loss = self.loss_func(target_v, Qsa)
losses += loss
self.optimizer.zero_grad()
losses.backward()
self.optimizer.step()
self.writer.add_scalar('loss/value_loss', losses / self.batch_size, self.update_count)
self.update_count += 1
if self.update_count % 500 == 0:
self.target_net.load_state_dict(self.act_net.state_dict())
torch.save(self.act_net.state_dict(), config.act_net_model_dir + str(self.update_count) + ".model")
from BiYeSheJi.Module.environment import Environment
env = Environment(Graph)
env.init_TreeList(config.dataSet + "/train.txt")
def main():
agentP = DQN()
find_target = 0
for i_ep in range(num_episodes):
root = env.reset()
if render: env.render()
# node_t 是经历一次蒙特卡洛树搜索后根据UCT选择的最好下一节点
path = [root.state.state_tup, ]
node_t = root
reward = 0
while node_t.state.current_node != root.state.target_node:
node_t = Policy_MCTS(node_t, agentP.act_net)
if node_t is not None:
def main():
from multiprocessing import Pool
from multiprocessing import Manager
pool = Pool(5)
isright = 0
precision_queue = Manager().Queue()
for i_ep, root in enumerate(env.TreeList):
pool.apply_async(func=sub_main, args=(root, precision_queue))
pool.close()
pool.join()
ql = precision_queue.qsize()
for i in range(ql):
isright += precision_queue.get_nowait()
print(isright/precision_queue.qsize())
# agentP.writer.add_scalar('find_target/step', find_target, global_step=i_ep)
agentP = DQN()
Policy_net = agentP.policy_net
if __name__ == '__main__':
from BiYeSheJi.Module.environment import Environment
env = Environment(Graph)
env.init_TreeList(config.dataSet + "/test.txt")
main()