def test(self, model_epoch: int = 0, should_render: bool = True):
train_provider, test_provider = self.data_provider.split_data_train_test(self.train_split_percentage)
del train_provider
test_env = SubprocVecEnv([make_env(test_provider, i) for i in range(self.n_envs)])
model_path = path.join('data', 'agents', f'{self.study_name}__{model_epoch}.pkl')
model = self.Model.load(model_path, env=test_env)
self.logger.info(f'Testing model ({self.study_name}__{model_epoch})')
state = None
obs, done, rewards = test_env.reset(), [False], []
while not all(done):
action, state = model.predict(obs, state=state)
obs, reward, done, _ = test_env.step(action)
rewards.append(reward)
if should_render and self.n_envs == 1:
test_env.render(mode='human')
self.logger.info(
f'Finished testing model ({self.study_name}__{model_epoch}): ${"{:.2f}".format(np.sum(rewards))}')
def main():
#env = SubprocVecEnv([(lambda i=i: SwocGym(i+1, GameServicePath, i, fieldWidth=10, fieldHeight=10, saveEpisode=True)) for i in range(1)])
env = SubprocVecEnv([
(lambda i=i: MazeGym(mazeWidth=10, mazeHeight=10, nrWallsToRemove=10))
for i in range(1)
])
try:
model = PPO2("MlpPolicy",
env,
verbose=1,
tensorboard_log='/home/ralph/swoc2019/log')
if SaveFile.exists():
print('loading...')
model.load_parameters(SaveFile)
else:
print('Warning: No save file loaded')
print('evaluating...', end='')
obs = env.reset()
totalRewards = None
for i in range(100):
action, _states = model.predict(obs)
obs, rewards, dones, info = env.step(action)
totalRewards = totalRewards + rewards if totalRewards is not None else rewards
env.render()
sleep(0.2)
print(f'mean reward: {np.mean(totalRewards)}')
except KeyboardInterrupt:
print('closing...')
finally:
env.close()
print('closed')
def attention_render(model_name, env_name, num_cpu, log_dir):
if not os.path.exists(log_dir):
raise ('log_dir not Exists')
env_id = env_name + 'NoFrameskip-v4'
env = SubprocVecEnv([make_env(env_id, i, log_dir) for i in range(num_cpu)])
# env = Monitor(env, log_dir, allow_early_resets=True)
if model_name == 'A2C_Attention':
model = A2C(AttentionPolicy,
env,
verbose=1,
tensorboard_log=log_dir + 'tensorboard/')
elif model_name == 'A2C_Attention2':
model = A2C(Attention2Policy,
env,
verbose=1,
tensorboard_log=log_dir + 'tensorboard/')
elif model_name == 'A2C':
model = A2C(LstmPolicy,
env,
verbose=1,
tensorboard_log=log_dir + 'tensorboard/')
else:
model = None
model = model.load(log_dir + model_name + '_' + env_name, env=env)
obs = env.reset()
# print(env.observation_space)
# cv2.imshow('test', RGB2BGR(obs[0]))
# cv2.waitKey(0)
while True:
action, _states = model.predict(obs)
obs, rewards, done, info = env.step(action)
attentions = model.get_attention(obs, _states, done)[0]
attentions_img = []
# print('attention', np.array(attention).shape)
for i, attention in enumerate(attentions):
attention = np.array(attention)
attention = np.reshape(attention, [
env.observation_space.shape[0] // 10,
env.observation_space.shape[1] // 10, 1
])
attention = np.repeat(attention, [10] * attention.shape[0], axis=0)
attention = np.repeat(attention, [10] * attention.shape[1], axis=1)
attention = attention * 255
attentions_img.append(attention)
# print(np.sum(attention))
attentions = tile_images(attentions_img)
cv2.imshow('attention', attentions)
cv2.waitKey(1)
# break
env.render()
return model
def test():
# Parallel environments
n_cpu = 4
env = SubprocVecEnv([lambda: RSEnv() for i in range(n_cpu)])
model = A2C(MlpPolicy, env, verbose=1)
model.learn(total_timesteps=600000, log_interval=10)
model.save("sba2c")
env = TestRSEnv()
obs = env.reset()
done = False
while not done:
action, _ = model.predict(obs)
obs, rewards, done, info = env.step(action)
env.render()
env.close()
def main():
agent_data = pd.read_csv('../output_EURUSD_M1_/agentData.csv')
agent_data = agent_data.drop(agent_data.columns[0], axis=1)
agent_data = agent_data.astype('float32')
env = SubprocVecEnv([lambda: ForexTradingEnv(agent_data)] * 10, )
#env = DummyVecEnv([lambda: ForexTradingEnv(agent_data)], )
# model = DQN(CustomDQNPolicy, env, gamma=0.95, verbose=1, tensorboard_log = "./tensorboard", entcoeff=0.005, adam_epsilon = 1e-6)
import tensorflow as tf
from TenorboardCallbacks import TensorboardCallback
checkpoint_callback = CheckpointCallback(save_freq=1000000,
save_path='./models/',
name_prefix='ppo2')
for curr in [1]:
model = PPO2(PPO2Policy_Basic,
env,
verbose=1,
tensorboard_log="./tensorboard",
vf_coef=1e-7,
ent_coef=1e-4,
n_steps=512,
gamma=0.99)
#model = PPO2.load("5_days_model/ppo2_999000000_steps.zip", policy=PPO2Policy_Basic, env = env,verbose=1, tensorboard_log = "./tensorboard")
model.learn(total_timesteps=10000000000,
log_interval=10000000,
callback=CallbackList(
[TensorboardCallback(env), checkpoint_callback]))
model.save(model_fileName)
obs = env.reset()
for i in range(2000000):
action, _states = model.predict(obs)
obs, rewards, done, info = env.step(action)
if i % 1 == 0:
env.render()
if done:
break
def run_baseline_ppo2(env_name, n_cpu=4, train=True):
from stable_baselines.common.policies import MlpPolicy
from stable_baselines.common.vec_env import SubprocVecEnv
from stable_baselines import PPO2
if train:
# multiprocess environment
env = SubprocVecEnv([lambda: gym.make(env_name) for i in range(n_cpu)])
model = PPO2(MlpPolicy, env, verbose=1)
model.learn(total_timesteps=100000)
model.save("checkpoints/ppo2_" + env_name)
else:
from stable_baselines.common.vec_env import DummyVecEnv
env = DummyVecEnv([lambda: gym.make(env_name)])
model = PPO2.load("checkpoints/ppo2_" + env_name)
obs = env.reset()
while True:
action, _states = model.predict(obs)
obs, rewards, dones, info = env.step(action)
env.render()
print("state: ", obs, " reward: ", rewards, " done: ", dones, "info: ", info)
class BitmexTradingStrategySBL(TradingStrategy):
"""A trading strategy capable of self tuning, training, and evaluating with stable-baselines.
Arguments:
environments: An instance of a trading environments for the agent to trade within.
model: The RL model to create the agent with.
Defaults to DQN.
policy: The RL policy to train the agent's model with.
Defaults to 'MlpPolicy'.
model_kwargs: Any additional keyword arguments to adjust the model.
kwargs: Optional keyword arguments to adjust the strategy.
"""
def __init__(self,
environment: BitmexEnvironment,
model: BaseRLModel = DQN,
policy: Union[str, BasePolicy] = 'MlpPolicy',
model_kwargs: any = {},
policy_kwargs: any = {},
n_env: int = 1,
**kwargs):
self._model = model
self._model_kwargs = model_kwargs
self._policy_kwargs = policy_kwargs
self._n_env = n_env
self.environment = environment
self._agent = self._model(policy,
self._environment,
**self._model_kwargs,
policy_kwargs=self._policy_kwargs)
@property
def environment(self) -> 'BitmexEnvironment':
"""A `BitmexEnvironment` instance for the agent to trade within."""
return self._environment
@environment.setter
def environment(self, environment: 'BitmexEnvironment'):
envs = [lambda: environment for _ in range(self._n_env)]
if self._n_env == 1:
self._environment = DummyVecEnv(envs)
else:
self._environment = SubprocVecEnv(envs)
def restore_agent(self, path: str, custom_objects: any = {}):
"""Deserialize the strategy's learning agent from a file.
Arguments:
path: The `str` path of the file the agent specification is stored in.
"""
self._custom_objects = custom_objects
self._agent = self._model.load(path,
env=self._environment,
custom_objects=self._custom_objects,
kwargs=self._model_kwargs)
def save_agent(self, path: str):
"""Serialize the learning agent to a file for restoring later.
Arguments:
path: The `str` path of the file to store the agent specification in.
"""
self._agent.save(path)
def tune(self,
steps: int = None,
episodes: int = None,
callback: Callable[[pd.DataFrame], bool] = None) -> pd.DataFrame:
raise NotImplementedError
def _train_callback(self, _locals, _globals):
# performance = self._environment.performance
#
# if self._episode_callback and self._environment.done():
# self._episode_callback(performance)
return True
def train(
self,
steps: int = None,
episodes: int = None,
render_mode: str = None,
episode_callback: Callable[[pd.DataFrame],
bool] = None) -> pd.DataFrame:
if steps is None:
raise ValueError(
'You must set the number of `steps` to train the strategy.')
self._agent.learn(steps, callback=self._train_callback)
return True
def test(
self,
steps: int = None,
episodes=None,
render_mode: str = None,
episode_callback: Callable[[pd.DataFrame],
bool] = None) -> pd.DataFrame:
if steps is None and episodes is None:
raise ValueError(
'You must set the number of `steps` or `episodes` to test the strategy.'
)
steps_completed, episodes_completed, average_reward = 0, 0, 0
obs, state, dones = self._environment.reset(), None, [False]
performance = {}
while (steps is not None and
(steps == 0 or steps_completed < steps)) or (
episodes is not None and episodes_completed < episodes):
actions, state = self._agent.predict(obs, state=state, mask=dones)
# actions, state = self._agent.predict(obs)
obs, rewards, dones, info = self._environment.step(actions)
steps_completed += 1
average_reward -= average_reward / steps_completed
average_reward += rewards[0] / (steps_completed + 1)
exchange_performance = info[0].get('exchange').performance
performance = exchange_performance if len(
exchange_performance) > 0 else performance
if render_mode is not None:
self._environment.render(mode=render_mode)
if dones[0]:
if episode_callback is not None and not episode_callback(
performance):
break
episodes_completed += 1
obs = self._environment.reset()
print("Finished running strategy.")
print("Total episodes: {} ({} timesteps).".format(
episodes_completed, steps_completed))
print("Average reward: {}.".format(average_reward))
return performance
policy = ''
model_tag = ''
if len(sys.argv) > 1:
policy = sys.argv[1]
model_tag = '_' + sys.argv[1]
if __name__ == '__main__':
env = SubprocVecEnv([lambda: BaseEnv() for i in range(4)])
env = VecFrameStack(env, 3)
model = PPO2(get_policy(policy),
env,
verbose=0,
nminibatches=1,
tensorboard_log=tensorboard_folder)
model.learn(total_timesteps=100000000, tb_log_name='PPO2' + model_tag)
model.save(model_folder + "PPO2" + model_tag)
del model
model = PPO2.load(model_folder + "PPO2" + model_tag)
done = False
states = None
obs = env.reset()
while not done:
action, states = model.predict(obs, states)
obs, _, done, info = env.step(action)
env.render()
def _init():
env = gym.make(env_id)
env.seed(seed + rank)
return env
set_global_seeds(seed)
return _init
if __name__ == '__main__':
env_id = "CartPole-v1"
num_cpu = 4 # Number of processes to use
# Create the vectorized environment
env = SubprocVecEnv([make_env(env_id, i) for i in range(num_cpu)])
# Stable Baselines provides you with make_vec_env() helper
# which does exactly the previous steps for you:
# env = make_vec_env(env_id, n_envs=num_cpu, seed=0)
model = ACKTR(MlpPolicy,
env,
verbose=1,
tensorboard_log="./a2c_cartpole_tensorboard/")
model.learn(total_timesteps=250000)
obs = env.reset()
for _ in range(1000):
action, _states = model.predict(obs)
obs, rewards, dones, info = env.step(action)
env.render(mode='rgb_array')
def main(args):
log_dir = args.log_path if (args.log_path is not None) else \
"/tmp/stable_baselines_" + time.strftime('%Y-%m-%d-%H-%M-%S')
configure_logger(log_dir)
set_global_seeds(args.seed)
n_cpu = get_num_workers(args.env) if not args.play else 1
env_kwargs = get_env_kwargs(args.env, args.random_ratio, args.sequential,
args.reward_type, args.n_object,
args.curriculum)
def make_thunk(rank):
return lambda: make_env(env_id=args.env,
rank=rank,
log_dir=log_dir,
flatten_dict=True,
kwargs=env_kwargs)
env = SubprocVecEnv([make_thunk(i) for i in range(n_cpu)])
eval_env_kwargs = env_kwargs.copy()
eval_env_kwargs['random_ratio'] = 0.0
if "use_cu" in eval_env_kwargs:
eval_env_kwargs['use_cu'] = False
eval_env = make_env(env_id=args.env,
rank=0,
flatten_dict=True,
kwargs=eval_env_kwargs)
print(eval_env)
if not args.play:
os.makedirs(log_dir, exist_ok=True)
train_kwargs = get_train_kwargs("ppo",
args,
parsed_action_noise=None,
eval_env=eval_env)
# policy = 'MlpPolicy'
from utils.attention_policy import AttentionPolicy
register_policy('AttentionPolicy', AttentionPolicy)
policy_kwargs = get_policy_kwargs("ppo", args)
print(policy_kwargs)
model = PPO2(args.policy,
env,
verbose=1,
nminibatches=32,
lam=0.95,
noptepochs=10,
ent_coef=0.01,
learning_rate=3e-4,
cliprange=0.2,
policy_kwargs=policy_kwargs,
**train_kwargs)
print(model.get_parameter_list())
def callback(_locals, _globals):
num_update = _locals["update"]
if 'FetchStack' in args.env:
mean_eval_reward = stack_eval_model(eval_env, _locals["self"])
else:
mean_eval_reward = eval_model(eval_env, _locals["self"])
log_eval(num_update, mean_eval_reward)
if num_update % 10 == 0:
model_path = os.path.join(log_dir,
'model_' + str(num_update // 10))
model.save(model_path)
print('model saved to', model_path)
return True
model.learn(total_timesteps=int(args.num_timesteps),
callback=callback,
seed=args.seed,
log_interval=1)
model.save(os.path.join(log_dir, 'final'))
else:
assert args.load_path is not None
model = PPO2.load(args.load_path)
fig, ax = plt.subplots(1, 1, figsize=(8, 8))
obs = env.reset()
goal_dim = env.get_attr('goal')[0].shape[0]
if 'FetchStack' in args.env:
while env.get_attr('current_nobject')[0] != env.get_attr('n_object')[0] or \
env.get_attr('task_mode')[0] != 1:
obs = env.reset()
elif 'FetchPush' in args.env:
while not (1.25 < obs[0][6] < 1.33 and obs[0][7] < 0.61
and 0.7 < obs[0][4] < 0.8):
obs = env.reset()
env.env_method('set_goal', np.array([1.2, 0.75, 0.425, 1, 0]))
obs = env.env_method('get_obs')
obs[0] = np.concatenate([
obs[0][key]
for key in ['observation', 'achieved_goal', 'desired_goal']
])
else:
while np.argmax(obs[0][-goal_dim + 3:]) != 0:
obs = env.reset()
print('achieved_goal', obs[0][-2 * goal_dim:-goal_dim], 'goal',
obs[0][-goal_dim:])
episode_reward = 0.0
num_episode = 0
frame_idx = 0
images = []
if 'max_episode_steps' not in env_kwargs.keys():
env_kwargs['max_episode_steps'] = 100
for i in range(env_kwargs['max_episode_steps'] * 10):
img = env.render(mode='rgb_array')
ax.cla()
ax.imshow(img)
if env.get_attr('goal')[0].shape[0] <= 3:
ax.set_title('episode ' + str(num_episode) + ', frame ' +
str(frame_idx))
else:
ax.set_title('episode ' + str(num_episode) + ', frame ' +
str(frame_idx) + ', goal idx ' +
str(np.argmax(env.get_attr('goal')[0][3:])))
if 'FetchStack' in args.env:
tasks = ['pick and place', 'stack']
ax.set_title('episode ' + str(num_episode) + ', frame ' +
str(frame_idx) + ', task: ' +
tasks[np.argmax(obs[0][-2 * goal_dim - 2:-2 *
goal_dim])])
images.append(img)
action, _ = model.predict(obs)
obs, reward, done, _ = env.step(action)
episode_reward += reward
frame_idx += 1
if not args.export_video:
plt.pause(0.1)
else:
plt.imsave(
os.path.join(os.path.dirname(args.load_path),
'tempimg%d.png' % i), img)
if done:
print('episode_reward', episode_reward)
if 'FetchStack' in args.env:
while env.get_attr('current_nobject')[0] != env.get_attr('n_object')[0] or \
env.get_attr('task_mode')[0] != 1:
obs = env.reset()
else:
while np.argmax(obs[0][-goal_dim + 3:]) != 0:
obs = env.reset()
print('goal', obs[0][-goal_dim:])
episode_reward = 0.0
frame_idx = 0
num_episode += 1
if num_episode >= 10:
break
if args.export_video:
os.system('ffmpeg -r 5 -start_number 0 -i ' +
os.path.dirname(args.load_path) +
'/tempimg%d.png -c:v libx264 -pix_fmt yuv420p ' +
os.path.join(os.path.dirname(args.load_path), args.env +
'.mp4'))
for i in range(env_kwargs['max_episode_steps'] * 10):
try:
os.remove(
os.path.join(os.path.dirname(args.load_path),
'tempimg' + str(i) + '.png'))
except:
pass
if __name__ == '__main__':
env_id = "TaxiDummy-v01"
DATA_PATH = os.path.join(os.environ['ALLDATA_PATH'], "macaoFiles",
"taxi_env_dummy")
if os.path.isdir(DATA_PATH):
shutil.rmtree(DATA_PATH)
os.makedirs(DATA_PATH)
num_cpu = 4 # Number of processes to use
# Create the vectorized environment
env = SubprocVecEnv([make_env(env_id, i) for i in range(num_cpu)])
model = A2C(MlpPolicy, env, verbose=1)
model.learn(total_timesteps=100000)
obs = env.reset()
images = []
img = env.render(mode="rgb_array")
images.append(img)
for _ in range(70):
action, _states = model.predict(obs)
obs, rewards, dones, info = env.step(action)
images.append(env.render(mode="rgb_array"))
imageio.mimwrite(os.path.join(DATA_PATH, 'taxi_dummy_a2c.gif'),
[np.array(img) for i, img in enumerate(images)],
format="GIF-PIL",
fps=5)
if __name__ == "__main__":
# Parallel environments
if parallel:
env = SubprocVecEnv([make_env(env_id, i) for i in range(num_cpu)])
else:
env = DummyVecEnv([make_env(env_id, i) for i in range(num_cpu)])
model = PPO2(CnnPolicy, env, verbose=1, tensorboard_log=log_dir)
model.learn(total_timesteps=max_steps)
model.save(log_dir + "ppo_minigrid")
del model # remove to demonstrate saving and loading
model = PPO2.load(log_dir + "ppo_minigrid")
env = make_env(env_id, 0)()
mean_reward, std_reward = evaluate_policy(model, env)
print("Mean Reward: {}, std_dev: {}".format(mean_reward, std_reward))
demo = input("Watch model? (q to quit)")
if demo != "q":
for _ in range(1000):
obs = env.reset()
t = 0
while t < 200:
action, _states = model.predict(obs)
obs, rewards, done, info = env.step(action)
print("Action: {}\tTimestep: {}".format(action, t))
env.render(mode='human')
t += 1
if done:
break
import gym
from stable_baselines.common.policies import MlpPolicy
from stable_baselines.common.vec_env import SubprocVecEnv
from stable_baselines import A2C
n_cpu = 4 # 支持4线程
env = SubprocVecEnv([lambda: gym.make('CartPole-v1') for i in range(n_cpu)])
model = A2C(MlpPolicy, env, verbose=1) # 使用MlpPolicy的A2C算法
model.learn(total_timesteps=25000) # 训练
obs = env.reset()
while True:
action, _states = model.predict(obs) # 预测
obs, rewards, dones, info = env.step(action) # 执行一步游戏
env.render() # 显示
