def get_forwards(feature_csvs):
global x_iter, predicts
if not type(feature_csvs) is list:
feature_csvs = [feature_csvs]
x_tests = []
for feature_csv in feature_csvs:
feature = [int(x) for x in feature_csv.split(',')]
board, ko, turn = play_go.FromFeature((feature + [0, 0, 0])[:84])
feature = play_go.ToFeature(board, ko, turn, 0, 0, True, True)
x_test, _ = train_lib.parse_row(feature, True)
x_tests.append(np.asarray(x_test, dtype=np.float32))
x_iter.update(x_tests)
if predicts == None:
predicts = estimator.predict(
x=x_iter) # Only calls once with opened iterator.
#predicts = estimator.predict(x=np.asarray(x_tests, dtype=np.float32))
rets = []
for _ in range(len(x_tests)): # predict in predicts:
#for predict in predicts:
probabilities = predicts.next()['probabilities']
actions = []
for i in range(len(probabilities)):
actions.append([
str(train_lib.UnpackAction(i))
if i > 0 and i < 82 else ACTION_CODE[i],
float(probabilities[i])
])
rets.append(sorted(actions, key=lambda x: x[1], reverse=True)[:10])
if len(rets) == 1:
return rets[0]
return rets
def get_forwards(feature_csvs):
if not type(feature_csvs) is list:
feature_csvs = [feature_csvs]
x_tests = []
for feature_csv in feature_csvs:
feature = [int(x) for x in feature_csv.split(',')]
board, ko, turn = play_go.FromFeature(feature + [0])
feature = play_go.ToFeature(board, ko, turn, 0, 0, True, True)
x_test, _ = train_lib.parse_row(feature, True)
x_tests.append(np.asarray(x_test, dtype=np.float32))
predicts = estimator.predict(np.array(x_tests))
rets = []
for predict in predicts:
probabilities = list(predict['probabilities'])
rets.append([float(x) for x in probabilities])
if len(rets) == 1:
return rets[0]
return rets
num_sequence = int(sys.argv[2])
num_predict = int(sys.argv[3])
print('Test %d sequecne with model %s' % (num_sequence, model_dir))
# Set up kifu
kifus = []
stones = [0] * 81
for i in range(num_sequence):
stones[i] = (i % 2) * (-2) + 1
tail = [(num_sequence % 2) * 2 - 1, 0, 0]
# Shuffle board and add to kifu list
for _ in range(num_predict):
random.shuffle(stones)
board, last_move, ko = play_go.FromFeature(stones + tail)
kifus.append(play_go.ToFeature(board, last_move, ko, 0, True,
True)[:-1]) # remove result column
x_test = np.array(kifus, dtype=np.float32)
# Load model and predict
model_fn = importlib.import_module('%s.model_fn' % model_dir)
estimator = model_fn.GetEstimator(model_dir)
ds_predict_tf = estimator.predict(x_test)
# Print out human readable.
def PrintBoard(feature, pred):
print('%s, predict(W(-1)~B(1)): %f\n' %
(play_go.SPrintBoard(feature), pred))
# Print out human readable.
logging.warning('Drops too short game')
continue
winner = GetWinner(summary, sequence)
surrendered = WinBySurrender(winner, summary)
win_count[winner] = win_count[winner] + 1
board, ko, turn = play_go.InitBoard()
seq_cnt = 0
for move in sequence:
# Print before play.
if len(move) == 0:
continue
turn, action = ParseMove(move)
discount = min(1, seq_cnt / (len(sequence) * 1.0))
feature = play_go.ToFeature(board, ko, turn, action, winner,
RICH_OUTPUT, RICH_OUTPUT)
print(','.join(list(map(str, feature))))
valid, ko = play_go.PlayGo(board, turn, action)
# TODO(neochio): fix encode.
#if ko == None:
# ko = 0
#else:
# ko = ko[0] * 9 + ko[1] - 9
seq_cnt = seq_cnt + 1
if surrendered:
feature = play_go.ToFeature(board, ko, play_go.FlipTurn(turn),
play_go.SURRENDER, winner, RICH_OUTPUT,
RICH_OUTPUT)
print(','.join(list(map(str, feature))))
logging.warning('win_count: %s' % str(win_count))
if len(sys.argv) == 3:
epsilon = float(sys.argv[2])
STONE_CODE = ' BW'
POS_CODE = ' abcdefghi'
ACTION_CODE = { 0: 'P', 82: 'S' }
move_sequences = []
# Load model and predict
model_fn = importlib.import_module('%s.model_fn' % model_dir)
estimator = model_fn.GetEstimator(model_dir)
board, ko, turn = play_go.InitBoard()
passed = False
while True:
feature = play_go.ToFeature(board, ko, turn, 0, 0, True, True)
print(play_go.SPrintBoard(feature[:-1]))
x_test, _ = train_lib.parse_row(feature, True) # TODO: make configuable
predict = estimator.predict(np.asarray(x_test, dtype=np.float32))
probabilities = list(predict)[0]['probabilities']
actions = []
for i in range(len(probabilities)):
actions.append([train_lib.UnpackAction(i) if i > 0 and i < 82
else ACTION_CODE[i], probabilities[i]])
actions[0][1] = actions[0][1] / 5 # suppress pass
sorted_actions = sorted(actions, key=lambda x:x[1], reverse = True)
print(sorted_actions[:5])
if sorted_actions[0][0] == 'P':
move_sequences.append('%s[]' % STONE_CODE[turn])
int) or len(feature) == 1: # Select from candidates mode.
feature = int(feature)
if feature == 0:
feature = [0] * 81 + [-1] + [0] * 3
else:
feature = forwards[int(feature) -
1] + ',0' # TODO: unify result column
if isinstance(feature, (list, tuple)):
feature = list(feature)
else:
feature = feature.split(',')
feature = list(map(float, feature))[:-1] # Remove result column
board, last_move, ko = play_go.FromFeature(feature[:81] +
feature[-3:]) # Uses board only
x_test = np.array(
[play_go.ToFeature(board, last_move, ko, 0, True, True)[:-1]],
dtype=np.float32)
PrintBoard(feature, list(estimator.predict(x_test))[0])
# get all features one step forward
features = play_go.FowardFeatures(feature)
# Batch eval (for performance)
x_test = np.array(features, dtype=np.float32)
pred_tf = estimator.predict(x_test)
move_scores = {}
idx = 0
for pred in pred_tf:
move_scores[','.join(list(map(str, features[idx])))] = pred
idx = idx + 1
surrender = False
forwards = []
def genmove():
global board, ko, turn
feature = play_go.ToFeature(board, ko, turn, 0, 0)
return get_forwards(','.join(str(x) for x in feature[:-2]))
move_count = 0
print('=')
print
elif tokens[0] == 'play':
turn = TURN_STR[tokens[1]]
if tokens[2] != 'pass':
pos0 = tokens[2][0].replace('j', 'i')
pos = (ord(pos0) - ord('a') + 1) * 10 + int(tokens[2][1])
valid, ko = play_go.PlayGo(board, turn, pos)
turn = play_go.FlipTurn(turn)
move_count += 1
print('=')
print
elif tokens[0] == 'genmove':
turn = TURN_STR[tokens[1]]
feature = play_go.ToFeature(board, ko, turn)[:-2]
actions = s.get_forwards(','.join(list(map(str, feature))))
prob_sum = 0.0
for ap in actions:
a = ap[0]
if a in ('P', 'S'):
continue
prob_sum += ap[1]
if move_count > 10 and (actions[0][0] == 'P' or prob_sum == 0.0):
print('= PASS')
else:
rand = random.random() * prob_sum
for ap in actions:
a = ap[0]
if a in ('P', 'S'):
continue
