def __init__(self,
model_path=None,
num_simul=800,
num_channel=128,
user=None):
# simul env
self.env_simul = tictactoe_env_simul.TicTacToeEnv()
# tree
self.tree = defaultdict(lambda: np.zeros((3, 3, 4), 'float'))
# model
self.pv_net = neural_net_5block.PolicyValueNet(num_channel).cuda()
if model_path is not None:
print(' ####### Model is loaded ####### ')
self.pv_net.load_state_dict(torch.load(model_path))
self.done = False
self.root = None
self.evaluate = None
self.player_color = None
self.num_simul = num_simul
self.user = user
# hyperparameter
self.c_puct = 5
self.epsilon = 0.25
self.alpha = 0.7
self.tau = None
# reset_step member
self.edge = None
self.total_visit = None
self.legal_move = None
self.no_legal_move = None
self.state = None
self.prob = None
self.value = None
self.current_user = None
# reset_episode member
self.node_memory = None
self.edge_memory = None
self.action_memory = None
self.action_count = None
# init
self.reset_step()
self._reset_episode()
def __init__(self, model_path=None):
# tree
self.tree = defaultdict(lambda: np.zeros((3, 3, 4), 'float'))
# model
self.pv_net = neural_net_5block.PolicyValueNet(CHANNEL)
if model_path is not None:
print(' ####### Model is loaded ####### ')
self.pv_net.load_state_dict(torch.load(model_path))
# hyperparameter
self.c_puct = 5
self.epsilon = 0.25
self.alpha = 0.7
# loop controller
self.done = False
# reset_step member
self.edge = None
self.total_visit = None
self.legal_move = None
self.no_legal_move = None
self.state = None
self.prob = None
self.value = None
self.current_user = None
# reset_episode member
self.node_memory = None
self.edge_memory = None
self.action_memory = None
self.action_count = None
# init
self.reset_step()
self._reset_episode()
EPOCHS = 64
BATCH_SIZE = 32
LR = 0.2
L2 = 0.0001
CHANNEL = 128
# data load
with open('data/train_dataset_s800_g800.pickle', 'rb') as f:
dataset = pickle.load(f)
train_dataset = data.DataLoader(dataset,
batch_size=BATCH_SIZE,
shuffle=True,
drop_last=True)
# 신경망 생성 및 최적화 인스턴스 생성
pv_net = neural_net_5block.PolicyValueNet(CHANNEL).cuda()
optimizer = torch.optim.SGD(pv_net.parameters(),
lr=LR,
momentum=0.9,
weight_decay=L2)
scheduler = lr_scheduler.ReduceLROnPlateau(optimizer,
'min',
min_lr=2e-4,
patience=5,
verbose=1)
# print spec
spec = {
'epoch': EPOCHS,
'batch size': BATCH_SIZE,
'optim': 'SGD',