fix multiproc callback, add continual learning to multiproc, fix totalprofit bug in env, set eval_freq automatically, improve default reward

This commit is contained in:
robcaulk 2022-08-25 11:46:18 +02:00
parent d1bee29b1e
commit 94cfc8e63f
6 changed files with 58 additions and 62 deletions

View File

@ -56,9 +56,9 @@
"freqai": { "freqai": {
"enabled": true, "enabled": true,
"model_save_type": "stable_baselines", "model_save_type": "stable_baselines",
"conv_width": 10, "conv_width": 4,
"purge_old_models": true, "purge_old_models": true,
"train_period_days": 10, "train_period_days": 5,
"backtest_period_days": 2, "backtest_period_days": 2,
"identifier": "unique-id", "identifier": "unique-id",
"data_kitchen_thread_count": 2, "data_kitchen_thread_count": 2,
@ -72,7 +72,7 @@
"30m" "30m"
], ],
"indicator_max_period_candles": 10, "indicator_max_period_candles": 10,
"indicator_periods_candles": [5, 10] "indicator_periods_candles": [5]
}, },
"data_split_parameters": { "data_split_parameters": {
"test_size": 0.5, "test_size": 0.5,
@ -85,13 +85,13 @@
"verbose": 1 "verbose": 1
}, },
"rl_config": { "rl_config": {
"train_cycles": 3, "train_cycles": 6,
"eval_cycles": 3,
"thread_count": 4, "thread_count": 4,
"max_trade_duration_candles": 100, "max_trade_duration_candles": 300,
"model_type": "PPO", "model_type": "PPO",
"policy_type": "MlpPolicy", "policy_type": "MlpPolicy",
"continual_retraining": true, "continual_learning": false,
"max_training_drawdown_pct": 0.5,
"model_reward_parameters": { "model_reward_parameters": {
"rr": 1, "rr": 1,
"profit_aim": 0.02, "profit_aim": 0.02,

View File

@ -1,3 +1,5 @@
# Example of a 3 action environment.
# import logging # import logging
# from enum import Enum # from enum import Enum

View File

@ -77,8 +77,7 @@ class Base5ActionRLEnv(gym.Env):
self._position = Positions.Neutral self._position = Positions.Neutral
self._position_history: list = [None] self._position_history: list = [None]
self.total_reward: float = 0 self.total_reward: float = 0
self._total_profit: float = 0 self._total_profit: float = 1
self._first_rendering: bool = False
self.history: dict = {} self.history: dict = {}
self.trade_history: list = [] self.trade_history: list = []
@ -101,7 +100,6 @@ class Base5ActionRLEnv(gym.Env):
self.total_reward = 0. self.total_reward = 0.
self._total_profit = 1. # unit self._total_profit = 1. # unit
self._first_rendering = True
self.history = {} self.history = {}
self.trade_history = [] self.trade_history = []
self.portfolio_log_returns = np.zeros(len(self.prices)) self.portfolio_log_returns = np.zeros(len(self.prices))
@ -165,7 +163,7 @@ class Base5ActionRLEnv(gym.Env):
{'price': self.current_price(), 'index': self._current_tick, {'price': self.current_price(), 'index': self._current_tick,
'type': trade_type}) 'type': trade_type})
if self._total_profit < 0.5: if self._total_profit < 1 - self.rl_config.get('max_training_drawdown_pct', 0.8):
self._done = True self._done = True
self._position_history.append(self._position) self._position_history.append(self._position)
@ -293,7 +291,6 @@ class Base5ActionRLEnv(gym.Env):
return 0. return 0.
def _update_profit(self, action): def _update_profit(self, action):
# if self._is_trade(action) or self._done:
if self._is_trade(action) or self._done: if self._is_trade(action) or self._done:
pnl = self.get_unrealized_profit() pnl = self.get_unrealized_profit()

View File

@ -23,8 +23,8 @@ logger = logging.getLogger(__name__)
torch.multiprocessing.set_sharing_strategy('file_system') torch.multiprocessing.set_sharing_strategy('file_system')
SB3_MODELS = ['PPO', 'A2C', 'DQN', 'TD3', 'SAC'] SB3_MODELS = ['PPO', 'A2C', 'DQN']
SB3_CONTRIB_MODELS = ['TRPO', 'ARS'] SB3_CONTRIB_MODELS = ['TRPO', 'ARS', 'RecurrentPPO', 'MaskablePPO']
class BaseReinforcementLearningModel(IFreqaiModel): class BaseReinforcementLearningModel(IFreqaiModel):
@ -41,7 +41,7 @@ class BaseReinforcementLearningModel(IFreqaiModel):
self.eval_callback: EvalCallback = None self.eval_callback: EvalCallback = None
self.model_type = self.freqai_info['rl_config']['model_type'] self.model_type = self.freqai_info['rl_config']['model_type']
self.rl_config = self.freqai_info['rl_config'] self.rl_config = self.freqai_info['rl_config']
self.continual_retraining = self.rl_config.get('continual_retraining', False) self.continual_learning = self.rl_config.get('continual_learning', False)
if self.model_type in SB3_MODELS: if self.model_type in SB3_MODELS:
import_str = 'stable_baselines3' import_str = 'stable_baselines3'
elif self.model_type in SB3_CONTRIB_MODELS: elif self.model_type in SB3_CONTRIB_MODELS:
@ -109,7 +109,6 @@ class BaseReinforcementLearningModel(IFreqaiModel):
""" """
train_df = data_dictionary["train_features"] train_df = data_dictionary["train_features"]
test_df = data_dictionary["test_features"] test_df = data_dictionary["test_features"]
eval_freq = self.freqai_info["rl_config"]["eval_cycles"] * len(test_df)
self.train_env = MyRLEnv(df=train_df, prices=prices_train, window_size=self.CONV_WIDTH, self.train_env = MyRLEnv(df=train_df, prices=prices_train, window_size=self.CONV_WIDTH,
reward_kwargs=self.reward_params, config=self.config) reward_kwargs=self.reward_params, config=self.config)
@ -117,7 +116,7 @@ class BaseReinforcementLearningModel(IFreqaiModel):
window_size=self.CONV_WIDTH, window_size=self.CONV_WIDTH,
reward_kwargs=self.reward_params, config=self.config)) reward_kwargs=self.reward_params, config=self.config))
self.eval_callback = EvalCallback(self.eval_env, deterministic=True, self.eval_callback = EvalCallback(self.eval_env, deterministic=True,
render=False, eval_freq=eval_freq, render=False, eval_freq=len(train_df),
best_model_save_path=str(dk.data_path)) best_model_save_path=str(dk.data_path))
@abstractmethod @abstractmethod
@ -138,6 +137,8 @@ class BaseReinforcementLearningModel(IFreqaiModel):
for trade in open_trades: for trade in open_trades:
if trade.pair == pair: if trade.pair == pair:
# FIXME: mypy typing doesnt like that strategy may be "None" (it never will be) # FIXME: mypy typing doesnt like that strategy may be "None" (it never will be)
# FIXME: get_rate and trade_udration shouldn't work with backtesting,
# we need to use candle dates and prices to compute that.
current_value = self.strategy.dp._exchange.get_rate( current_value = self.strategy.dp._exchange.get_rate(
pair, refresh=False, side="exit", is_short=trade.is_short) pair, refresh=False, side="exit", is_short=trade.is_short)
openrate = trade.open_rate openrate = trade.open_rate
@ -256,7 +257,7 @@ def make_env(env_id: str, rank: int, seed: int, train_df: DataFrame, price: Data
env = MyRLEnv(df=train_df, prices=price, window_size=window_size, env = MyRLEnv(df=train_df, prices=price, window_size=window_size,
reward_kwargs=reward_params, id=env_id, seed=seed + rank, config=config) reward_kwargs=reward_params, id=env_id, seed=seed + rank, config=config)
if monitor: if monitor:
env = Monitor(env, ".") env = Monitor(env)
return env return env
set_random_seed(seed) set_random_seed(seed)
return _init return _init
@ -272,18 +273,19 @@ class MyRLEnv(Base5ActionRLEnv):
# first, penalize if the action is not valid # first, penalize if the action is not valid
if not self._is_valid(action): if not self._is_valid(action):
return -15 return -2
pnl = self.get_unrealized_profit() pnl = self.get_unrealized_profit()
rew = np.sign(pnl) * (pnl + 1) rew = np.sign(pnl) * (pnl + 1)
factor = 100 factor = 100
# reward agent for entering trades # reward agent for entering trades
if action in (Actions.Long_enter.value, Actions.Short_enter.value): if action in (Actions.Long_enter.value, Actions.Short_enter.value) \
and self._position == Positions.Neutral:
return 25 return 25
# discourage agent from not entering trades # discourage agent from not entering trades
if action == Actions.Neutral.value and self._position == Positions.Neutral: if action == Actions.Neutral.value and self._position == Positions.Neutral:
return -15 return -1
max_trade_duration = self.rl_config.get('max_trade_duration_candles', 300) max_trade_duration = self.rl_config.get('max_trade_duration_candles', 300)
trade_duration = self._current_tick - self._last_trade_tick trade_duration = self._current_tick - self._last_trade_tick
@ -294,8 +296,8 @@ class MyRLEnv(Base5ActionRLEnv):
factor *= 0.5 factor *= 0.5
# discourage sitting in position # discourage sitting in position
if self._position in (Positions.Short, Positions.Long): if self._position in (Positions.Short, Positions.Long) and action == Actions.Neutral.value:
return -50 * trade_duration / max_trade_duration return -1 * trade_duration / max_trade_duration
# close long # close long
if action == Actions.Long_exit.value and self._position == Positions.Long: if action == Actions.Long_exit.value and self._position == Positions.Long:

View File

@ -27,7 +27,7 @@ class ReinforcementLearner(BaseReinforcementLearningModel):
policy_kwargs = dict(activation_fn=th.nn.ReLU, policy_kwargs = dict(activation_fn=th.nn.ReLU,
net_arch=[512, 512, 256]) net_arch=[512, 512, 256])
if dk.pair not in self.dd.model_dictionary or not self.continual_retraining: if dk.pair not in self.dd.model_dictionary or not self.continual_learning:
model = self.MODELCLASS(self.policy_type, self.train_env, policy_kwargs=policy_kwargs, model = self.MODELCLASS(self.policy_type, self.train_env, policy_kwargs=policy_kwargs,
tensorboard_log=Path(dk.data_path / "tensorboard"), tensorboard_log=Path(dk.data_path / "tensorboard"),
**self.freqai_info['model_training_parameters'] **self.freqai_info['model_training_parameters']
@ -61,7 +61,6 @@ class ReinforcementLearner(BaseReinforcementLearningModel):
""" """
train_df = data_dictionary["train_features"] train_df = data_dictionary["train_features"]
test_df = data_dictionary["test_features"] test_df = data_dictionary["test_features"]
eval_freq = self.freqai_info["rl_config"]["eval_cycles"] * len(test_df)
self.train_env = MyRLEnv(df=train_df, prices=prices_train, window_size=self.CONV_WIDTH, self.train_env = MyRLEnv(df=train_df, prices=prices_train, window_size=self.CONV_WIDTH,
reward_kwargs=self.reward_params, config=self.config) reward_kwargs=self.reward_params, config=self.config)
@ -69,7 +68,7 @@ class ReinforcementLearner(BaseReinforcementLearningModel):
window_size=self.CONV_WIDTH, window_size=self.CONV_WIDTH,
reward_kwargs=self.reward_params, config=self.config)) reward_kwargs=self.reward_params, config=self.config))
self.eval_callback = EvalCallback(self.eval_env, deterministic=True, self.eval_callback = EvalCallback(self.eval_env, deterministic=True,
render=False, eval_freq=eval_freq, render=False, eval_freq=len(train_df),
best_model_save_path=str(dk.data_path)) best_model_save_path=str(dk.data_path))
@ -83,18 +82,19 @@ class MyRLEnv(Base5ActionRLEnv):
# first, penalize if the action is not valid # first, penalize if the action is not valid
if not self._is_valid(action): if not self._is_valid(action):
return -15 return -2
pnl = self.get_unrealized_profit() pnl = self.get_unrealized_profit()
rew = np.sign(pnl) * (pnl + 1) rew = np.sign(pnl) * (pnl + 1)
factor = 100 factor = 100
# reward agent for entering trades # reward agent for entering trades
if action in (Actions.Long_enter.value, Actions.Short_enter.value): if action in (Actions.Long_enter.value, Actions.Short_enter.value) \
and self._position == Positions.Neutral:
return 25 return 25
# discourage agent from not entering trades # discourage agent from not entering trades
if action == Actions.Neutral.value and self._position == Positions.Neutral: if action == Actions.Neutral.value and self._position == Positions.Neutral:
return -15 return -1
max_trade_duration = self.rl_config.get('max_trade_duration_candles', 300) max_trade_duration = self.rl_config.get('max_trade_duration_candles', 300)
trade_duration = self._current_tick - self._last_trade_tick trade_duration = self._current_tick - self._last_trade_tick
@ -105,8 +105,8 @@ class MyRLEnv(Base5ActionRLEnv):
factor *= 0.5 factor *= 0.5
# discourage sitting in position # discourage sitting in position
if self._position in (Positions.Short, Positions.Long): if self._position in (Positions.Short, Positions.Long) and action == Actions.Neutral.value:
return -50 * trade_duration / max_trade_duration return -1 * trade_duration / max_trade_duration
# close long # close long
if action == Actions.Long_exit.value and self._position == Positions.Long: if action == Actions.Long_exit.value and self._position == Positions.Long:

View File

@ -26,12 +26,19 @@ class ReinforcementLearner_multiproc(BaseReinforcementLearningModel):
# model arch # model arch
policy_kwargs = dict(activation_fn=th.nn.ReLU, policy_kwargs = dict(activation_fn=th.nn.ReLU,
net_arch=[512, 512, 256]) net_arch=[256, 256])
if dk.pair not in self.dd.model_dictionary or not self.continual_learning:
model = self.MODELCLASS(self.policy_type, self.train_env, policy_kwargs=policy_kwargs, model = self.MODELCLASS(self.policy_type, self.train_env, policy_kwargs=policy_kwargs,
tensorboard_log=Path(dk.full_path / "tensorboard"), tensorboard_log=Path(dk.full_path / "tensorboard"),
**self.freqai_info['model_training_parameters'] **self.freqai_info['model_training_parameters']
) )
else:
logger.info('Continual training activated - starting training from previously '
'trained agent.')
model = self.dd.model_dictionary[dk.pair]
model.tensorboard_log = Path(dk.data_path / "tensorboard")
model.set_env(self.train_env)
model.learn( model.learn(
total_timesteps=int(total_timesteps), total_timesteps=int(total_timesteps),
@ -57,8 +64,6 @@ class ReinforcementLearner_multiproc(BaseReinforcementLearningModel):
test_df = data_dictionary["test_features"] test_df = data_dictionary["test_features"]
eval_freq = self.freqai_info["rl_config"]["eval_cycles"] * len(test_df) eval_freq = self.freqai_info["rl_config"]["eval_cycles"] * len(test_df)
# environments
if not self.train_env:
env_id = "train_env" env_id = "train_env"
num_cpu = int(self.freqai_info["rl_config"]["thread_count"] / 2) num_cpu = int(self.freqai_info["rl_config"]["thread_count"] / 2)
self.train_env = SubprocVecEnv([make_env(env_id, i, 1, train_df, prices_train, self.train_env = SubprocVecEnv([make_env(env_id, i, 1, train_df, prices_train,
@ -74,13 +79,3 @@ class ReinforcementLearner_multiproc(BaseReinforcementLearningModel):
self.eval_callback = EvalCallback(self.eval_env, deterministic=True, self.eval_callback = EvalCallback(self.eval_env, deterministic=True,
render=False, eval_freq=eval_freq, render=False, eval_freq=eval_freq,
best_model_save_path=dk.data_path) best_model_save_path=dk.data_path)
else:
self.train_env.env_method('reset')
self.eval_env.env_method('reset')
self.train_env.env_method('reset_env', train_df, prices_train,
self.CONV_WIDTH, self.reward_params)
self.eval_env.env_method('reset_env', train_df, prices_train,
self.CONV_WIDTH, self.reward_params)
self.eval_callback.__init__(self.eval_env, deterministic=True,
render=False, eval_freq=eval_freq,
best_model_save_path=dk.data_path)