def test_chinese_handicap_handling(self):
intermediate_board = utils_test.load_board('''
.........
.........
......X..
.........
.........
.........
.........
.........
.........
''')
intermediate_position = go.Position(
utils_test.BOARD_SIZE,
intermediate_board,
n=1,
komi=5.5,
caps=(0, 0),
recent=(go.PlayerMove(go.BLACK, coords.from_kgs(
utils_test.BOARD_SIZE, 'G7')),),
to_play=go.BLACK,
)
final_board = utils_test.load_board('''
....OX...
.O.OOX...
O.O.X.X..
.OXXX....
OX...XX..
.X.XXO...
X.XOOXXX.
XXXO.OOX.
.XOOX.O..
''')
final_position = go.Position(
utils_test.BOARD_SIZE,
final_board,
n=50,
komi=5.5,
caps=(7, 2),
ko=None,
recent=(
go.PlayerMove(
go.WHITE, coords.from_kgs(utils_test.BOARD_SIZE, 'E9')),
go.PlayerMove(
go.BLACK, coords.from_kgs(utils_test.BOARD_SIZE, 'F9')),),
to_play=go.WHITE
)
positions_w_context = list(replay_sgf(
utils_test.BOARD_SIZE, CHINESE_HANDICAP_SGF))
self.assertEqualPositions(
intermediate_position, positions_w_context[1].position)
self.assertEqual(
positions_w_context[1].next_move, coords.from_kgs(
utils_test.BOARD_SIZE, 'C3'))
final_replayed_position = positions_w_context[-1].position.play_move(
positions_w_context[-1].next_move)
self.assertEqualPositions(final_position, final_replayed_position)
def test_japanese_handicap_handling(self):
intermediate_board = utils_test.load_board('''
.........
.........
......X..
.........
....O....
.........
..X......
.........
.........
''')
intermediate_position = go.Position(
utils_test.BOARD_SIZE,
intermediate_board,
n=1,
komi=5.5,
caps=(0, 0),
recent=(go.PlayerMove(go.WHITE, coords.from_kgs(
utils_test.BOARD_SIZE, 'E5')),),
to_play=go.BLACK,
)
final_board = utils_test.load_board('''
.........
.........
......X..
.........
....O....
.........
..XX.....
.........
.........
''')
final_position = go.Position(
utils_test.BOARD_SIZE,
final_board,
n=2,
komi=5.5,
caps=(0, 0),
recent=(
go.PlayerMove(go.WHITE, coords.from_kgs(
utils_test.BOARD_SIZE, 'E5')),
go.PlayerMove(go.BLACK, coords.from_kgs(
utils_test.BOARD_SIZE, 'D3')),),
to_play=go.WHITE,
)
positions_w_context = list(replay_sgf(
utils_test.BOARD_SIZE, JAPANESE_HANDICAP_SGF))
self.assertEqualPositions(
intermediate_position, positions_w_context[1].position)
final_replayed_position = positions_w_context[-1].position.play_move(
positions_w_context[-1].next_move)
self.assertEqualPositions(final_position, final_replayed_position)
def test_inference(self):
with tempfile.TemporaryDirectory() as working_dir, \
tempfile.TemporaryDirectory() as export_dir:
dualnet.bootstrap(working_dir, model_params.DummyMiniGoParams())
exported_model = os.path.join(export_dir, 'bootstrap-model')
dualnet.export_model(working_dir, exported_model)
n1 = dualnet.DualNetRunner(
exported_model, model_params.DummyMiniGoParams())
n1.run(go.Position(utils_test.BOARD_SIZE))
n2 = dualnet.DualNetRunner(
exported_model, model_params.DummyMiniGoParams())
n2.run(go.Position(utils_test.BOARD_SIZE))
def test_add_child(self):
root = MCTSNode(utils_test.BOARD_SIZE,
go.Position(utils_test.BOARD_SIZE))
child = root.maybe_add_child(17)
self.assertIn(17, root.children)
self.assertEqual(child.parent, root)
self.assertEqual(child.fmove, 17)
def test_only_check_game_end_once(self):
# When presented with a situation where the last move was a pass,
# and we have to decide whether to pass, it should be the first thing
# we check, but not more than that.
white_passed_pos = go.Position(utils_test.BOARD_SIZE, ).play_move(
(3, 3) # b plays
).play_move(
(3, 4) # w plays
).play_move(
(4, 3) # b plays
).pass_move() # w passes - if B passes too, B would lose by komi.
player = MCTSPlayerMixin(utils_test.BOARD_SIZE, DummyNet())
player.initialize_game(white_passed_pos)
# initialize the root
player.tree_search()
# explore a child - should be a pass move.
player.tree_search()
pass_move = utils_test.BOARD_SIZE * utils_test.BOARD_SIZE
self.assertEqual(player.root.children[pass_move].N, 1)
self.assertEqual(player.root.child_N[pass_move], 1)
player.tree_search()
# check that we didn't visit the pass node any more times.
self.assertEqual(player.root.child_N[pass_move], 1)
def test_is_game_over(self):
root = go.Position(utils_test.BOARD_SIZE)
self.assertFalse(root.is_game_over())
first_pass = root.play_move(None)
self.assertFalse(first_pass.is_game_over())
second_pass = first_pass.play_move(None)
self.assertTrue(second_pass.is_game_over())
def test_add_child_idempotency(self):
root = MCTSNode(utils_test.BOARD_SIZE,
go.Position(utils_test.BOARD_SIZE))
child = root.maybe_add_child(17)
current_children = copy.copy(root.children)
child2 = root.maybe_add_child(17)
self.assertEqual(child, child2)
self.assertEqual(current_children, root.children)
def initialize_game(self, position=None):
if position is None:
position = go.Position(self.board_size)
self.root = MCTSNode(self.board_size, position)
self.result = 0
self.result_string = None
self.comments = []
self.searches_pi = []
self.qs = []
def test_action_flipping(self):
np.random.seed(1)
probs = np.array([.02] *
(utils_test.BOARD_SIZE * utils_test.BOARD_SIZE + 1))
probs += np.random.random(
[utils_test.BOARD_SIZE * utils_test.BOARD_SIZE + 1]) * 0.001
black_root = MCTSNode(utils_test.BOARD_SIZE,
go.Position(utils_test.BOARD_SIZE))
white_root = MCTSNode(
utils_test.BOARD_SIZE,
go.Position(utils_test.BOARD_SIZE, to_play=go.WHITE))
black_root.select_leaf().incorporate_results(probs, 0, black_root)
white_root.select_leaf().incorporate_results(probs, 0, white_root)
# No matter who is to play, when we know nothing else, the priors
# should be respected, and the same move should be picked
black_leaf = black_root.select_leaf()
white_leaf = white_root.select_leaf()
self.assertEqual(black_leaf.fmove, white_leaf.fmove)
self.assertEqualNPArray(black_root.child_action_score,
white_root.child_action_score)
def test_tree_search_failsafe(self):
# Test that the failsafe works correctly. It can trigger if the MCTS
# repeatedly visits a finished game state.
probs = np.array([.001] *
(utils_test.BOARD_SIZE * utils_test.BOARD_SIZE + 1))
probs[-1] = 1 # Make the dummy net always want to pass
player = MCTSPlayerMixin(utils_test.BOARD_SIZE,
DummyNet(fake_priors=probs))
pass_position = go.Position(utils_test.BOARD_SIZE).pass_move()
player.initialize_game(pass_position)
player.tree_search(num_parallel=1)
self.assertNoPendingVirtualLosses(player.root)
def clear(self):
if self.position and len(self.position.recent) > 1:
try:
sgf = self.to_sgf()
with open(datetime.datetime.now().strftime(
'%Y-%m-%d-%H:%M.sgf'), 'w') as f:
f.write(sgf)
except NotImplementedError:
pass
except:
print('Error saving sgf', file=sys.stderr, flush=True)
self.position = go.Position(komi=self.komi)
self.initialize_game(self.position)
def test_dont_pick_unexpanded_child(self):
probs = np.array([0.001] *
(utils_test.BOARD_SIZE * utils_test.BOARD_SIZE + 1))
# make one move really likely so that tree search goes down that path twice
# even with a virtual loss
probs[17] = 0.999
root = MCTSNode(utils_test.BOARD_SIZE,
go.Position(utils_test.BOARD_SIZE))
root.incorporate_results(probs, 0, root)
leaf1 = root.select_leaf()
self.assertEqual(leaf1.fmove, 17)
leaf1.add_virtual_loss(up_to=root)
# the second select_leaf pick should return the same thing, since the child
# hasn't yet been sent to neural net for eval + result incorporation
leaf2 = root.select_leaf()
self.assertIs(leaf1, leaf2)
def test_do_not_explore_past_finish(self):
probs = np.array([0.02] *
(utils_test.BOARD_SIZE * utils_test.BOARD_SIZE + 1),
dtype=np.float32)
root = MCTSNode(utils_test.BOARD_SIZE,
go.Position(utils_test.BOARD_SIZE))
root.select_leaf().incorporate_results(probs, 0, root)
first_pass = root.maybe_add_child(
coords.to_flat(utils_test.BOARD_SIZE, None))
first_pass.incorporate_results(probs, 0, root)
second_pass = first_pass.maybe_add_child(
coords.to_flat(utils_test.BOARD_SIZE, None))
with self.assertRaises(AssertionError):
second_pass.incorporate_results(probs, 0, root)
node_to_explore = second_pass.select_leaf()
# should just stop exploring at the end position.
self.assertEqual(node_to_explore, second_pass)
def test_long_game_tree_search(self):
player = MCTSPlayerMixin(utils_test.BOARD_SIZE, DummyNet())
endgame = go.Position(utils_test.BOARD_SIZE,
board=TT_FTW_BOARD,
n=MAX_DEPTH - 2,
komi=2.5,
ko=None,
recent=(go.PlayerMove(go.BLACK, (0, 1)),
go.PlayerMove(go.WHITE, (0, 8))),
to_play=go.BLACK)
player.initialize_game(endgame)
# Test that an almost complete game
for i in range(10):
player.tree_search(num_parallel=8)
self.assertNoPendingVirtualLosses(player.root)
self.assertGreater(player.root.Q, 0)
def test_replay_position(self):
sgf_positions = list(sgf_wrapper.replay_sgf(
utils_test.BOARD_SIZE, NO_HANDICAP_SGF))
initial = sgf_positions[0]
self.assertEqual(initial.result, go.WHITE)
final = sgf_positions[-1].position.play_move(
sgf_positions[-1].next_move)
# sanity check to ensure we're working with the right position
final_board = utils_test.load_board('''
.OXX.....
O.OX.X...
.OOX.....
OOOOXXXXX
XOXXOXOOO
XOOXOO.O.
XOXXXOOXO
XXX.XOXXO
X..XOO.O.
''')
expected_final_position = go.Position(
utils_test.BOARD_SIZE,
final_board,
n=62,
komi=6.5,
caps=(3, 2),
ko=None,
recent=tuple(),
to_play=go.BLACK
)
self.assertEqualPositions(expected_final_position, final)
self.assertEqual(final.n, len(final.recent))
replayed_positions = list(go.replay_position(
utils_test.BOARD_SIZE, final, 1))
for sgf_pos, replay_pos in zip(sgf_positions, replayed_positions):
self.assertEqualPositions(sgf_pos.position, replay_pos.position)
def test_make_dataset_from_sgf(self):
with tempfile.NamedTemporaryFile() as sgf_file, \
tempfile.NamedTemporaryFile() as record_file:
sgf_file.write(TEST_SGF.encode('utf8'))
sgf_file.seek(0)
preprocessing.make_dataset_from_sgf(utils_test.BOARD_SIZE,
sgf_file.name,
record_file.name)
recovered_data = self.extract_data(record_file.name)
start_pos = go.Position(utils_test.BOARD_SIZE)
first_move = coords.from_sgf('fd')
next_pos = start_pos.play_move(first_move)
second_move = coords.from_sgf('cf')
expected_data = [
(features.extract_features(utils_test.BOARD_SIZE, start_pos),
preprocessing._one_hot(
utils_test.BOARD_SIZE,
coords.to_flat(utils_test.BOARD_SIZE, first_move)), -1),
(features.extract_features(utils_test.BOARD_SIZE, next_pos),
preprocessing._one_hot(
utils_test.BOARD_SIZE,
coords.to_flat(utils_test.BOARD_SIZE, second_move)), -1)
]
self.assertEqualData(expected_data, recovered_data)
tf.logging.set_verbosity(tf.logging.ERROR)
EMPTY_ROW = '.' * utils_test.BOARD_SIZE + '\n'
TEST_BOARD = utils_test.load_board('''
.X.....OO
X........
XXXXXXXXX
''' + EMPTY_ROW * 6)
TEST_POSITION = go.Position(
utils_test.BOARD_SIZE,
board=TEST_BOARD,
n=3,
komi=6.5,
caps=(1, 2),
ko=None,
recent=(go.PlayerMove(go.BLACK, (0, 1)), go.PlayerMove(go.WHITE, (0, 8)),
go.PlayerMove(go.BLACK, (1, 0))),
to_play=go.BLACK,
)
TEST_BOARD2 = utils_test.load_board('''
.XOXXOO..
XO.OXOX..
XXO..X...
''' + EMPTY_ROW * 6)
TEST_POSITION2 = go.Position(
utils_test.BOARD_SIZE,
board=TEST_BOARD2,
.XXOOOOOO
X.XOO...O
.XXOO...O
X.XOO...O
.XXOO..OO
X.XOOOOOO
.XXOOOOOO
X.XXXXXXX
XXXXXXXXX
''')
SEND_TWO_RETURN_ONE = go.Position(utils_test.BOARD_SIZE,
board=ALMOST_DONE_BOARD,
n=70,
komi=2.5,
caps=(1, 4),
ko=None,
recent=(go.PlayerMove(go.BLACK, (0, 1)),
go.PlayerMove(go.WHITE, (0, 8))),
to_play=go.BLACK)
# 505 moves for 19x19, 113 for 9x9
MAX_DEPTH = (utils_test.BOARD_SIZE**2) * 1.4
class DummyNet():
def __init__(self, fake_priors=None, fake_value=0):
if fake_priors is None:
fake_priors = np.ones((utils_test.BOARD_SIZE**2) +
1) / (utils_test.BOARD_SIZE**2 + 1)
self.fake_priors = fake_priors
