class Play_Sekiro_Online:
def __init__(self,
save_memory_path=None,
load_memory_path=None,
save_weights_path=None,
load_weights_path=None):
self.save_memory_path = save_memory_path # 指定记忆/经验保存的路径。默认为None，不保存。
self.load_memory_path = load_memory_path # 指定记忆/经验加载的路径。默认为None，不加载。
self.sekiro_agent = Sekiro_Agent(
save_weights_path=save_weights_path, # 指定模型权重保存的路径。默认为None，不保存。
load_weights_path=load_weights_path # 指定模型权重加载的路径。默认为None，不加载。
)
if not save_weights_path: # 注：默认也是测试模式，若设置该参数，就会开启训练模式
self.train = False
self.sekiro_agent.step = self.sekiro_agent.replay_start_size + 1
else:
self.train = True
self.reward_system = RewardSystem()
self.i = 0 # 计步器
self.screens = deque(maxlen=in_depth * 2) # 用双端队列存放图像
if self.load_memory_path:
self.load_memory() # 加载记忆/经验
def load_memory(self):
if os.path.exists(self.load_memory_path):
last_time = time.time()
self.sekiro_agent.replayer.memory = pd.read_json(
self.load_memory_path) # 从json文件加载记忆/经验。
print(
f'Load {self.load_memory_path}. Took {round(time.time()-last_time, 3):>5} seconds.'
)
i = self.sekiro_agent.replayer.memory.action.count()
self.sekiro_agent.replayer.i = i
self.sekiro_agent.replayer.count = i
self.sekiro_agent.step = i
else:
print('No memory to load.')
def get_S(self):
for _ in range(in_depth):
self.screens.append(get_screen()) # 先进先出，右进左出
def img_processing(self, screens):
return np.array([
cv2.resize(
roi(cv2.cvtColor(screen, cv2.COLOR_BGR2GRAY), x, x_w, y, y_h),
(in_height, in_width)) for screen in screens
])
def round(self):
observation = self.img_processing(list(self.screens)[:in_depth]) # S
action = self.action = self.sekiro_agent.choose_action(
observation) # A
self.get_S()
reward = self.reward_system.get_reward(
cur_status=get_status(list(self.screens)[in_depth - 1])[:4],
next_status=get_status(list(self.screens)[in_depth * 2 -
1])[:4]) # R
next_observation = self.img_processing(list(
self.screens)[in_depth:]) # S'
if self.train:
self.sekiro_agent.replayer.store(observation, action, reward,
next_observation)
if self.sekiro_agent.replayer.count >= self.sekiro_agent.replay_start_size:
self.sekiro_agent.learn()
def run(self):
paused = True
print("Ready!")
while True:
last_time = time.time()
keys = key_check()
if paused:
if 'T' in keys:
self.get_S()
paused = False
print('\nStarting!')
else: # 按 'T' 之后，马上下一轮就进入这里
self.i += 1
self.round()
print(
f'\r {self.sekiro_agent.who_play:>4} , step: {self.i:>6} . Loop took {round(time.time()-last_time, 3):>5} seconds. action {self.action:>1} , total_reward: {self.reward_system.total_reward:>10.3f} , memory: {self.sekiro_agent.replayer.count:7>} .',
end='')
if 'P' in keys:
if self.train:
self.sekiro_agent.save_evaluate_network(
) # 学习完毕，保存网络权重
self.sekiro_agent.replayer.memory.to_json(
self.save_memory_path) # 保存经验
self.reward_system.save_reward_curve(
) # 绘制 reward 曲线并保存在当前目录
break
print('\nDone!')
class Play_Sekiro_Online:
def __init__(self,
save_memory_path=None,
load_memory_path=None,
save_weights_path=None,
load_weights_path=None):
self.save_memory_path = save_memory_path # Specify the path to save the memory/experience. The default is None, do not save.
self.load_memory_path = load_memory_path # Specify the path to load the memory/experience. The default is None, do not load.
self.sekiro_agent = Sekiro_Agent(
save_weights_path=
save_weights_path, # Specify the path to save the model weight. The default is None, do not save.
load_weights_path=
load_weights_path # Specify the path to load the model weight. The default is None, do not load.
)
if not save_weights_path: # Note: The default is also the test mode, if you set this parameter, the training mode will be turned on.
self.train = False
self.sekiro_agent.step = self.sekiro_agent.replay_start_size + 1
else:
self.train = True
self.reward_system = RewardSystem()
self.i = 0 # Pedometer
self.screens = deque(maxlen=in_depth * 2) # Use deque to store images.
if self.load_memory_path:
self.load_memory() # load the memory/experience.
def load_memory(self):
if os.path.exists(self.load_memory_path):
last_time = time.time()
self.sekiro_agent.replayer.memory = pd.read_json(
self.load_memory_path
) # load the memory/experience from json file.
print(
f'Load {self.load_memory_path}. Took {round(time.time()-last_time, 3):>5} seconds.'
)
i = self.sekiro_agent.replayer.memory.action.count()
self.sekiro_agent.replayer.i = i
self.sekiro_agent.replayer.count = i
self.sekiro_agent.step = i
else:
print('No memory to load.')
def get_S(self):
for _ in range(in_depth):
self.screens.append(
get_screen()) # First in first out, right in left out.
def img_processing(self, screens):
return np.array([
cv2.resize(
roi(cv2.cvtColor(screen, cv2.COLOR_BGR2GRAY), x, x_w, y, y_h),
(in_height, in_width)) for screen in screens
])
def round(self):
observation = self.img_processing(list(self.screens)[:in_depth]) # S
action = self.action = self.sekiro_agent.choose_action(
observation) # A
self.get_S()
reward = self.reward_system.get_reward(
cur_status=get_status(list(self.screens)[in_depth - 1])[:4],
next_status=get_status(list(self.screens)[in_depth * 2 -
1])[:4]) # R
next_observation = self.img_processing(list(
self.screens)[in_depth:]) # S'
if self.train:
self.sekiro_agent.replayer.store(observation, action, reward,
next_observation)
if self.sekiro_agent.replayer.count >= self.sekiro_agent.replay_start_size:
self.sekiro_agent.learn()
def run(self):
paused = True
print("Ready!")
while True:
last_time = time.time()
keys = key_check()
if paused:
if 'T' in keys:
self.get_S()
paused = False
print('\nStarting!')
else: # After pressing'T', it will enter here immediately in the next round
self.i += 1
self.round()
print(
f'\r {self.sekiro_agent.who_play:>4} , step: {self.i:>6} . Loop took {round(time.time()-last_time, 3):>5} seconds. action {self.action:>1} , total_reward: {self.reward_system.total_reward:>10.3f} , memory: {self.sekiro_agent.replayer.count:7>} .',
end='')
if 'P' in keys:
if self.train:
self.sekiro_agent.save_evaluate_network()
self.sekiro_agent.replayer.memory.to_json(
self.save_memory_path)
self.reward_system.save_reward_curve()
break
print('\nDone!')