def run_game(b1_role, mcts_role, b1_N, mcts_N, model, rand=0):
print(f'b1 ({" ".join(b1_role)}), N={b1_N}')
print(f'mcts ({" ".join(mcts_role)}), N={mcts_N}')
curb1 = B1Node(propnet, data, model=model)
curmcts = MCTSNode(propnet, data)
board = [list('.' * 8) for i in range(6)]
for step in range(1000):
print(*(''.join(b) for b in board[::-1]), sep='\n')
legal = curb1.propnet.legal_moves_dict(curb1.data)
b1_moves = choose_move(curb1, b1_role, b1_N, legal, step < rand)
mcts_moves = choose_move(curmcts, mcts_role, mcts_N, legal,
step < rand)
taken_moves = dict(
list(zip(b1_role, b1_moves)) + list(zip(mcts_role, mcts_moves)))
moves = tuple(taken_moves[role] for role in propnet.roles)
curb1 = curb1.get_or_make_child(moves)
curmcts = curmcts.get_or_make_child(moves)
print('Moves were:')
for move in propnet.legal:
if move.id in moves and move.move_gdl.strip() != 'noop':
print(move.move_role, move.move_gdl)
if 'drop' in move.move_gdl:
col = int(move.move_gdl.split()[2]) - 1
for i in range(len(board)):
if board[i][col] == '.':
board[i][col] = move.move_role[0]
break
if curb1.terminal:
print(*(''.join(b) for b in board[::-1]), sep='\n')
break
print('Results:', curb1.scores)
return (sum(curb1.scores[role] for role in b1_role),
sum(curb1.scores[role] for role in mcts_role))
def play(self, endtime):
root = B1Node(self.propnet, model=self.model)
for i in range(500):
simulation(root)
root.print_node()
best, choice = -1, None
for i, c in root.move_counts[self.role].items():
if c > best:
best, choice = c, i
move = self.propnet.id_to_move[choice].move_gdl
print('Made move', move)
return move
import tensorflow as tf
import re
import pickle
if len(sys.argv) < 4:
print("Usage: transfer.py <game> <ckpt> <total_rounds>")
game = sys.argv[1]
expertckpt = sys.argv[2]
totalRounds = int(sys.argv[3])
data, propnet = load_propnet(game)
m = Model(propnet, create=False)
m.load('./models/' + game + '/step-%06d.ckpt' % int(expertckpt))
cur = [None]
start = time.time()
gameIndex = 0
for i in range(totalRounds + 1):
cur[0] = B1Node(propnet, data, model=m)
print('Game number', i)
start_game = time.time()
do_game(cur, propnet, m, z=0.5)
print("took ", time.time() - start_game, "seconds to play game")
pickle.dump(
m.getBuffer(),
open(
"buffers/" + game + expertckpt + "for" + str(totalRounds) + "rounds.p",
"wb"))
model.perform_transfer('./models/' + from_game+ '/step-%06d.ckpt' % int(ckpt), reuse_output=reuse, pad_0=zeroPad, breakthroughMap=breakthroughMap)
if clear:
model.clear_output_layer()
## train
cur = [None]
set_pauser({
'cur': cur,
'model': model,
'propnet': new_prop,
})
model_name = to_game + "from" + from_game + ckpt + mode
model.save(model_name,0, transfer=True)
start = time.time()
for i in range(int(train_to)):
cur[0] = B1Node(new_prop, data, model=model)
print('Game number', i)
start_game = time.time()
do_game(cur, new_prop, model, z=0.5)
print("took ", time.time()-start_game, "seconds to play game")
start_train = time.time()
model.train(epochs=10)
print("took ", time.time()-start_train, "seconds to train")
if i and i % 50 == 0:
model.save(model_name, i, transfer=True)
with open(f'models_transfer/times-{to_game}', 'a') as f:
f.write(f'{i} {time.time()-start}\n')
human_win = myfont.render("You won!", False, (0, 0, 0))
draw = myfont.render("Draw!", False, (0, 0, 0))
size = [1550, 1000]
screen = pygame.display.set_mode(size)
pygame.display.set_caption('Connect-4 against AlphaZero')
clock = pygame.time.Clock()
HUMAN = your_role[0]
my_role = list(set(propnet.roles) - {your_role})[0]
COMPUTER = my_role[0]
left_player, right_player = human_text, a0_text
N = 500
col = None
done = False
cur = B1Node(propnet, data, model=model)
state = [list('.' * 7) for i in range(6)]
if my_role == 'white':
show_state(state)
move = 3 # get_computer_move()
drop(state, move, COMPUTER)
moves = make_moves(move, None)
cur = cur.get_or_make_child(moves)
left_player, right_player = right_player, left_player
while not done:
for event in pygame.event.get(): # User did something
if event.type == pygame.QUIT: # If user clicked close
done = True