3:
board.play_move(go.PASS)
score = strategies.fast_score(board)
board.result = score / 2
break
else:
board.play_move(go.PASS)
continue
elif move == go.RESIGN:
board.play_move(move)
board.result = go.N * go.N + 1 if c == go.WHITE else -go.N * go.N - 1
'''msg = '%s方第%d手投子认负，对局结束, %s。' % (go.get_color_str(c),
board.step, go.result_str(board.result))
if qp:
qp.show_message(msg=msg)'''
break
illegal, caps = board.play_move(move)
if illegal == 0:
passbw = 0
if qp:
qp.update(board)
if sgfdir:
dt = sgfdir + '/self_' + time.strftime('%Y-%m-%d_%H_%M_%S') + '.sgf'
'''msg = '%.1f：\t保存sgf棋谱文件到%s' % (time.time() - policy.start_time, dt)
if qp:
qp.show_message(msg=msg)'''
if not os.path.exists(sgfdir):
os.makedirs(sgfdir)
sgfparser.save_board(board, dt)
return board
def replay(board, data, queue, times=80, qp=None):
final_score = strategies.fast_score(board)
mcts = strategies.RandomPlayer()
rboard = board.copy()
name = mp.current_process().name
config.running = True
for i in range(times):
if not config.running:
break
if rboard.step < 2:
break
rboard.undo()
curboard = rboard.copy()
if qp:
qp.start(curboard)
c = curboard.to_move
while config.running and not curboard.is_gameover:
move, move_values, points = mcts.suggest_move(curboard)
curboard.win_rate = points if c == go.BLACK else -points
data.add_from_node(curboard, move_values, points)
if data.data_size % 1024 == 5:
queue.put("{}-- data size: {}".format(name, data.data_size))
if data.data_size > 12800:
queue.put(name + "--保存训练数据……")
data.save()
data.clear()
queue.put(name + "--保存完毕！")
curboard.play_move(move)
if qp:
qp.update(curboard)
score = curboard.win_rate if c == go.BLACK else -curboard.win_rate
cscore = score - final_score
cscore = cscore if c == go.BLACK else -cscore
if i % 20 == 15:
queue.put("{}--倒退{}步，全探查结果：{}，{}{}了{:.1f}子, {}步".format(
name, i + 1, go.result_str(score), go.get_color_str(c),
"增加" if cscore > 0 else "减少", abs(cscore), curboard.step))
def search_tree(self, node, board, depth=0):
depth += 1
if len(self.data) > 1000:
print("保存到数据库，depth:",
depth,
", self.data len:",
len(self.data),
file=sys.stderr)
self.sql.save(self.data)
self.data = []
if node.search_over:
return node.points / (node.visits if node.visits > 0 else 1)
if depth > go.N * go.N * 2:
node.search_over = True
return 0
p = self.sql.search(board)
if p > -9999:
node.update(p)
return p
move_values = [0.0 for i in range(go.N * go.N + 2)]
tomove = board.to_move
if node.expanded and (board.is_gameover or not node.childs):
node.search_over = True
score = strategies.fast_score(board)
ps = score if tomove == go.BLACK else -score
node.update(ps)
self.data.append(TrainData(board, ps))
move_values[-2] = 1.0
self.datas.add_from_node(board, move_values, ps)
if self.datas.data_size > 6400:
self.datas.save()
self.datas.clear()
return ps
node.points = 0.0
for child in node.childs:
if not config.running:
break
b = board.copy()
b.play_move(child.move)
v = 0.0
if not child.expanded:
self.expand(child, b)
v = self.search_tree(child, b, depth)
elif not child.search_over:
v = self.search_tree(child, b.depth)
ind = go.flatten_coords(child.move)
move_values[ind] = v
visits = node.visits if node.visits > 0 else 1
#maxnode = max(node.childs, key=lambda x:x.points)
#move_values[go.flatten_coords(maxnode.move)] = 1.0
if config.running and depth < int(go.N * go.N * 2):
node.search_over = True
self.data.append(TrainData(board, node.points / visits))
self.datas.add_from_node(board, move_values, node.points / visits)
if self.datas.data_size > 6400:
self.datas.save()
#print("保存训练数据，depth:", depth, ", self.datas len:", self.datas.data_size)
self.datas.clear()
return node.points / visits
def test_play(net, old_net, qp=None, sgfdir=None, bw=go.BLACK):
test_start = time.time()
pos=board.Board(komi=7.5)
if qp:
qp.start(pos)
move = None
passbw = 0
while config.running:
#time.sleep(0.5)
c = pos.to_move
if pos.step>go.N*go.N*2:
pos.result = 0
msg = '双方超过722手，对局作为和棋结束。'
if qp:
qp.show_message(msg=msg)
break
if c==bw:
move_probs, win_rate = net.run(pos)
else:
move_probs, win_rate = old_net.run(pos)
pos.win_rate= win_rate*go.N*go.N
move = strategies.select_most_likely(pos, move_probs, True, True) #True:检查是否合法, True:拒绝PASS
if move==go.RESIGN:
pos.result = -go.N*go.N-1 if c==1 else go.N*go.N+1
msg = '{}方第{}手认输，对局结束。{}中盘胜。'.format(go.get_color_str(c), pos.step, go.get_color_str(-c))
if qp:
qp.show_message(msg=msg)
break
if move is None or move == go.PASS:
move = go.PASS
passbw |= (1 if c==go.BLACK else 2)
if passbw==3:
pos.play_move(go.PASS)
msg = '%s方第%d手PASS，对局结束。' % (go.get_color_str(c), pos.step)
if qp:
qp.show_message(msg=msg)
break
else:
pos.play_move(go.PASS)
continue
illegal, caps = pos.play_move(move)
if illegal == 0:
passbw = 0
score = pos.score()
if qp:
qp.update(pos)
if pos.step>go.N*go.N*2:
return 0
if abs(pos.result)<go.N*go.N:
n = pos.step
score = strategies.fast_score(pos)
print("fast score: {}, n: {}/{}".format(score, n, pos.step), file=sys.stderr)
else:
score = pos.result*2
score *= 1 if bw==go.BLACK else -1
eloa = net.elo
elob = old_net.elo
ea=1/(1+10.0**((elob-eloa)/400))
eb=1/(1+10.0**((eloa-elob)/400))
win = 1 if score>0 else 0
xs = 14.5 + 17500.0/(6000+net.elo)
addea=xs*(win-ea)
addeb=xs*(1-win-eb)
print('old elo1: {:.2f}, 2: {:.2f}, add elo 1: {:.2f}, 2: {:.2f}'.format(
eloa, elob, addea, addeb), file=sys.stderr)
net.elo += addea
net.elo = max(-5000, net.elo)
old_net.elo += addeb
old_net.elo = max(-5000, old_net.elo)
pos.result = score / 2
msg = '新权重：{},用时：{},{}手，{}方胜{:.1f}子, 增加{:.2f}/{:.2f}分。'.format(go.get_color_str(bw),
time.strftime("%Mm%Ss", time.localtime(float(time.time() - test_start))),
pos.step+1, '新权重' if score>0 else '老权重', abs(score/2.0), addea, addeb)
if qp:
qp.show_message(msg=msg)
if sgfdir:
dt = sgfdir +'/test_' + time.strftime('%Y-%m-%d_%H_%M_%S')+'.sgf'
msg = '%.1f：\t保存sgf棋谱文件到%s' % (time.time()-policy.start_time, dt)
if qp:
qp.show_message(msg=msg)
if not os.path.exists(sgfdir):
os.makedirs(sgfdir)
sgfparser.save_board(pos, dt)
return score/2