def mini_batch(hcpevec):
features1 = np.empty((len(hcpevec), FEATURES1_NUM, 9, 9), dtype=np.float32)
features2 = np.empty((len(hcpevec), FEATURES2_NUM, 9, 9), dtype=np.float32)
move = np.empty((len(hcpevec)), dtype=np.int32)
result = np.empty((len(hcpevec)), dtype=np.int32)
value = np.empty((len(hcpevec)), dtype=np.float32)
cppshogi.hcpe_decode_with_value(hcpevec, features1, features2, move,
result, value)
return (Variable(cuda.to_gpu(features1)), Variable(cuda.to_gpu(features2)),
Variable(cuda.to_gpu(result.reshape((len(hcpevec), 1)))),
Variable(cuda.to_gpu(value.reshape((len(value), 1)))))
def __getitem__(self, idx):
feature1 = np.empty((FEATURES1_NUM, 9, 9), dtype=np.float32)
feature2 = np.empty((FEATURES2_NUM, 9, 9), dtype=np.float32)
move = np.empty(1, dtype=np.int32)
result = np.empty(1, dtype=np.float32)
value = np.empty_like(result)
# 要素を普通に取り出すとnp.void型になってしまう
cppshogi.hcpe_decode_with_value(
self.data[idx:idx + 1], feature1, feature2, move, result, value
)
z = result - value + 0.5
return feature1, feature2, np.int64(move), result, z, value
def mini_batch(self, hcpevec):
cppshogi.hcpe_decode_with_value(
hcpevec, self.features1, self.features2, self.move,
self.result, self.value
)
# aobaで評価値がない場合はvalueが-1になるようにした
# 負のadvantageは良くないように見えるので、
# そのままのvalueでpolicy gradientの目的関数にする
z = np.where(self.value == -1, 0.5, self.value)
# z = self.result - self.value + 0.5
return (self.torch_features1.to(self.device),
self.torch_features2.to(self.device),
self.torch_move.to(self.device),
self.torch_result.to(self.device),
torch.tensor(z).to(self.device),
self.torch_value.to(self.device))
def mini_batch(hcpevec):
features1 = np.empty((len(hcpevec), FEATURES1_NUM, 9, 9), dtype=np.float32)
features2 = np.empty((len(hcpevec), FEATURES2_NUM, 9, 9), dtype=np.float32)
move = np.empty((len(hcpevec)), dtype=np.int32)
result = np.empty((len(hcpevec)), dtype=np.float32)
value = np.empty((len(hcpevec)), dtype=np.float32)
cppshogi.hcpe_decode_with_value(hcpevec, features1, features2, move,
result, value)
z = result.astype(np.float32) - value + 0.5
return (torch.tensor(features1).to(device),
torch.tensor(features2).to(device),
torch.tensor(move.astype(np.int64)).to(device),
torch.tensor(result.reshape(
(len(hcpevec), 1))).to(device), torch.tensor(z).to(device),
torch.tensor(value.reshape((len(value), 1))).to(device))
parser = argparse.ArgumentParser()
parser.add_argument('--test_data', type=str, default=r'H:\src\DeepLearningShogi\x64\Release_NoOpt\test.hcpe', help='test data file')
parser.add_argument('--initmodel', '-m', default=r'H:\src\DeepLearningShogi\dlshogi\model_rl_val_wideresnet10_110_1', help='Initialize the model from given file')
args = parser.parse_args()
model = PolicyValueNetwork()
model.to_gpu()
print('Load model from', args.initmodel)
serializers.load_npz(args.initmodel, model)
hcpevec = np.fromfile(args.test_data, dtype=HuffmanCodedPosAndEval)
features1 = np.empty((len(hcpevec), FEATURES1_NUM, 9, 9), dtype=np.float32)
features2 = np.empty((len(hcpevec), FEATURES2_NUM, 9, 9), dtype=np.float32)
move = np.empty((len(hcpevec)), dtype=np.int32)
result = np.empty((len(hcpevec)), dtype=np.int32)
value = np.empty((len(hcpevec)), dtype=np.float32)
cppshogi.hcpe_decode_with_value(hcpevec, features1, features2, move, result, value)
x1 = Variable(cuda.to_gpu(features1))
x2 = Variable(cuda.to_gpu(features2))
with chainer.no_backprop_mode():
with chainer.using_config('train', False):
y1, y2 = model(x1, x2)
print(y1.data)
print(F.sigmoid(y2).data)