def make_env():
run_name = "yt1"
saves_path = AppConfig.SAVES_DIR / f"simple-{run_name}"
saves_path.mkdir(parents=True, exist_ok=True)
data_path = pathlib.Path(AppConfig.STOCKS)
val_path = pathlib.Path(AppConfig.VAL_STOCKS)
year = 2016
if year is not None or data_path.is_file():
if year is not None:
print('load stock data...')
stock_data = BarData.load_year_data(year)
else:
stock_data = {"YNDX": BarData.load_relative(data_path)}
env = MinuteBarEnv(stock_data,
bars_count=AppConfig.BARS_COUNT,
volumes=True)
env_tst = MinuteBarEnv(stock_data,
bars_count=AppConfig.BARS_COUNT,
volumes=True)
elif data_path.is_dir():
env = MinuteBarEnv.from_dir(data_path,
bars_count=AppConfig.BARS_COUNT)
env_tst = MinuteBarEnv.from_dir(data_path,
bars_count=AppConfig.BARS_COUNT)
else:
raise RuntimeError("No data to train on")
env = gym.wrappers.TimeLimit(env, max_episode_steps=1000)
val_data = {"YNDX": BarData.load_relative(val_path)}
env_val = MinuteBarEnv(val_data,
bars_count=AppConfig.BARS_COUNT,
volumes=True)
return env, env_val, env_tst
def test_State_main(self):
print('生成环境状态类')
year = 2016
instrument = 'data\\YNDX_160101_161231.csv'
stock_data = BarData.load_year_data(year)
print('stock_data: {0};'.format(stock_data[instrument]))
st = State(bars_count=10,
commission_perc=0.1,
reset_on_close=True,
reward_on_close=True,
volumes=True)
st.reset(stock_data[instrument], offset=AppConfig.BARS_COUNT + 1)
obs = st.encode()
print('initial observation: type:{0}; shape:{1};'.format(
type(obs), obs))
# 购买股票
action = AssetActions.Buy
reward, done = st.step(action=action)
obs = st.encode()
info = {'instrument': 'YNDX', 'offset': st._offset}
self._print_State_step_result(reward, done, obs, info)
# 持有
action = AssetActions.Skip
reward, done = st.step(action=action)
obs = st.encode()
info = {'instrument': 'YNDX', 'offset': st._offset}
self._print_State_step_result(reward, done, obs, info)
# 卖出
action = AssetActions.Sell
reward, done = st.step(action=action)
obs = st.encode()
info = {'instrument': 'YNDX', 'offset': st._offset}
self._print_State_step_result(reward, done, obs, info)
print('^_^')
self.assertTrue(1 > 0)
def test_MinuteBarEnv_main(self):
'''
研究市场环境类
'''
year = 2016
instrument = 'data\\YNDX_160101_161231.csv'
stock_data = BarData.load_year_data(year)
print('stock_data: {0};'.format(stock_data[instrument]))
env = MinuteBarEnv(stock_data,
bars_count=AppConfig.BARS_COUNT,
volumes=True)
obs = env.reset()
seq = 1
while True:
if seq > 3:
done = True
if 1 == seq:
action = AssetActions.Buy
elif 2 == seq:
action = AssetActions.Sell
else:
action = AssetActions.Skip
obs, reward, done, info = env.step(action)
if done:
break
env.render(mode='human', obs=obs, reward=reward, info=info)
seq += 1
print('observation: {0};'.format(obs))
def test_exp(self):
#
device = torch.device("cuda:0")
year = 2016
stock_data = BarData.load_year_data(year)
env = MinuteBarEnv(
stock_data, bars_count=AppConfig.BARS_COUNT)
env = gym.wrappers.TimeLimit(env, max_episode_steps=1000)
net = SimpleFFDQN(env.observation_space.shape[0],
env.action_space.n).to(device)
selector = rll.actions.EpsilonGreedyActionSelector(AppConfig.EPS_START)
agt = DQNAgent(net, selector, device=device)
obs = env.reset()
'''
def test_exp(self):
#
device = torch.device("cuda:0")
year = 2016
stock_data = BarData.load_year_data(year)
env = MinuteBarEnv(stock_data,
bars_count=AppConfig.BARS_COUNT,
volumes=True)
env = gym.wrappers.TimeLimit(env, max_episode_steps=1000)
net = SimpleFFDQN(env.observation_space.shape[0],
env.action_space.n).to(device)
selector = rll.actions.EpsilonGreedyActionSelector(AppConfig.EPS_START)
agent = DQNAgent(net, selector, device=device)
exp_source = rll.experience.ExperienceSourceFirstLast(
env, agent, AppConfig.GAMMA, steps_count=AppConfig.REWARD_STEPS)
src_itr = iter(exp_source)
v1 = next(src_itr)
def test_exp(self):
#
device = torch.device("cuda:0")
year = 2016
stock_data = BarData.load_year_data(year)
env = MinuteBarEnv(
stock_data, bars_count=AppConfig.BARS_COUNT, volumes=True)
env = gym.wrappers.TimeLimit(env, max_episode_steps=1000)
net = SimpleFFDQN(env.observation_space.shape[0],
env.action_space.n).to(device)
selector = rll.actions.EpsilonGreedyActionSelector(AppConfig.EPS_START)
agent = DQNAgent(net, selector, device=device)
exp_source = rll.experience.ExperienceSourceFirstLast(
env, agent, AppConfig.GAMMA, steps_count=AppConfig.REWARD_STEPS)
replay_buffer = ExperienceReplayBuffer(
exp_source, AppConfig.REPLAY_SIZE)
replay_buffer.populate(1000)
batch_size = 16
X = replay_buffer.sample(batch_size)
print('X: {0}; {1};'.format(type(X), X))
def from_dir(cls, data_dir, **kwargs):
prices = {
file: BarData.load_relative(file)
for file in BarData.price_files(data_dir)
}
return MinuteBarEnv(prices, **kwargs)