def pretrain(self,
memory_pretrained_fn,
batch_size=512,
iterations=int(1e5)):
memory_pretrained = pickle.load(open(memory_pretrained_fn, 'rb'))
self._memory = ReplayMemory(len(memory_pretrained))
self._memory.memory = memory_pretrained
self.train_q_batch_size = batch_size
start_time = time.time()
for i in range(iterations):
if i % 500 == 0:
if i > 1:
for key in self._Qt.state_dict():
try:
print(
sum(self._Qt.state_dict()[key] -
self._Q.state_dict()[key]))
except TypeError as e:
print(e)
self._Qt = copy.deepcopy(self._Q)
if i % 10000 == 0:
torch.save(self._Q.state_dict(), f'{self.save_name}_{i}.pth')
pickle.dump(self.loss,
open(f'{self.save_name}_loss_{i}.pkl', 'wb'))
print(f'Training iteration {i}...', )
self.train_q(squeeze=True)
end_time = time.time()
print(f'Training completed. Took {start_time - end_time} seconds')
torch.save(self._Q.state_dict(), f'{self.save_name}_{iterations}.pth')
pickle.dump(self.loss,
open(f'{self.save_name}_loss_{iterations}.pkl', 'wb'))
def __init__(self, save_name='./data/', load_name=None):
super(BaseRLAgent, self).__init__()
self.training = False
self.max_frames = 10000000
self._epsilon = Epsilon(start=0.9, end=0.1, update_increment=0.0001)
self.gamma = 0.99
self.train_q_per_step = 4
self.train_q_batch_size = 256
self.steps_before_training = 5000
self.target_q_update_frequency = 10000
self.save_name = save_name
if load_name is None:
self.load_name = self.save_name
else:
self.load_name = load_name
self._Q = None
self._Qt = None
self._optimizer = None
self.device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
self._criterion = nn.MSELoss()
self._memory = ReplayMemory(100000)
self._loss = deque(maxlen=int(1e5))
self._max_q = deque(maxlen=int(1e5))
self.loss = []
self.max_q = []
self.reward = []
self._action = None
self._screen = None
self._screen_size = 32
self.n_episodes = 0
self.features = None
def __init__(self, mapname):
super(BattleAgentScripted, self).__init__()
self.max_frames = int(1e5)
self._memory = ReplayMemory(self.max_frames)
self.obs = None
self.features = [_PLAYER_RELATIVE, _UNIT_TYPE, _UNIT_HIT_POINTS]
self.mapname = mapname
self.screen_size = 32
def __init__(self, mapname):
super(BattleAgentScriptedBeacon, self).__init__()
self.max_frames = int(1e5)
self._memory = ReplayMemory(self.max_frames)
self.obs = None
self.features = 5
self.mapname = mapname
self.screen_size = 32