def self_play(tracker_queue,
net,
replay_queue,
probs_queue,
loop_count,
device="cpu"):
mcts_stores = [mcts.MCTS(), mcts.MCTS()]
for j in range(SELF_PLAY_PERIOD):
i = loop_count * SELF_PLAY_PERIOD + j
t = time.perf_counter()
status = ""
_, steps = model.play_game(mcts_stores,
replay_queue,
probs_queue,
net,
net,
steps_before_tau_0=STEPS_BEFORE_TAU_0,
mcts_searches=MCTS_SEARCHES,
mcts_batch_size=MCTS_BATCH_SIZE,
device=device,
status=status)
game_steps = steps
dt = time.perf_counter() - t
speed_steps = game_steps / dt
status = "episode #{}, steps {:3d}, processing time {:5.2f} [s], steps/s {:5.2f}".format(
i, steps, dt, speed_steps)
tracker_queue.put(("speed_steps", speed_steps, i))
print("episode #%d, steps %3d, steps/s %5.2f" %
(i, game_steps, speed_steps))
def evaluate(net1, net2, rounds, device="cpu"):
n1_win, n2_win = 0, 0
mcts_stores = [mcts.MCTS(), mcts.MCTS()]
for r_idx in range(rounds):
r, _ = model.play_game(mcts_stores=mcts_stores, replay_buffer=None, net1=net1, net2=net2,
steps_before_tau_0=0, mcts_searches=20, mcts_batch_size=16,
device=device)
if r < -0.5:
n2_win += 1
elif r > 0.5:
n1_win += 1
return n1_win / (n1_win + n2_win)
def eval(val, lock, net1, net2, device, cpuf):
if cpuf: net1.to(device); net2.to(device)
mcts_stores = [mcts.MCTS(), mcts.MCTS()]
while True:
are = random.randrange(0, 2)
r, _ = model.play_game(val, mcts_stores, None, net1=net1 if are<1 else net2,
net2=net2 if are<1 else net1, steps_before_tau_0=20,
mcts_searches=40, mcts_batch_size=40, best_idx=-1, device=device)
bf = False
lock.acquire()
if r!=None:
val[1 if (r > 0.5 and are<1) or (r<-0.5 and are>=1) else 2] += 1
print("%d:%d %d/%d"%(are,r,val[1],val[2]),end=' ', flush=True)
if (val[1]+val[2]) % 5 <1: print()
if val[0]<=0: bf=True
lock.release()
if bf: break
def __init__(self, model_file, player_moves_first, player_id):
self.model_file = model_file
self.model = model.Net(input_shape=model.OBS_SHAPE, actions_n=game.GAME_COLS)
self.model.load_state_dict(torch.load(model_file, map_location=lambda storage, loc: storage))
self.state = game.INITIAL_STATE
self.value = None
self.player_moves_first = player_moves_first
self.player_id = player_id
self.moves = []
self.mcts_store = mcts.MCTS()
def evaluate(net1, net2, rounds, device="cpu"):
n1_win, n2_win = 0, 0
mcts_stores = [mcts.MCTS(), mcts.MCTS()]
for r_idx in range(rounds):
r, step = model.play_game(None,
mcts_stores,
None,
net1 if r_idx < rounds // 2 else net2,
net2 if r_idx < rounds // 2 else net1,
steps_before_tau_0=game.MAX_TURN,
mcts_searches=40,
mcts_batch_size=40,
best_idx=-1,
device=device)
if (r > 0 and r_idx < rounds // 2) or (r < 0 and r_idx >= rounds // 2):
n1_win += 1
if r != 0: n2_win += 1
print(r_idx, r, step)
return (n1_win / n2_win) if n2_win > 0 else 0.5
parser.add_argument("--cuda", default=False, action="store_true", help="Enable CUDA")
args = parser.parse_args()
device = torch.device("cuda" if args.cuda else "cpu")
saves_path = os.path.join("saves", args.name)
os.makedirs(saves_path, exist_ok=True)
writer = SummaryWriter(comment="-" + args.name)
net = model.Net(input_shape=model.OBS_SHAPE, actions_n=game.GAME_COLS).to(device)
best_net = ptan.agent.TargetNet(net)
print(net)
optimizer = optim.SGD(net.parameters(), lr=LEARNING_RATE, momentum=0.9)
replay_buffer = collections.deque(maxlen=REPLAY_BUFFER)
mcts_store = mcts.MCTS()
step_idx = 0
best_idx = 0
with ptan.common.utils.TBMeanTracker(writer, batch_size=10) as tb_tracker:
while True:
t = time.time()
prev_nodes = len(mcts_store)
game_steps = 0
for _ in range(PLAY_EPISODES):
_, steps = model.play_game(mcts_store, replay_buffer, best_net.target_model, best_net.target_model,
steps_before_tau_0=STEPS_BEFORE_TAU_0, mcts_searches=MCTS_SEARCHES,
mcts_batch_size=MCTS_BATCH_SIZE, device=device)
game_steps += steps
game_nodes = len(mcts_store) - prev_nodes
dt = time.time() - t
def play_game(mcts_stores, replay_buffer, net1, net2, steps_before_tau_0, mcts_searches, mcts_batch_size,
net1_plays_first=None, cuda=False):
"""
Play one single game, memorizing transitions into the replay buffer
:param mcts_stores: could be None or single MCTS or two MCTSes for individual net
:param replay_buffer: queue with (state, probs, values), if None, nothing is stored
:param net1: player1
:param net2: player2
:return: value for the game in respect to player1 (+1 if p1 won, -1 if lost, 0 if draw)
"""
assert isinstance(replay_buffer, (collections.deque, type(None)))
assert isinstance(mcts_stores, (mcts.MCTS, type(None), list))
assert isinstance(net1, Net)
assert isinstance(net2, Net)
assert isinstance(steps_before_tau_0, int) and steps_before_tau_0 >= 0
assert isinstance(mcts_searches, int) and mcts_searches > 0
assert isinstance(mcts_batch_size, int) and mcts_batch_size > 0
if mcts_stores is None:
mcts_stores = [mcts.MCTS(), mcts.MCTS()]
elif isinstance(mcts_stores, mcts.MCTS):
mcts_stores = [mcts_stores, mcts_stores]
state = game.INITIAL_STATE
nets = [net1, net2]
if net1_plays_first is None:
cur_player = np.random.choice(2)
else:
cur_player = 0 if net1_plays_first else 1
step = 0
tau = 1 if steps_before_tau_0 > 0 else 0
game_history = []
result = None
net1_result = None
while result is None:
mcts_stores[cur_player].search_batch(mcts_searches, mcts_batch_size, state, cur_player, nets[cur_player], cuda=cuda)
probs, _ = mcts_stores[cur_player].get_policy_value(state, tau=tau)
game_history.append((state, cur_player, probs))
action = np.random.choice(game.GAME_COLS, p=probs)
if action not in game.possible_moves(state):
print("Impossible action selected")
state, won = game.move(state, action, cur_player)
if won:
result = 1
net1_result = 1 if cur_player == 0 else -1
break
cur_player = 1-cur_player
# check the draw case
if len(game.possible_moves(state)) == 0:
result = 0
net1_result = 0
break
step += 1
if step >= steps_before_tau_0:
tau = 0
if replay_buffer is not None:
for state, cur_player, probs in reversed(game_history):
replay_buffer.append((state, cur_player, probs, result))
result = -result
return net1_result, step
def play_game(net1, steps_before_tau_0, mcts_batch_size, device="cpu"):
assert isinstance(net1, model.Net)
assert isinstance(steps_before_tau_0, int) and steps_before_tau_0 >= 0
assert isinstance(mcts_batch_size, int) and mcts_batch_size > 0
global mcts_searches
pan = game.encode_lists([list(i) for i in game.INITIAL_STATE], 0)
historystr = []
cur_player = 0
step = 0
mctsi = mcts.MCTS()
result = None; exitf = False
a0 = ord('1')
while True:
s=input('플레이하려는 진영을 선택하세요 0) 초, 1)한 ?')
if s.find('level') >= 0:
mcts_searches = LEVELC * int(s[6:])
print('OK', flush=True)
else:
player_human = 0 if int(s)<1 else 1
break
while result is None:
movelist = game.possible_moves(pan, cur_player, step)
if step>9 and historystr[-4][:90]==historystr[-8][:90]:
p = game.decode_binary(pan)
for idx, m in enumerate(movelist):
spos = m // 100; tpos = m % 100; y0 = spos // 9; x0 = spos % 9; y1 = tpos // 9; x1 = tpos % 9
captured = p[y1][x1]; p[y1][x1] = p[y0][x0]; p[y0][x0] = 0
ps = game.encode_lists(p, step+1)
if ps[:90]==historystr[-4][:90]:
del movelist[idx]; break
p[y0][x0] = p[y1][x1]; p[y1][x1] = captured
if (step<2 and cur_player != player_human) or (step>1 and cur_player == player_human):
if step < 2:
print("마상 차림을 선택하세요 0) "+masang[0]+", 1) "+masang[1]+", 2) "+masang[2]+", 3) "+masang[3])
else:
render(pan, player_human)
if step==2 or step==3:
print("")
print("옮기고자 하는 기물의 세로 번호, 가로 번호, 목적지의 세로 번호, 가로 번호 ex) 0010 한수 쉬기: 0")
action = -1
while action<0:
s=input((str(step-1) if step>1 else '')+' ? ')
if s=="new": exitf=True; break
elif s.find('level')>=0:
mcts_searches=LEVELC*int(s[6:]); print('OK', flush=True)
elif step<2:
if len(s)==1 and s[0]>='0' and s[0]<'4': action = int(s) + 10000
elif len(s)==1: action = 0
elif s=='undo' and step>3:
step-=2; historystr.pop(); historystr.pop(); pan=historystr[-1]
movelist = game.possible_moves(pan, cur_player, step)
render(pan, player_human)
elif len(s)==4 and s[0]>='0' and s[0]<='9' and s[1]>'0' and s[1]<='9' and s[2]>='0' and s[2]<='9' and s[3]>'0' and s[3]<='9':
b1=9-ord(s[0])+a0 if s[0]>'0' else 0
if player_human<1: b1=9-b1
b2 = ord(s[1]) - a0
if player_human < 1: b2 = 8 - b2
b3 = 9-ord(s[2]) + a0 if s[2]>'0' else 0
if player_human < 1: b3 = 9 - b3
b4 = ord(s[3]) - a0
if player_human < 1: b4 = 8 - b4
action = (b1*9 + b2)*100 + b3*9+b4
if action not in movelist: action = -1
else: print('OK', flush=True)
else:
mctsi.search_batch(mcts_searches, mcts_batch_size, pan,
cur_player, net1, step, device=device)
probs, values = mctsi.get_policy_value(pan, movelist, cur_player)
chList = actionTable.choList if cur_player < 1 else actionTable.hanList
n = np.random.choice(actionTable.AllMoveLength, p=probs) if step<steps_before_tau_0 else np.argmax(probs)
action = chList[n]
"""for m in movelist:
print('%04d %.2f' % (m, probs[chList.index(m)]), end=', ')
print()"""
if step<2:
print(('한: ' if step<1 else '초: ')+masang[action-10000]+' '+str(values[n]), flush=True)
if step==1: render(pan, player_human)
else:
if action<1: print('한수쉼'+' '+str(values[n]))
else:
b1=action//100//9
if player_human<1: b1=9-b1
b2 = action//100%9
if player_human < 1: b2 = 8 - b2
b3 = action%100//9
if player_human < 1: b3 = 9 - b3
b4 = action%100%9
if player_human < 1: b4 = 8 - b4
print((chr(9-b1+a0) if b1>0 else '0')+chr(b2+a0)+(chr(9-b3+a0) if b3>0 else '0')+chr(b4+a0)+' '+str(values[n]))
if exitf: break
pan, won = game.move(pan, action, step)
historystr.append(pan)
if won>0:
render(pan, player_human)
print(('초' if won==1 else '한')+' 승')
break
cur_player = 1-cur_player
step += 1
def play_game(value, mcts_stores, queue, net1, net2, steps_before_tau_0, mcts_searches, mcts_batch_size,
best_idx, url=None, username=None, device="cpu"):
assert isinstance(mcts_stores, (mcts.MCTS, type(None), list))
assert isinstance(net1, Net)
assert isinstance(net2, Net)
assert isinstance(steps_before_tau_0, int) and steps_before_tau_0 >= 0
assert isinstance(mcts_searches, int) and mcts_searches > 0
assert isinstance(mcts_batch_size, int) and mcts_batch_size > 0
if mcts_stores is None:
mcts_stores = [mcts.MCTS(), mcts.MCTS()]
elif isinstance(mcts_stores, mcts.MCTS):
mcts_stores.clear()
mcts_stores = [mcts_stores, mcts_stores]
else: mcts_stores[0].clear(); mcts_stores[1].clear()
state = game.encode_lists([list(i) for i in game.INITIAL_STATE], 0)
nets = [net1, net2]
cur_player = 0
step = 0
tau = 1 if steps_before_tau_0 > 0 else 0
game_history = []
net1_result = None
result = None
while net1_result is None and (value==None or value[0]>0):
mcts_stores[cur_player].search_batch(mcts_searches, mcts_batch_size, state,
cur_player, nets[cur_player], step, device=device)
movel = game.possible_moves(state, cur_player, step)
probs, _ = mcts_stores[cur_player].get_policy_value(state, movel, cur_player, tau=tau)
chList = actionTable.choList if cur_player < 1 else actionTable.hanList
action = chList[np.random.choice(actionTable.AllMoveLength, p=probs)]
game_history.append((action, probs) if queue is None else (state, step, probs))
if action not in movel:
print("Impossible action selected")
state, won = game.move(state, action, step)
if step%3<1: print('.', end='', flush=True)
if won>0:
net1_result = 1 if won == 1 else -1
result = -net1_result
break
step += 1
cur_player = 1-cur_player
if step >= steps_before_tau_0:
tau = 0
if net1_result !=None:
print()
if queue is not None:
dequeuef = isinstance(queue, collections.deque)
for state, hstep, probs in game_history:
queue.append((state, hstep, probs, result)) if dequeuef else\
queue.put((state, hstep, probs, result))
if hstep!=1: result = -result
elif best_idx>=0:
gh = []
for (action, probs) in game_history:
prar = []
for idx, prob in enumerate(probs):
if prob>0: prar.append([idx, prob])
gh.append((action, prar))
js = {"netIdx":best_idx, "result":net1_result, "username":username, "action":gh}
hr = webFunction.http_request(url, True, json.dumps(js))
if hr == None: sys.exit()
elif hr['status'] == 'error': print('error occured')
else: print("game is uploaded")
return net1_result, step if net1_result!=None else 0
def play_game1(net1,
net2,
steps_before_tau_0,
mcts_searches,
mcts_batch_size,
device="cpu"):
"""
Play one single game, memorizing transitions into the replay buffer
:param net1: player1
:param net2: player2
:return: value for the game in respect to player1 (+1 if p1 won, -1 if lost, 0 if draw)
"""
# assert isinstance(replay_queue, (Queue, type(None)))
assert isinstance(net1, Net)
assert isinstance(net2, Net)
assert isinstance(steps_before_tau_0, int) and steps_before_tau_0 >= 0
mcts_stores = [mcts.MCTS(), mcts.MCTS()]
state = game_c.INITIAL_STATE
nets = [net1, net2]
cur_player = game.PLAYER_BLACK
step = 0
tau = 1 if steps_before_tau_0 > 0 else 0
game_history = []
result = None
net1_result = None
n_trees = [[], []]
pass_count = 0
while result is None:
mcts_stores[cur_player].search_batch(mcts_searches[cur_player],
mcts_batch_size[cur_player],
state,
cur_player,
nets[cur_player],
device=device)
n_trees[cur_player].append(len(mcts_stores[cur_player]))
probs, _ = mcts_stores[cur_player].get_policy_value(state, tau=tau)
game_history.append((state, cur_player, probs))
action = np.random.choice(game.BOARD_SIZE**2 + 1, p=probs)
del probs
if not game_c.is_possible_move(state, action):
# in this case, game.move function raise AssertionError
print("Impossible action selected")
state, field = game_c.move(state, action)
mcts_stores[0].clear_subtrees(state)
mcts_stores[1].clear_subtrees(state)
if action == game.BOARD_SIZE**2:
pass_count += 1
else:
pass_count = 0
if pass_count == 2 or (field != 2).all():
result = game_c.calc_result_f(field, cur_player)
net1_result = result if cur_player == 0 else -result
cur_player = 1 - cur_player
step += 1
if step >= steps_before_tau_0:
tau = 0
game_history.clear()
mcts_stores[0].clear()
mcts_stores[1].clear()
del game_history
del nets
mcts_stores.clear()
del mcts_stores
return net1_result, step, n_trees
def play_game(mcts_stores,
replay_queue,
probs_queue,
net1,
net2,
steps_before_tau_0,
mcts_searches,
mcts_batch_size,
device="cpu",
status=""):
"""
Play one single game, memorizing transitions into the replay buffer
:param mcts_stores: could be None or single MCTS or two MCTSes for individual net
:param replay_buffer: queue with (state, probs, values), if None, nothing is stored
:param net1: player1
:param net2: player2
:return: value for the game in respect to player1 (+1 if p1 won, -1 if lost, 0 if draw)
"""
# assert isinstance(replay_queue, (Queue, type(None)))
assert isinstance(mcts_stores, (mcts.MCTS, type(None), list))
assert isinstance(net1, Net)
assert isinstance(net2, Net)
assert isinstance(steps_before_tau_0, int) and steps_before_tau_0 >= 0
assert isinstance(mcts_searches, int) and mcts_searches > 0
assert isinstance(mcts_batch_size, int) and mcts_batch_size > 0
if mcts_stores is None:
mcts_stores = [mcts.MCTS(), mcts.MCTS()]
elif isinstance(mcts_stores, mcts.MCTS):
mcts_stores = [mcts_stores, mcts_stores]
state = game_c.INITIAL_STATE
nets = [net1, net2]
cur_player = game.PLAYER_BLACK
step = 0
tau = 1 if steps_before_tau_0 > 0 else 0
game_history = []
result = None
net1_result = None
pass_count = 0
while result is None:
t = time.perf_counter()
mcts_stores[cur_player].search_batch(mcts_searches,
mcts_batch_size,
state,
cur_player,
nets[cur_player],
device=device)
probs, _ = mcts_stores[cur_player].get_policy_value(state, tau=tau)
game_history.append((state, cur_player, probs))
action = np.random.choice(game.BOARD_SIZE**2 + 1, p=probs)
del probs
if not game_c.is_possible_move(state, action):
# in this case, game.move function raise AssertionError
print("Impossible action selected")
state, field = game_c.move(state, action)
mcts_stores[0].clear_subtrees(state)
mcts_stores[1].clear_subtrees(state)
if action == game.BOARD_SIZE**2:
# print("pass: player{}, #{}".format(cur_player, step + 1))
# render(state)
pass_count += 1
else:
pass_count = 0
# render(state)
# print(status)
# print("{:.2f} [s/move]".format(time.perf_counter() - t))
if pass_count == 2 or (field != 2).all() or \
(field != 0).all() or (field != 1).all():
result = game_c.calc_result_f(field, cur_player)
net1_result = result if cur_player == 0 else -result
cur_player = 1 - cur_player
step += 1
if step >= steps_before_tau_0:
tau = 0
if replay_queue is not None:
for state, cur_player, probs in reversed(game_history):
# replay_queue.put((state, cur_player, result))
# probs_queue.put(probs)
augment_data = game.augment_data(state, probs)
for arg_state, arg_probs in zip(*augment_data):
replay_queue.put((arg_state, cur_player, result))
probs_queue.put(arg_probs)
augment_data[0].clear()
augment_data[1].clear()
del state, cur_player, probs
result = -0.95 * result
game_history.clear()
mcts_stores[0].clear()
mcts_stores[1].clear()
del game_history
del nets
return net1_result, step
