本文整理汇总了Python中agent.Agent.learn方法的典型用法代码示例。如果您正苦于以下问题：Python Agent.learn方法的具体用法？Python Agent.learn怎么用？Python Agent.learn使用的例子？那么恭喜您, 这里精选的方法代码示例或许可以为您提供帮助。您也可以进一步了解该方法所在类agent.Agent的用法示例。
在下文中一共展示了Agent.learn方法的3个代码示例，这些例子默认根据受欢迎程度排序。您可以为喜欢或者感觉有用的代码点赞，您的评价将有助于系统推荐出更棒的Python代码示例。
示例1: __init__
# 需要导入模块: from agent import Agent [as 别名]
# 或者: from agent.Agent import learn [as 别名]
class Environment:
def __init__(self, config):
config.model_type = Config.MODEL_TYPE_CONV2D
print(config.device)
self.config = config
self.env = wrap_deepmind(make_atari(config.env), frame_stack=True)
self.num_states = self.env.observation_space.shape[-1]
self.num_actions = self.env.action_space.n
self.agent = Agent(config, self.num_states, self.num_actions, self.config.num_atoms)
self.total_step = np.zeros(100)
self.data_path = config.data_path
if self.data_path != Config.DATA_PATH_DEFAULT:
self.agent.load_model()
def prepro(self, observation):
ret = np.zeros((4, 84, 84))
ret[0] = observation[:, :, 0]
ret[1] = observation[:, :, 1]
ret[2] = observation[:, :, 2]
ret[3] = observation[:, :, 3]
return ret
def run_episode(self, episode, steps_accumulated=0):
start_time = time.time()
total_reward = 0
observation = self.prepro(self.env.reset())
state = torch.from_numpy(observation).to(self.config.device, dtype=torch.uint8).unsqueeze(0)
for step in range(self.config.num_steps):
if self.config.is_render:
time.sleep(0.064)
self.env.render()
action = self.agent.get_action(state, step + steps_accumulated)
observation_next, reward, done, _ = self.env.step(action.item())
if done:
state_next = None
self.total_step = np.hstack((self.total_step[1:], step + 1))
else:
state_next = self.prepro(observation_next)
state_next = torch.from_numpy(state_next).to(self.config.device, dtype=torch.uint8).unsqueeze(0)
total_reward += reward
reward = torch.tensor([reward], dtype=torch.uint8, device=self.config.device)
if not self.config.is_render:
self.agent.observe(state, action, state_next, reward)
if step % self.config.replay_interval == 0:
self.agent.learn(episode)
state = state_next
if done:
elapsed_time = round(time.time() - start_time, 3)
print('episode: {0}, steps: {1}, mean steps {2}, time: {3}, reward: {4}'.format(episode, step, self.total_step.mean(), elapsed_time, total_reward))
return step + 1
return self.config.num_steps
def run(self):
if not self.config.is_render:
steps = 0
while True:
steps += self.run_episode(-1)
if self.config.steps_learning_start <= steps:
break
steps = 0
for episode in range(self.config.num_episodes):
steps += self.run_episode(episode, steps)
self.env.close()
if self.config.is_saved:
self.agent.save_model()示例2: __init__
# 需要导入模块: from agent import Agent [as 别名]
# 或者: from agent.Agent import learn [as 别名]
class Environment:
def __init__(self, config):
print(config.device)
self.config = config
self.env = gym.make(ENV)
# self.env = wrappers.Monitor(self.env, '/tmp/gym/cartpole_dqn', force=True)
self.num_states = self.env.observation_space.shape[0]
self.num_actions = self.env.action_space.n
self.agent = Agent(config, self.num_states, self.num_actions, config.num_atoms)
self.total_step = np.zeros(100)
def is_success_episode(self, step):
return NUM_STEPS_TO_SUCCEED <= step
def run_episode(self, episode, steps_accumulated=0):
start_time = time.time()
observation = self.env.reset()
state = torch.from_numpy(observation).to(self.config.device, dtype=torch.float32).unsqueeze(0)
for step in range(MAX_STEPS):
action = self.agent.get_action(state, step + steps_accumulated)
observation_next, _, done, _ = self.env.step(action.item())
if done:
state_next = None
self.total_step = np.hstack((self.total_step[1:], step + 1))
if self.is_success_episode(step):
reward = torch.tensor([1.0], dtype=torch.float32, device=self.config.device)
else:
reward = torch.tensor([-1.0], dtype=torch.float32, device=self.config.device)
else:
reward = torch.tensor([0.0], dtype=torch.float32, device=self.config.device)
state_next = torch.from_numpy(observation_next).to(self.config.device, dtype=torch.float32).unsqueeze(0)
self.agent.observe(state, action, state_next, reward)
if step % self.config.replay_interval == 0:
self.agent.learn(episode)
state = state_next
if done:
elapsed_time = round(time.time() - start_time, 3)
print('episode: {0}, steps: {1}, mean steps {2}, time: {3}'.format(episode, step, self.total_step.mean(), elapsed_time))
return step + 1
return MAX_STEPS
def run(self):
steps = 0
while True:
steps += self.run_episode(-1)
if self.config.steps_learning_start < steps:
break
steps = 0
for episode in range(self.config.num_episodes):
if MEAN_STEPS_TO_SUCCEED <= self.total_step.mean():
print('over {0} steps of average last 100 episodes, last episode: {1}, steps: {2}'.format(MEAN_STEPS_TO_SUCCEED, episode, steps))
break
steps += self.run_episode(episode, steps)
self.env.close()示例3: Model
# 需要导入模块: from agent import Agent [as 别名]
# 或者: from agent.Agent import learn [as 别名]
model = Model(lr=1e-4, load=None)
env = Environment(env_type=Environment.TYPE_PONG, render=False)
agent = Agent(env=env, model=model, epsilon_decay=1e-4, min_epsilon=.1)
episode = 0
scores = []
losses = []
qs = []
eps = []
while True:
episode += 1
print "Episode #%d"%(episode)
warmup = 10 if episode == 1 else 0
score, loss, mean_q, epsilon, predictions = agent.learn(overfit=False, games=1, warmup=warmup)
print "Loss %f, mean q %f"%(loss, mean_q)
print "Predictions 0: %d%%, 1: %d%%"%(predictions[0]/sum(predictions)*100, predictions[1]/sum(predictions)*100)
model.save('model.h5')
plt.close()
scores.append(score)
losses.append(loss)
qs.append(mean_q)
eps.append(epsilon)
s_score = plt.subplot(411)
s_loss = plt.subplot(412)