def test_flatten(self):
self.assertEqual(coords.to_flat((0, 0)), 0)
self.assertEqual(coords.to_flat((0, 3)), 3)
self.assertEqual(coords.to_flat((3, 0)), 27)
self.assertEqual(coords.from_flat(27), (3, 0))
self.assertEqual(coords.from_flat(10), (1, 1))
self.assertEqual(coords.from_flat(80), (8, 8))
self.assertEqual(coords.to_flat(
coords.from_flat(10)), 10)
self.assertEqual(coords.from_flat(
coords.to_flat((5, 4))), (5, 4))
def test_flatten(self):
self.assertEqual(coords.to_flat(utils_test.BOARD_SIZE, (0, 0)), 0)
self.assertEqual(coords.to_flat(utils_test.BOARD_SIZE, (0, 3)), 3)
self.assertEqual(coords.to_flat(utils_test.BOARD_SIZE, (3, 0)), 27)
self.assertEqual(coords.from_flat(utils_test.BOARD_SIZE, 27), (3, 0))
self.assertEqual(coords.from_flat(utils_test.BOARD_SIZE, 10), (1, 1))
self.assertEqual(coords.from_flat(utils_test.BOARD_SIZE, 80), (8, 8))
self.assertEqual(coords.to_flat(
utils_test.BOARD_SIZE, coords.from_flat(utils_test.BOARD_SIZE, 10)), 10)
self.assertEqual(coords.from_flat(
utils_test.BOARD_SIZE, coords.to_flat(
utils_test.BOARD_SIZE, (5, 4))), (5, 4))
def test_proper_move_transform(self):
# Check that the reinterpretation of 362 = 19*19 + 1 during symmetry
# application is consistent with coords.from_flat
move_array = np.arange(go.N ** 2 + 1)
coord_array = np.zeros([go.N, go.N])
for c in range(go.N ** 2):
coord_array[coords.from_flat(c)] = c
for s in symmetries.SYMMETRIES:
with self.subTest(symmetry=s):
transformed_moves = apply_p(s, move_array)
transformed_board = apply_f(s, coord_array)
for new_coord, old_coord in enumerate(transformed_moves[:-1]):
self.assertEqual(
old_coord,
transformed_board[coords.from_flat(new_coord)])
def describe(self):
sort_order = list(range(go.N * go.N + 1))
sort_order.sort(key=lambda i: (
self.child_N[i], self.child_action_score[i]), reverse=True)
soft_n = self.child_N / sum(self.child_N)
p_delta = soft_n - self.child_prior
p_rel = p_delta / self.child_prior
# Dump out some statistics
output = []
output.append("{q:.4f}\n".format(q=self.Q))
output.append(self.most_visited_path())
output.append(
"move: action Q U P P-Dir N soft-N p-delta p-rel\n")
output.append("\n".join(["{!s:6}: {: .3f}, {: .3f}, {:.3f}, {:.3f}, {:.3f}, {:4d} {:.4f} {: .5f} {: .2f}".format(
coords.to_kgs(coords.from_flat(key)),
self.child_action_score[key],
self.child_Q[key],
self.child_U[key],
self.child_prior[key],
self.original_prior[key],
int(self.child_N[key]),
soft_n[key],
p_delta[key],
p_rel[key])
for key in sort_order][:15]))
return ''.join(output)
def maybe_add_child(self, fcoord):
""" Adds child node for fcoord if it doesn't already exist, and returns it. """
if fcoord not in self.children:
new_position = self.position.play_move(
coords.from_flat(fcoord))
self.children[fcoord] = MCTSNode(
new_position, fmove=fcoord, parent=self)
return self.children[fcoord]
def test_pass(self):
self.assertEqual(coords.from_sgf(''), None)
self.assertEqual(coords.from_flat(utils_test.BOARD_SIZE, 81), None)
self.assertEqual(coords.from_kgs(utils_test.BOARD_SIZE, 'pass'), None)
self.assertEqual(coords.from_pygtp(utils_test.BOARD_SIZE, (0, 0)), None)
self.assertEqual(coords.to_sgf(None), '')
self.assertEqual(coords.to_flat(utils_test.BOARD_SIZE, None), 81)
self.assertEqual(coords.to_kgs(utils_test.BOARD_SIZE, None), 'pass')
self.assertEqual(coords.to_pygtp(utils_test.BOARD_SIZE, None), (0, 0))
def test_topleft(self):
self.assertEqual(coords.from_sgf('ia'), (0, 8))
self.assertEqual(coords.from_flat(utils_test.BOARD_SIZE, 8), (0, 8))
self.assertEqual(coords.from_kgs(utils_test.BOARD_SIZE, 'J9'), (0, 8))
self.assertEqual(coords.from_pygtp(utils_test.BOARD_SIZE, (9, 9)), (0, 8))
self.assertEqual(coords.to_sgf((0, 8)), 'ia')
self.assertEqual(coords.to_flat(utils_test.BOARD_SIZE, (0, 8)), 8)
self.assertEqual(coords.to_kgs(utils_test.BOARD_SIZE, (0, 8)), 'J9')
self.assertEqual(coords.to_pygtp(utils_test.BOARD_SIZE, (0, 8)), (9, 9))
def test_upperleft(self):
self.assertEqual(coords.from_sgf('aa'), (0, 0))
self.assertEqual(coords.from_flat(utils_test.BOARD_SIZE, 0), (0, 0))
self.assertEqual(coords.from_kgs(utils_test.BOARD_SIZE, 'A9'), (0, 0))
self.assertEqual(coords.from_pygtp(utils_test.BOARD_SIZE, (1, 9)), (0, 0))
self.assertEqual(coords.to_sgf((0, 0)), 'aa')
self.assertEqual(coords.to_flat(utils_test.BOARD_SIZE, (0, 0)), 0)
self.assertEqual(coords.to_kgs(utils_test.BOARD_SIZE, (0, 0)), 'A9')
self.assertEqual(coords.to_pygtp(utils_test.BOARD_SIZE, (0, 0)), (1, 9))
def mvp_gg(self):
""" Returns most visited path in go-gui VAR format e.g. 'b r3 w c17..."""
node = self
output = []
while node.children and max(node.child_N) > 1:
next_kid = np.argmax(node.child_N)
node = node.children[next_kid]
output.append("%s" % coords.to_kgs(
coords.from_flat(node.fmove)))
return ' '.join(output)
def test_pass(self):
self.assertEqual(coords.from_sgf(''), None)
self.assertEqual(coords.from_flat(81), None)
self.assertEqual(coords.from_kgs('pass'), None)
self.assertEqual(coords.from_pygtp((0, 0)), None)
self.assertEqual(coords.to_sgf(None), '')
self.assertEqual(coords.to_flat(None), 81)
self.assertEqual(coords.to_kgs(None), 'pass')
self.assertEqual(coords.to_pygtp(None), (0, 0))
def test_topleft(self):
self.assertEqual(coords.from_sgf('ia'), (0, 8))
self.assertEqual(coords.from_flat(8), (0, 8))
self.assertEqual(coords.from_kgs('J9'), (0, 8))
self.assertEqual(coords.from_pygtp((9, 9)), (0, 8))
self.assertEqual(coords.to_sgf((0, 8)), 'ia')
self.assertEqual(coords.to_flat((0, 8)), 8)
self.assertEqual(coords.to_kgs((0, 8)), 'J9')
self.assertEqual(coords.to_pygtp((0, 8)), (9, 9))
def test_upperleft(self):
self.assertEqual(coords.from_sgf('aa'), (0, 0))
self.assertEqual(coords.from_flat(0), (0, 0))
self.assertEqual(coords.from_kgs('A9'), (0, 0))
self.assertEqual(coords.from_pygtp((1, 9)), (0, 0))
self.assertEqual(coords.to_sgf((0, 0)), 'aa')
self.assertEqual(coords.to_flat((0, 0)), 0)
self.assertEqual(coords.to_kgs((0, 0)), 'A9')
self.assertEqual(coords.to_pygtp((0, 0)), (1, 9))
def test_legal_moves(self):
board = test_utils.load_board('''
.O.O.XOX.
O..OOOOOX
......O.O
OO.....OX
XO.....X.
.O.......
OX.....OO
XX...OOOX
.....O.X.
''')
position = Position(board=board, to_play=BLACK)
illegal_moves = coords_from_kgs_set('A9 E9 J9')
legal_moves = coords_from_kgs_set('A4 G1 J1 H7') | {None}
for move in illegal_moves:
with self.subTest(type='illegal', move=move):
self.assertFalse(position.is_move_legal(move))
for move in legal_moves:
with self.subTest(type='legal', move=move):
self.assertTrue(position.is_move_legal(move))
# check that the bulk legal test agrees with move-by-move illegal test.
bulk_legality = position.all_legal_moves()
for i, bulk_legal in enumerate(bulk_legality):
with self.subTest(type='bulk', move=coords.from_flat(i)):
self.assertEqual(
bulk_legal, position.is_move_legal(coords.from_flat(i)))
# flip the colors and check that everything is still (il)legal
position = Position(board=-board, to_play=WHITE)
for move in illegal_moves:
with self.subTest(type='illegal', move=move):
self.assertFalse(position.is_move_legal(move))
for move in legal_moves:
with self.subTest(type='legal', move=move):
self.assertTrue(position.is_move_legal(move))
bulk_legality = position.all_legal_moves()
for i, bulk_legal in enumerate(bulk_legality):
with self.subTest(type='bulk', move=coords.from_flat(i)):
self.assertEqual(
bulk_legal, position.is_move_legal(coords.from_flat(i)))
def most_visited_path(self):
node = self
output = []
while node.children:
next_kid = np.argmax(node.child_N)
node = node.children.get(next_kid)
if node is None:
output.append("GAME END")
break
output.append("%s (%d) ==> " % (coords.to_kgs(
coords.from_flat(node.fmove)),
node.N))
output.append("Q: {:.5f}\n".format(node.Q))
return ''.join(output)
def most_visited_path(self):
node = self
output = []
while node.children:
next_kid = np.argmax(node.child_N)
node = node.children.get(next_kid)
if node is None:
output.append("GAME END")
break
output.append(
"%s (%d) ==> " %
(coords.to_kgs(coords.from_flat(node.fmove)), node.N))
output.append("Q: {:.5f}\n".format(node.Q))
return ''.join(output)
def pick_move(self):
'''Picks a move to play, based on MCTS readout statistics.
Highest N is most robust indicator. In the early stage of the game, pick
a move weighted by visit count; later on, pick the absolute max.'''
if self.root.position.n >= self.temp_threshold:
fcoord = np.argmax(self.root.child_N)
else:
cdf = self.root.child_N.cumsum()
cdf /= cdf[-2] # Prevents passing via softpick.
selection = random.random()
fcoord = cdf.searchsorted(selection)
assert self.root.child_N[fcoord] != 0
return coords.from_flat(fcoord)
def apply(action, history):
""" Apply the action as the next move of given history.
action: legal move, given as flat coordinates.
history: history of the game so far.
"""
board = history[-1].copy()
to_play = -1 if len(history) % 2 == 0 else 1
# if not pass
if action is not board.size:
p = go.Position(board=board, to_play=to_play)
p.play_move(coords.from_flat(action), mutate=True)
if p.ko is not None:
board[p.ko] = 4
history.append(board)
def pick_move(self):
'''Picks a move to play, based on MCTS readout statistics.
Highest N is most robust indicator. In the early stage of the game, pick
a move weighted by visit count; later on, pick the absolute max.'''
if self.root.position.n > self.temp_threshold:
fcoord = np.argmax(self.root.child_N)
else:
cdf = self.root.child_N.cumsum()
cdf /= cdf[-1]
selection = random.random()
fcoord = cdf.searchsorted(selection)
assert self.root.child_N[fcoord] != 0
return coords.from_flat(fcoord)
def most_visited_path(self):
node = self
output = []
while node.children:
next_kid = np.argmax(node.child_N)
node = node.children.get(next_kid)
if node is None:
output.append('GAME END')
break
output.append('{} ({}) ==> '.format(
coords.to_kgs(self.board_size,
coords.from_flat(self.board_size, node.fmove)),
node.N))
output.append('Q: {:.5f}\n'.format(node.Q))
return ''.join(output)
def most_visited_path(self):
node = self
output = []
while node.children:
next_kid = np.argmax(node.child_N)
node = node.children.get(next_kid)
if node is None:
output.append('GAME END')
break
output.append('{} ({}) ==> '.format(
coords.to_kgs(
self.board_size,
coords.from_flat(self.board_size, node.fmove)), node.N))
output.append('Q: {:.5f}\n'.format(node.Q))
return ''.join(output)
def extract_move_data(root_node, worker_id, completed_time, board_size):
current_node = root_node.next
move_data = []
move_num = 1
while current_node is not None:
props = current_node.properties
if 'B' in props:
to_play = 1
move_played = props['B'][0]
elif 'W' in props:
to_play = -1
move_played = props['W'][0]
else:
import pdb; pdb.set_trace()
move_played = coords.to_flat(coords.from_sgf(move_played))
post_Q, debug_rows = parse_comment_node(props['C'][0])
policy_prior = [0] * (board_size * board_size + 1)
policy_prior_orig = policy_prior[:]
mcts_visit_counts = policy_prior[:]
mcts_visit_counts_norm = policy_prior[:]
for debug_row in debug_rows:
move = debug_row.move
policy_prior[move] = debug_row.prior
policy_prior_orig[move] = debug_row.orig_prior
mcts_visit_counts[move] = debug_row.N
mcts_visit_counts_norm[move] = debug_row.soft_N
move_data.append({
'worker_id': worker_id,
'completed_time': completed_time,
'move_num': move_num,
'turn_to_play': to_play,
'move': move_played,
'move_kgs': coords.to_kgs(coords.from_flat(move_played)),
'prior_Q': None,
'post_Q': post_Q,
'policy_prior': policy_prior,
'policy_prior_orig': policy_prior_orig,
'mcts_visit_counts': mcts_visit_counts,
'mcts_visit_counts_norm': mcts_visit_counts_norm,
})
move_num += 1
current_node = current_node.next
return move_data
def run(self, model, position):
# assert position is of type Position from go.py
root = Node(0, position.to_play)
boards, playerCaps, opponentCaps = gamesToData([[position, 1]])
action_probs = model.callPol(boards, playerCaps, opponentCaps)[0]
value = model.callVal(boards, playerCaps, opponentCaps)[0]
valid_moves = position.all_legal_moves()
action_probs = action_probs * valid_moves # mask invalid moves
action_probs /= np.sum(action_probs)
root.expand(position, action_probs)
for _ in range(self.number_of_sim):
node = root
search_path = [node]
while node.expanded():
action, node = node.select_child()
search_path.append(node)
parent = search_path[-2]
position = parent.position
next_position = position.play_move(coords.from_flat(action))
if not next_position.is_game_over():
new_boards, new_playerCaps, new_opponentCaps = gamesToData(
[[next_position, 1]])
action_probs = model.callPol(
new_boards, new_playerCaps, new_opponentCaps)[0]
value = model.callVal(
new_boards, new_playerCaps, new_opponentCaps)[0]
valid_moves = next_position.all_legal_moves()
action_probs = action_probs * valid_moves # mask invalid moves
action_probs /= np.sum(action_probs)
node.expand(next_position, action_probs)
else:
if next_position.to_play == 1:
value = next_position.result()
else:
value = next_position.result()*-1
self.backpropagate(search_path, value, next_position.to_play)
return root
def _from_flat(flat_coords):
return coords.from_flat(utils_test.BOARD_SIZE, flat_coords)
def heatmap(self, sort_order, node, prop):
return "\n".join(["{!s:6} {}".format(
coords.to_kgs(coords.from_flat(key)),
node.__dict__.get(prop)[key])
for key in sort_order if node.child_N[key] > 0][:20])
def extract_move_data(root_node, worker_id, completed_time, board_size):
current_node = root_node.next
move_data = []
move_num = 1
while current_node is not None:
props = current_node.properties
if 'B' in props:
to_play = 1
move_played = props['B'][0]
elif 'W' in props:
to_play = -1
move_played = props['W'][0]
else:
import pdb
pdb.set_trace()
move_played = coords.to_flat(coords.from_sgf(move_played))
post_Q, debug_rows = parse_comment_node(props['C'][0])
def get_row_data(debug_row):
column_names = ["prior", "orig_prior", "N", "soft_N"]
return [getattr(debug_row, field) for field in column_names]
if FLAGS.only_top_move:
assert len(debug_rows) <= 1
row_data = list(map(get_row_data, debug_rows))
else:
row_data = [[0] * 4 for _ in range(board_size * board_size + 1)]
for debug_row in debug_rows:
move = debug_row.move
row_data[move] = get_row_data(debug_row)
policy_prior, policy_prior_orig, mcts_visits, mcts_visits_norm = \
zip(*row_data)
move_data.append({
'worker_id':
worker_id,
'completed_time':
completed_time,
'move_num':
move_num,
'turn_to_play':
to_play,
'move':
move_played,
'move_kgs':
coords.to_gtp(coords.from_flat(move_played)),
'prior_Q':
None,
'post_Q':
post_Q,
'policy_prior':
policy_prior,
'policy_prior_orig':
policy_prior_orig,
'mcts_visit_counts':
mcts_visits,
'mcts_visit_counts_norm':
mcts_visits_norm,
})
move_num += 1
current_node = current_node.next
return move_data
def _heatmap(self, sort_order, node, prop):
return "\n".join([
"{!s:6} {}".format(coords.to_kgs(coords.from_flat(key)),
node.__dict__.get(prop)[key])
for key in sort_order if node.child_N[key] > 0
][:20])
def play_mcts(network, board=None):
pos = Position(board=board)
player = get_mcts_player(network, pos)
node = player.root
children = node.rank_children()
soft_n = node.child_N / max(1, sum(node.child_N))
original_moves = {}
heatmap = np.zeros((N, N), dtype=np.float)
a_b = None
for i in children:
if node.child_N[i] == 0:
break
if a_b is None:
a_b = coords.from_flat(i)
original_moves[coords.to_gtp(coords.from_flat(i))] = soft_n[i]
a_b = player.pick_move()
# player.play_move(move)
a_b_coords = a_b
a_b = coords.to_gtp(a_b)
print(original_moves)
print("best action: ", a_b)
print(node.position)
p = original_moves[a_b]
print(p)
for i in range(N):
for j in range(N):
if board[i][j] == -1 or board[i][j] == 1:
new_board = np.copy(board)
new_board[i, j] = 0
new_pos = perturb_position(pos, new_board)
if new_pos.is_move_legal(a_b_coords):
player = get_mcts_player(network, new_pos)
node = player.root
print(node.position)
new_moves = {}
children = node.rank_children()
soft_n = node.child_N / max(1, sum(node.child_N))
for ch in children:
if node.child_N[ch] == 0:
break
new_moves[coords.to_gtp(coords.from_flat(ch))] = soft_n[ch]
new_a_b = player.pick_move()
# player.play_move(move)
new_a_b = coords.to_gtp(new_a_b)
# if new_a_b == 'F5':
print("---------------------")
# print("Moves: ", new_moves)
if a_b in new_moves:
new_p = new_moves[a_b]
else:
new_p = 0.
print("New best move", new_a_b)
print("p", new_p)
print("------------------")
K = cross_entropy_mcts(original_moves, new_moves, a_b)
if K == -1:
print("index", i, j)
heatmap[i, j] = -1.0
continue
dP = p - new_p
if dP > 0:
heatmap[i, j] = 2.0*dP/(1. + dP*K)
else:
heatmap[i, j] = -1.0
heatmap[heatmap == -1] = np.max(heatmap)
heatmap[heatmap<np.max(heatmap)/1.5] = 0
plt.imshow(heatmap, cmap='jet')
plt.colorbar()
plt.show()
return player
def play_network(network, board=None):
'''
Generates saliency maps of 3 methods given a board position
'''
pos = Position(board=board)
original_moves = {}
heatmap = np.zeros((N,N), dtype=np.float)
policy, V = network.run(pos)
best_move = np.argmax(policy)
print("Best Move is", coords.to_gtp(coords.from_flat(best_move)))
p = np.max(policy)
player = get_mcts_player(network, pos)
node = player.root
old_Q = node.child_Q[best_move]
atariV = np.zeros([N, N], dtype=np.float)
atariP = np.zeros([N, N], dtype=np.float)
delQ = np.zeros([N, N], dtype=np.float)
heatmap = np.zeros([N, N], dtype=np.float)
for i in range(N):
for j in range(N):
if board[i, j] == 1 or board[i, j] == -1:
print(i, j)
print("---------------------")
new_board = np.copy(board)
new_board[i, j] = 0
new_pos = perturb_position(pos, new_board)
new_policy, new_V = network.run(new_pos)
new_p = new_policy[best_move]
player = get_mcts_player(network, pos)
node = player.root
# print(node.describe())
new_Q = node.child_Q[best_move]
atariV[i, j] = 0.5*((V - new_V)**2)
atariP[i, j] = 0.5*np.linalg.norm(policy - new_policy)
dP = p - new_p
dQ = old_Q - new_Q
K = cross_entropy(policy, new_policy, best_move)
if dP>0:
heatmap[i, j] = 2*dP/(1 + dP*K)
if dQ>0:
delQ[i, j] = dQ
atariV = (atariV - np.min(atariV))/(np.max(atariV) - np.min(atariV))
atariP = (atariP - np.min(atariP))/(np.max(atariP) - np.min(atariP))
# heatmap[heatmap < np.max(heatmap)/3] = 0
# atariV[atariV < np.max(atariV)/3] = 0
# atariP[atariP < np.max(atariP)/3] = 0
# delQ[delQ < np.max(delQ)/3] = 0
frame = np.zeros((N,N,3))
frame = saliency_combine(atariV, frame, blur=256, channel=2)
frame = saliency_combine(atariP, frame, blur=256, channel=0)
plt.figure(1)
plt.imshow(atariV, cmap = 'Reds')
plt.colorbar()
plt.savefig(save_path + 'atariV.png')
plt.show()
plt.figure(2)
plt.imshow(atariP, cmap= 'Reds')
plt.colorbar()
plt.savefig(save_path + 'atariP.png')
plt.show()
plt.figure(3)
plt.imshow(frame)
plt.savefig(save_path + 'atari.png')
plt.show()
plt.figure(4)
plt.imshow(delQ, cmap = 'Reds')
plt.colorbar()
plt.savefig(save_path + 'deltaQ.png')
plt.show()
plt.figure(5)
plt.imshow(heatmap, cmap = 'Reds')
plt.colorbar()
plt.savefig(save_path + 'entropy.png')
plt.show()
def eval_player(player, positions, moves, results): probs, values = batch_run_many(player, positions) policy_moves = [coords.from_flat(c) for c in np.argmax(probs, axis=1)] top_move_agree = [moves[idx] == policy_moves[idx] for idx in range(len(moves))] square_err = (values - results)**2/4 return top_move_agree, square_err
def play(network, readouts, resign_threshold, verbosity=0):
''' Plays out a self-play match, returning
- the final position
- the n x 362 tensor of floats representing the mcts search probabilities
- the n-ary tensor of floats representing the original value-net estimate
where n is the number of moves in the game'''
player = MCTSPlayer(network,
resign_threshold=resign_threshold,
verbosity=verbosity,
num_parallel=SIMULTANEOUS_LEAVES)
global_n = 0
# Disable resign in 5% of games
if random.random() < 0.05:
player.resign_threshold = -1.0
player.initialize_game()
# Must run this once at the start, so that noise injection actually
# affects the first move of the game.
first_node = player.root.select_leaf()
prob, val = network.run(first_node.position)
# print("prob", prob)
# print("val", val)
first_node.incorporate_results(prob, val, first_node)
while True:
start = time.time()
player.root.inject_noise()
current_readouts = player.root.N
# we want to do "X additional readouts", rather than "up to X readouts".
while player.root.N < current_readouts + readouts:
player.tree_search()
if (verbosity >= 3):
print(player.root.position)
print(player.root.describe())
if player.should_resign():
player.set_result(-1 * player.root.position.to_play,
was_resign=True)
break
move = player.pick_move()
player.play_move(move)
if player.root.is_done():
player.set_result(player.root.position.result(), was_resign=False)
break
if (verbosity >= 2) or (verbosity >= 1
and player.root.position.n % 10 == 9):
print("Q: {:.5f}".format(player.root.Q))
dur = time.time() - start
print("%d: %d readouts, %.3f s/100. (%.2f sec)" %
(player.root.position.n, readouts, dur / readouts * 100.0,
dur),
flush=True)
if verbosity >= 3:
print("Played >>",
coords.to_kgs(coords.from_flat(player.root.fmove)))
if verbosity >= 2:
print("%s: %.3f" % (player.result_string, player.root.Q),
file=sys.stderr)
print(player.root.position,
player.root.position.score(),
file=sys.stderr)
return player
def _from_flat(flat_coords): return coords.from_flat(utils_test.BOARD_SIZE, flat_coords)
def play(network):
''' Plays out a self-play match, returning a MCTSPlayer object containing:
- the final position
- the n x 362 tensor of floats representing the mcts search probabilities
- the n-ary tensor of floats representing the original value-net estimate
where n is the number of moves in the game'''
readouts = FLAGS.num_readouts # defined in strategies.py
# Disable resign in 5% of games
if random.random() < FLAGS.resign_disable_pct:
resign_threshold = -1.0
else:
resign_threshold = None
player = MCTSPlayer(network, resign_threshold=resign_threshold)
player.initialize_game()
# Must run this once at the start to expand the root node.
first_node = player.root.select_leaf()
prob, val = network.run(first_node.position)
first_node.incorporate_results(prob, val, first_node)
while True:
start = time.time()
player.root.inject_noise()
current_readouts = player.root.N
# we want to do "X additional readouts", rather than "up to X readouts".
while player.root.N < current_readouts + readouts:
player.tree_search()
if FLAGS.verbose >= 3:
print(player.root.position)
print(player.root.describe())
if player.should_resign():
player.set_result(-1 * player.root.position.to_play,
was_resign=True)
break
move = player.pick_move()
player.play_move(move)
if player.root.is_done():
player.set_result(player.root.position.result(), was_resign=False)
break
if (FLAGS.verbose >= 2) or (FLAGS.verbose >= 1
and player.root.position.n % 10 == 9):
print("Q: {:.5f}".format(player.root.Q))
dur = time.time() - start
print("%d: %d readouts, %.3f s/100. (%.2f sec)" %
(player.root.position.n, readouts, dur / readouts * 100.0,
dur),
flush=True)
if FLAGS.verbose >= 3:
print("Played >>",
coords.to_kgs(coords.from_flat(player.root.fmove)))
if FLAGS.verbose >= 2:
utils.dbg("%s: %.3f" % (player.result_string, player.root.Q))
utils.dbg(player.root.position, player.root.position.score())
return player
def testAgainstRandom(model, matches):
#untrained models do not play at random, they have random weight initializations and then alawys play in terms of those
#this function takes a model (a trained one) and plays it against a player who makes a random move every time.
#it plays matches number of matches
veteran = model
veteranWins = 0
beginnerWins = 0
white = None
black = None
for i in range(matches):
if i % 2 == 0:
black = veteran
else:
white = veteran
position = go.Position()
while not position.is_game_over():
if position.n >= 100:
position = position.pass_move()
else:
if position.to_play == 1:
if black == veteran:
boards, playerCaps, opponentCaps = gamesToData(
[[position, 1]])
actions = black.callPol(boards, playerCaps,
opponentCaps)[0]
pdist = tf.nn.softmax(
tf.cast(actions, dtype=tf.float64))
legalMoves = position.all_legal_moves()
move = np.random.choice(np.arange(0, len(pdist)),
p=pdist)
if legalMoves[move] == 0:
actions = actions * legalMoves
move = tf.math.argmax(actions).numpy()
position = position.play_move(coords.from_flat(move))
else:
position = choose_and_play_move(position)
else:
if white == veteran:
boards, playerCaps, opponentCaps = gamesToData(
[[position, 1]])
actions = white.callPol(boards, playerCaps,
opponentCaps)[0]
pdist = tf.nn.softmax(
tf.cast(actions, dtype=tf.float64))
legalMoves = position.all_legal_moves()
move = np.random.choice(np.arange(0, len(pdist)),
p=pdist)
if legalMoves[move] == 0:
actions = actions * legalMoves
move = tf.math.argmax(actions).numpy()
position = position.play_move(coords.from_flat(move))
else:
position = choose_and_play_move(position)
if black == veteran:
if position.result() == 1:
veteranWins += 1
elif position.result() == -1:
beginnerWins += 1
else:
print("No one wins!!")
else:
if position.result() == 1:
beginnerWins += 1
elif position.result() == -1:
veteranWins += 1
else:
print("No one wins!!")
print("The model wins " + str(veteranWins))
print("The random wins " + str(beginnerWins))
return veteranWins - beginnerWins
def play(board_size, network, readouts, resign_threshold, simultaneous_leaves,
verbosity=0):
"""Plays out a self-play match.
Args:
board_size: the go board size
network: the DualNet model
readouts: the number of readouts in MCTS
resign_threshold: the threshold to resign at in the match
simultaneous_leaves: the number of simultaneous leaves in MCTS
verbosity: the verbosity of the self-play match
Returns:
the final position
the n x 362 tensor of floats representing the mcts search probabilities
the n-ary tensor of floats representing the original value-net estimate
where n is the number of moves in the game.
"""
player = MCTSPlayer(board_size, network, resign_threshold=resign_threshold,
verbosity=verbosity, num_parallel=simultaneous_leaves)
# Disable resign in 5% of games
if random.random() < 0.05:
player.resign_threshold = -1.0
player.initialize_game()
# Must run this once at the start, so that noise injection actually
# affects the first move of the game.
first_node = player.root.select_leaf()
prob, val = network.run(first_node.position)
first_node.incorporate_results(prob, val, first_node)
while True:
start = time.time()
player.root.inject_noise()
current_readouts = player.root.N
# we want to do "X additional readouts", rather than "up to X readouts".
while player.root.N < current_readouts + readouts:
player.tree_search()
if verbosity >= 3:
print(player.root.position)
print(player.root.describe())
if player.should_resign():
player.set_result(-1 * player.root.position.to_play, was_resign=True)
break
move = player.pick_move()
player.play_move(move)
if player.root.is_done():
player.set_result(player.root.position.result(), was_resign=False)
break
if (verbosity >= 2) or (
verbosity >= 1 and player.root.position.n % 10 == 9):
print("Q: {:.5f}".format(player.root.Q))
dur = time.time() - start
print("%d: %d readouts, %.3f s/100. (%.2f sec)" % (
player.root.position.n, readouts, dur / readouts * 100.0, dur))
if verbosity >= 3:
print("Played >>",
coords.to_kgs(coords.from_flat(player.root.fmove)))
if verbosity >= 2:
print("%s: %.3f" % (player.result_string, player.root.Q), file=sys.stderr)
print(player.root.position,
player.root.position.score(), file=sys.stderr)
return player