def go_cmp(binary=False):
b = board
if not b.turn:
b = b.mirror()
legal_moves = list(b.legal_moves)
best_move = legal_moves[0]
for _ in range(5):
for mv in legal_moves:
if mv is best_move:
continue
b.push(best_move)
best = util.state_to_cnn(util.board_to_state(b))
b.pop()
b.push(mv)
comp = util.state_to_cnn(util.board_to_state(b))
b.pop()
with torch.no_grad():
model.train()
t = torch.tensor(np.array([best, comp]),
dtype=torch.float).to(device)
x = model(t).detach().cpu()
if not binary and x[0].argmax() < 10:
best_move = mv
elif binary and x[0] > 0:
best_move = mv
uci = best_move.uci()
if not board.turn:
uci = util.flip_uci(uci)
print("bestmove " + uci)
def go_state(mode='buckets'):
global board
b = board
if not board.turn:
b = board.mirror()
moves = list(b.legal_moves)
cnn = []
for mv in moves:
b.push(mv)
state = util.board_to_state(b)
cnn.append(util.state_to_cnn(state))
b.pop()
model.train()
with torch.no_grad():
y = model(torch.tensor(cnn, dtype=torch.float, device=device)).detach()
idx = 0
if mode is 'buckets':
idx = y.argmax(dim=1).argmax(dim=0).cpu().numpy()
elif mode is 'value':
idx = y.argmax().cpu().numpy()
uci = moves[idx].uci()
if not board.turn:
uci = util.flip_uci(uci)
print("bestmove " + uci)
def precompute(self):
with open("data/depth18_gamma0.200000/moves_%s.csv" % (self.mode),
"r") as csv_file:
r = csv.reader(csv_file)
self.n_items = 0
self.cnn = []
self.cp_loss = []
self.mask = []
self.legal_mask = []
self.num_of_splits = 0
print('Loading data...')
for fen_without_count, dstr in progressbar.progressbar(r):
cpdict = ast.literal_eval(dstr)
fen = fen_without_count + " 0 1"
board = chess.Board(fen=fen)
state = util.board_to_state(board)
legal_mask = util.movelist_to_actionmask(board.legal_moves)
cnn = util.state_to_cnn(state)
cp_loss, mask = util.cpdict_to_loss_mask(cpdict)
self.cnn.append(cnn)
self.cp_loss.append(cp_loss)
self.mask.append(mask)
self.legal_mask.append(legal_mask)
self.n_items += 1
if self.n_items % 10000 == 0:
self.save_precomputed()
self.save_precomputed()
def random_move(boards, p_dict):
cnn_t = torch.zeros(n_par_games,
17,
8,
8,
dtype=torch.float,
device=device)
mask_t = torch.zeros(n_par_games,
64 * 64,
dtype=torch.float,
device=device)
fen_without_count = [None for _ in boards]
for i, b in enumerate(boards):
turn = b.turn
#p = None
#fen_without_count[i] = ' '.join(board.fen().split()[:-2])
#if fen_without_count[i] in p_dict.keys():
# p = p_dict[fen_without_count[i]]
if not turn:
b = b.mirror()
state = util.board_to_state(b)
cnn = util.state_to_cnn(state)
cnn_t[i] = torch.tensor(cnn).to(device)
mask_t[i] = torch.tensor(util.movelist_to_actionmask(b.legal_moves),
dtype=torch.float,
device=device)
y = model(cnn_t).detach()
y_masked = torch.nn.functional.softmax(y, dim=1) * mask_t
y_masked = torch.nn.functional.normalize(y_masked, p=1)
p = y_masked.cpu().numpy()
ucis = []
for i, b in enumerate(boards):
#p_dict[fen_without_count[i]] = p[i]
idx = np.random.choice(64 * 64, p=p[i])
uci = util.idx_to_uci(idx)
if not b.turn:
uci = util.flip_uci(uci)
try:
b.push_uci(uci)
b.pop()
except:
uci += np.random.choice(['q', 'r', 'b', 'n'])
ucis.append(uci)
return ucis
def __init__(self):
super(ChessMoveDataset_cp,self).__init__()
with open("data/depth18_gamma0.200000/moves.csv", "r") as csv_file:
r = csv.reader(csv_file)
data = []
print('Loading data...')
for fen_without_count,dstr in progressbar.progressbar(r):
cpdict = ast.literal_eval(dstr)
fen = fen_without_count + " 0 1"
state = util.board_to_state(chess.Board(fen=fen))
cnn = util.state_to_cnn(state)
cp_loss,mask = util.cpdict_to_loss_mask(cpdict)
data.append((cnn,cp_loss,mask))
self.data = np.array(data)
def __iter__(self):
it = None
worker_info = torch.utils.data.get_worker_info()
if worker_info is None:
it = range(0,self.num_of_splits,1)
else:
worker_id = worker_info.id
num_workers = worker_info.num_workers
it = range(worker_id, self.num_of_splits, num_workers)
for idx in it:
b = np.load('data/pre_pov/boards_%s_%d.npy'%(self.mode,idx))
m = np.load('data/pre_pov/moves_%s_%d.npy'%(self.mode,idx))
for j in range(len(b)):
yield torch.tensor(util.state_to_cnn(b[j]),dtype=torch.float) ,torch.tensor(m[j], dtype=torch.long)
def precompute(self):
with open("data/depth18_gamma0.200000/moves_statevalue_%s.csv"%(self.mode), "r") as csv_file:
r = csv.reader(csv_file)
self.n_items = 0
self.cnn = []
self.values = []
self.num_of_splits = 0
print('Loading data...')
for fen,value in progressbar.progressbar(r):
board = chess.Board(fen=fen)
state = util.board_to_state(board)
cnn = util.state_to_cnn(state)
self.cnn.append(cnn)
self.values.append(int(value))
self.n_items += 1
if self.n_items % 10000 == 0:
self.save_precomputed()
self.save_precomputed()
def go():
global board
b = board
if not board.turn:
b = board.mirror()
state = util.board_to_state(b)
cnn = util.state_to_cnn(state)
model.train()
with torch.no_grad():
d = {}
for _ in range(100):
y = model(
torch.tensor(cnn, dtype=torch.float,
device=device).unsqueeze(0)).detach()
y = y - y.min()
y_masked = y.cpu().numpy() * util.movelist_to_actionmask(
b.legal_moves)
uci = util.action_to_uci(y_masked)
if not board.turn:
uci = util.flip_uci(uci)
try:
board.push_uci(uci)
board.pop()
except:
uci += 'q'
if uci not in d.keys():
d[uci] = 1
else:
d[uci] += 1
maxcount, bestuci = 0, None
for uci, count in d.items():
if maxcount < count:
bestuci = uci
maxcount = count
print("bestmove " + uci)