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 load_dataset(filename):
with gfile.Open(filename) as csv_file:
data_file = csv.reader(csv_file)
data, target = [], []
for row in data_file:
if len(row) >= 81 * 13:
x, y = row[:81 * 13], int(row[-1])
else:
x, y = train_lib.parse_row(row, produce_dominion)
data.append(np.asarray(x, dtype=np.float32))
target.append(y)
return data, target
def load_dataset(filename):
with gfile.Open(filename) as csv_file:
data_file = csv.reader(csv_file)
data, target = [], []
for row in data_file:
if len(row) >= 81 * 13:
x, y = row[:81 * 13], row[-(len(row) % 81):]
else:
x, y = train_lib.parse_row(row, produce_dominion)
data.append(np.asarray(x, dtype=np.float32))
target.append(int(y[0])) # Uses action only on policy net.
return data, target
def load_csv(filename):
line_cnt = sum(1 for row in open(filename))
with gfile.Open(filename) as csv_file:
data_file = csv.reader(csv_file)
data, action, result = [''] * line_cnt, [0] * line_cnt, [0.0
] * line_cnt
idx = 0
for row in data_file:
if len(row) >= 81 * 13:
x, a, r = row[:81 * 13], row[81 * 13], row[81 * 13 + 1]
else:
# TODO: deprecated.
x, y = train_lib.parse_row(row, produce_dominion)
data[idx] = np.asarray(x, dtype=np.float32)
action[idx] = int(a)
result[idx] = float(r)
idx += 1
return data, action, result
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
# Generates training data from rich feature csv.
# As feature getting larger, it takes 30 minutes to initialize.
#
# Usage: python generate_train_data.py <in_csv>
import numpy as np
import csv
import sys
from tensorflow.python.platform import gfile
import train_lib
in_csv = sys.argv[1]
with gfile.Open(in_csv) as inf:
data_file = csv.reader(inf)
data, target = [], []
for row in data_file:
x, y = train_lib.parse_row(row, True)
print(','.join(list(map(str, x + [y]))))
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])
ko = 0
if passed: