persist a single training environment.

config_examples/config_freqai-rl.example.json

@@ -61,7 +61,7 @@
         "train_period_days": 10,
         "backtest_period_days": 2,
         "identifier": "unique-id",
-        "data_kitchen_thread_count": 4,
+        "data_kitchen_thread_count": 2,
         "feature_parameters": {
             "include_corr_pairlist": [
                 "BTC/USDT",

freqtrade/freqai/RL/Base5ActionRLEnv.py

@@ -7,7 +7,7 @@ import numpy as np
 from gym import spaces
 from gym.utils import seeding
 from pandas import DataFrame
-
+import pandas as pd
 logger = logging.getLogger(__name__)
 
 
@@ -47,6 +47,9 @@ class Base5ActionRLEnv(gym.Env):
 
         self.id = id
         self.seed(seed)
+        self.reset_env(df, prices, window_size, reward_kwargs, starting_point)
+
+    def reset_env(self, df, prices, window_size, reward_kwargs, starting_point=True):
         self.df = df
         self.signal_features = self.df
         self.prices = prices
@@ -178,10 +181,15 @@ class Base5ActionRLEnv(gym.Env):
         return observation, step_reward, self._done, info
 
     def _get_observation(self):
-        features_and_state = self.signal_features[(
+        features_window = self.signal_features[(
             self._current_tick - self.window_size):self._current_tick]
+        features_and_state = DataFrame(np.zeros((len(features_window), 2)),
+                                       columns=['current_profit_pct', 'position'],
+                                       index=features_window.index)
+
         features_and_state['current_profit_pct'] = self.get_unrealized_profit()
         features_and_state['position'] = self._position.value
+        features_and_state = pd.concat([features_window, features_and_state], axis=1)
         return features_and_state
 
     def get_unrealized_profit(self):

freqtrade/freqai/RL/BaseReinforcementLearningModel.py

@@ -8,9 +8,10 @@ from pandas import DataFrame
 from abc import abstractmethod
 from freqtrade.freqai.data_kitchen import FreqaiDataKitchen
 from freqtrade.freqai.freqai_interface import IFreqaiModel
-from freqtrade.freqai.RL.Base3ActionRLEnv import Base3ActionRLEnv, Actions, Positions
+from freqtrade.freqai.RL.Base5ActionRLEnv import Base5ActionRLEnv, Actions, Positions
 from freqtrade.persistence import Trade
 import torch.multiprocessing
+from stable_baselines3.common.monitor import Monitor
 import torch as th
 logger = logging.getLogger(__name__)
 
@@ -26,6 +27,7 @@ class BaseReinforcementLearningModel(IFreqaiModel):
         super().__init__(config=kwargs['config'])
         th.set_num_threads(self.freqai_info.get('data_kitchen_thread_count', 4))
         self.reward_params = self.freqai_info['rl_config']['model_reward_parameters']
+        self.train_env: Base5ActionRLEnv = None
 
     def train(
         self, unfiltered_dataframe: DataFrame, pair: str, dk: FreqaiDataKitchen
@@ -65,15 +67,37 @@ class BaseReinforcementLearningModel(IFreqaiModel):
         )
         logger.info(f'Training model on {len(data_dictionary["train_features"])} data points')
 
-        model = self.fit_rl(data_dictionary, pair, dk, prices_train, prices_test)
+        self.set_train_and_eval_environments(data_dictionary, prices_train, prices_test)
+
+        model = self.fit_rl(data_dictionary, dk)
 
         logger.info(f"--------------------done training {pair}--------------------")
 
         return model
 
+    def set_train_and_eval_environments(self, data_dictionary, prices_train, prices_test):
+        """
+        User overrides this in their prediction model if they are custom a MyRLEnv. Othwerwise
+        leaving this will default to Base5ActEnv
+        """
+        train_df = data_dictionary["train_features"]
+        test_df = data_dictionary["test_features"]
+
+        # environments
+        if not self.train_env:
+            self.train_env = MyRLEnv(df=train_df, prices=prices_train, window_size=self.CONV_WIDTH,
+                                     reward_kwargs=self.reward_params)
+            self.eval_env = Monitor(MyRLEnv(df=test_df, prices=prices_test,
+                                    window_size=self.CONV_WIDTH,
+                                    reward_kwargs=self.reward_params), ".")
+        else:
+            self.train_env.reset_env(train_df, prices_train, self.CONV_WIDTH, self.reward_params)
+            self.eval_env.reset_env(train_df, prices_train, self.CONV_WIDTH, self.reward_params)
+            self.train_env.reset()
+            self.eval_env.reset()
+
     @abstractmethod
-    def fit_rl(self, data_dictionary: Dict[str, Any], pair: str, dk: FreqaiDataKitchen,
+    def fit_rl(self, data_dictionary: Dict[str, Any], dk: FreqaiDataKitchen):
-               prices_train: DataFrame, prices_test: DataFrame):
         """
         Agent customizations and abstract Reinforcement Learning customizations
         go in here. Abstract method, so this function must be overridden by
@@ -193,66 +217,39 @@ class BaseReinforcementLearningModel(IFreqaiModel):
         return
 
 
-class MyRLEnv(Base3ActionRLEnv):
+class MyRLEnv(Base5ActionRLEnv):
+    """
+    User can override any function in BaseRLEnv and gym.Env. Here the user
+    Adds 5 actions.
+    """
 
-    def step(self, action):
+    def calculate_reward(self, action):
-        self._done = False
-        self._current_tick += 1
 
-        if self._current_tick == self._end_tick:
+        if self._last_trade_tick is None:
-            self._done = True
+            return 0.
 
-        self.update_portfolio_log_returns(action)
+        # close long
+        if action == Actions.Long_sell.value and self._position == Positions.Long:
+            last_trade_price = self.add_buy_fee(self.prices.iloc[self._last_trade_tick].open)
+            current_price = self.add_sell_fee(self.prices.iloc[self._current_tick].open)
+            return float(np.log(current_price) - np.log(last_trade_price))
 
-        self._update_profit(action)
+        if action == Actions.Long_sell.value and self._position == Positions.Long:
-        step_reward = self._calculate_reward(action)
+            if self.close_trade_profit[-1] > self.profit_aim * self.rr:
-        self.total_reward += step_reward
+                last_trade_price = self.add_buy_fee(self.prices.iloc[self._last_trade_tick].open)
+                current_price = self.add_sell_fee(self.prices.iloc[self._current_tick].open)
+                return float((np.log(current_price) - np.log(last_trade_price)) * 2)
 
-        trade_type = None
+        # close short
-        if self.is_tradesignal(action):  # exclude 3 case not trade
+        if action == Actions.Short_buy.value and self._position == Positions.Short:
-            # Update position
+            last_trade_price = self.add_sell_fee(self.prices.iloc[self._last_trade_tick].open)
-            """
+            current_price = self.add_buy_fee(self.prices.iloc[self._current_tick].open)
-            Action: Neutral, position: Long ->  Close Long
+            return float(np.log(last_trade_price) - np.log(current_price))
-            Action: Neutral, position: Short -> Close Short
 
-            Action: Long, position: Neutral -> Open Long
+        if action == Actions.Short_buy.value and self._position == Positions.Short:
-            Action: Long, position: Short -> Close Short and Open Long
+            if self.close_trade_profit[-1] > self.profit_aim * self.rr:
+                last_trade_price = self.add_sell_fee(self.prices.iloc[self._last_trade_tick].open)
+                current_price = self.add_buy_fee(self.prices.iloc[self._current_tick].open)
+                return float((np.log(last_trade_price) - np.log(current_price)) * 2)
 
-            Action: Short, position: Neutral -> Open Short
+        return 0.
-            Action: Short, position: Long -> Close Long and Open Short
-            """
-
-            if action == Actions.Neutral.value:
-                self._position = Positions.Neutral
-                trade_type = "neutral"
-            elif action == Actions.Long.value:
-                self._position = Positions.Long
-                trade_type = "long"
-            elif action == Actions.Short.value:
-                self._position = Positions.Short
-                trade_type = "short"
-            else:
-                print("case not defined")
-
-            # Update last trade tick
-            self._last_trade_tick = self._current_tick
-
-            if trade_type is not None:
-                self.trade_history.append(
-                    {'price': self.current_price(), 'index': self._current_tick,
-                     'type': trade_type})
-
-        if self._total_profit < 0.2:
-            self._done = True
-
-        self._position_history.append(self._position)
-        observation = self._get_observation()
-        info = dict(
-            tick=self._current_tick,
-            total_reward=self.total_reward,
-            total_profit=self._total_profit,
-            position=self._position.value
-        )
-        self._update_history(info)
-
-        return observation, step_reward, self._done, info

freqtrade/freqai/prediction_models/ReinforcementLearningPPO_multiproc.py

@@ -3,9 +3,7 @@ from typing import Any, Dict  # , Tuple
 
 import numpy as np
 # import numpy.typing as npt
-# import pandas as pd
 import torch as th
-# from pandas import DataFrame
 from stable_baselines3.common.monitor import Monitor
 from typing import Callable
 from stable_baselines3 import PPO
@@ -16,7 +14,6 @@ from freqtrade.freqai.RL.Base3ActionRLEnv import Base3ActionRLEnv, Actions, Posi
 from freqtrade.freqai.RL.BaseReinforcementLearningModel import BaseReinforcementLearningModel
 from freqtrade.freqai.data_kitchen import FreqaiDataKitchen
 import gym
-from pandas import DataFrame
 
 logger = logging.getLogger(__name__)
 
@@ -48,26 +45,15 @@ class ReinforcementLearningPPO_multiproc(BaseReinforcementLearningModel):
     User created Reinforcement Learning Model prediction model.
     """
 
-    def fit_rl(self, data_dictionary: Dict[str, Any], pair: str, dk: FreqaiDataKitchen,
+    def fit_rl(self, data_dictionary: Dict[str, Any], dk: FreqaiDataKitchen):
-               prices_train: DataFrame, prices_test: DataFrame):
 
         train_df = data_dictionary["train_features"]
         test_df = data_dictionary["test_features"]
         eval_freq = self.freqai_info["rl_config"]["eval_cycles"] * len(test_df)
         total_timesteps = self.freqai_info["rl_config"]["train_cycles"] * len(train_df)
 
-        env_id = "train_env"
-        num_cpu = int(dk.thread_count / 2)
-        train_env = SubprocVecEnv([make_env(env_id, i, 1, train_df, prices_train,
-                                   self.reward_params, self.CONV_WIDTH) for i in range(num_cpu)])
-
-        eval_env_id = 'eval_env'
-        eval_env = SubprocVecEnv([make_env(eval_env_id, i, 1, test_df, prices_test,
-                                  self.reward_params, self.CONV_WIDTH, monitor=True) for i in
-                                  range(num_cpu)])
-
         path = dk.data_path
-        eval_callback = EvalCallback(eval_env, best_model_save_path=f"{path}/",
+        eval_callback = EvalCallback(self.eval_env, best_model_save_path=f"{path}/",
                                      log_path=f"{path}/ppo/logs/", eval_freq=int(eval_freq),
                                      deterministic=True, render=False)
 
@@ -75,7 +61,7 @@ class ReinforcementLearningPPO_multiproc(BaseReinforcementLearningModel):
         policy_kwargs = dict(activation_fn=th.nn.ReLU,
                              net_arch=[512, 512, 512])
 
-        model = PPO('MlpPolicy', train_env, policy_kwargs=policy_kwargs,
+        model = PPO('MlpPolicy', self.train_env, policy_kwargs=policy_kwargs,
                     tensorboard_log=f"{path}/ppo/tensorboard/",
                     **self.freqai_info['model_training_parameters']
                     )
@@ -87,10 +73,37 @@ class ReinforcementLearningPPO_multiproc(BaseReinforcementLearningModel):
 
         best_model = PPO.load(dk.data_path / "best_model")
         print('Training finished!')
-        eval_env.close()
 
         return best_model
 
+    def set_train_and_eval_environments(self, data_dictionary, prices_train, prices_test):
+        """
+        User overrides this in their prediction model if they are custom a MyRLEnv. Othwerwise
+        leaving this will default to Base5ActEnv
+        """
+        train_df = data_dictionary["train_features"]
+        test_df = data_dictionary["test_features"]
+
+        # environments
+        if not self.train_env:
+            env_id = "train_env"
+            num_cpu = int(self.freqai_info["data_kitchen_thread_count"] / 2)
+            self.train_env = SubprocVecEnv([make_env(env_id, i, 1, train_df, prices_train,
+                                            self.reward_params, self.CONV_WIDTH) for i
+                                            in range(num_cpu)])
+
+            eval_env_id = 'eval_env'
+            self.eval_env = SubprocVecEnv([make_env(eval_env_id, i, 1, test_df, prices_test,
+                                           self.reward_params, self.CONV_WIDTH, monitor=True) for i
+                                           in range(num_cpu)])
+        else:
+            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.train_env.env_method('reset')
+            self.eval_env.env_method('reset')
+
 
 class MyRLEnv(Base3ActionRLEnv):
     """

freqtrade/freqai/prediction_models/ReinforcementLearningTDQN.py

@@ -9,8 +9,7 @@ from freqtrade.freqai.RL.TDQNagent import TDQN
 from stable_baselines3 import DQN
 from stable_baselines3.common.buffers import ReplayBuffer
 import numpy as np
-from pandas import DataFrame
+import gc
-
 from freqtrade.freqai.data_kitchen import FreqaiDataKitchen
 
 logger = logging.getLogger(__name__)
@@ -21,24 +20,15 @@ class ReinforcementLearningTDQN(BaseReinforcementLearningModel):
     User created Reinforcement Learning Model prediction model.
     """
 
-    def fit_rl(self, data_dictionary: Dict[str, Any], pair: str, dk: FreqaiDataKitchen,
+    def fit_rl(self, data_dictionary: Dict[str, Any], dk: FreqaiDataKitchen):
-               prices_train: DataFrame, prices_test: DataFrame):
 
         train_df = data_dictionary["train_features"]
         test_df = data_dictionary["test_features"]
         eval_freq = self.freqai_info["rl_config"]["eval_cycles"] * len(test_df)
         total_timesteps = self.freqai_info["rl_config"]["train_cycles"] * len(train_df)
 
-        # environments
-        train_env = MyRLEnv(df=train_df, prices=prices_train, window_size=self.CONV_WIDTH,
-                            reward_kwargs=self.reward_params)
-        eval = MyRLEnv(df=test_df, prices=prices_test,
-                       window_size=self.CONV_WIDTH, reward_kwargs=self.reward_params)
-        eval_env = Monitor(eval, ".")
-        eval_env.reset()
-
         path = dk.data_path
-        eval_callback = EvalCallback(eval_env, best_model_save_path=f"{path}/",
+        eval_callback = EvalCallback(self.eval_env, best_model_save_path=f"{path}/",
                                      log_path=f"{path}/tdqn/logs/", eval_freq=int(eval_freq),
                                      deterministic=True, render=False)
 
@@ -46,7 +36,7 @@ class ReinforcementLearningTDQN(BaseReinforcementLearningModel):
         policy_kwargs = dict(activation_fn=th.nn.ReLU,
                              net_arch=[256, 256, 128])
 
-        model = TDQN('TMultiInputPolicy', train_env,
+        model = TDQN('TMultiInputPolicy', self.train_env,
                      tensorboard_log=f"{path}/tdqn/tensorboard/",
                      policy_kwargs=policy_kwargs,
                      replay_buffer_class=ReplayBuffer,
@@ -58,12 +48,33 @@ class ReinforcementLearningTDQN(BaseReinforcementLearningModel):
             callback=eval_callback
         )
 
+        del model
         best_model = DQN.load(dk.data_path / "best_model")
 
         print('Training finished!')
-
+        gc.collect()
         return best_model
 
+    def set_train_and_eval_environments(self, data_dictionary, prices_train, prices_test):
+        """
+        User overrides this as shown here if they are using a custom MyRLEnv
+        """
+        train_df = data_dictionary["train_features"]
+        test_df = data_dictionary["test_features"]
+
+        # environments
+        if not self.train_env:
+            self.train_env = MyRLEnv(df=train_df, prices=prices_train, window_size=self.CONV_WIDTH,
+                                     reward_kwargs=self.reward_params)
+            self.eval_env = Monitor(MyRLEnv(df=test_df, prices=prices_test,
+                                    window_size=self.CONV_WIDTH,
+                                    reward_kwargs=self.reward_params), ".")
+        else:
+            self.train_env.reset_env(train_df, prices_train, self.CONV_WIDTH, self.reward_params)
+            self.eval_env.reset_env(train_df, prices_train, self.CONV_WIDTH, self.reward_params)
+            self.train_env.reset()
+            self.eval_env.reset()
+
 
 # User can inherit and customize 5 action environment
 class MyRLEnv(Base5ActionRLEnv):

freqtrade/freqai/prediction_models/ReinforcementLearningTDQN_multiproc.py

@@ -4,8 +4,8 @@ import torch as th
 import numpy as np
 import gym
 from typing import Callable
-from stable_baselines3.common.callbacks import (
+from stable_baselines3.common.callbacks import EvalCallback
-    EvalCallback, StopTrainingOnNoModelImprovement, StopTrainingOnRewardThreshold)
+# EvalCallback , StopTrainingOnNoModelImprovement, StopTrainingOnRewardThreshold
 from stable_baselines3.common.monitor import Monitor
 from stable_baselines3.common.vec_env import SubprocVecEnv
 from stable_baselines3.common.utils import set_random_seed
@@ -15,7 +15,6 @@ from freqtrade.freqai.RL.BaseReinforcementLearningModel import BaseReinforcement
 from freqtrade.freqai.RL.TDQNagent import TDQN
 from stable_baselines3.common.buffers import ReplayBuffer
 from freqtrade.freqai.data_kitchen import FreqaiDataKitchen
-from pandas import DataFrame
 
 logger = logging.getLogger(__name__)
 
@@ -47,46 +46,23 @@ class ReinforcementLearningTDQN_multiproc(BaseReinforcementLearningModel):
     User created Reinforcement Learning Model prediction model.
     """
 
-    def fit_rl(self, data_dictionary: Dict[str, Any], pair: str, dk: FreqaiDataKitchen,
+    def fit_rl(self, data_dictionary: Dict[str, Any], dk: FreqaiDataKitchen):
-               prices_train: DataFrame, prices_test: DataFrame):
 
         train_df = data_dictionary["train_features"]
         test_df = data_dictionary["test_features"]
         eval_freq = self.freqai_info["rl_config"]["eval_cycles"] * len(test_df)
         total_timesteps = self.freqai_info["rl_config"]["train_cycles"] * len(train_df)
 
-        env_id = "train_env"
-        num_cpu = int(dk.thread_count / 2)
-        train_env = SubprocVecEnv([make_env(env_id, i, 1, train_df, prices_train,
-                                   self.reward_params, self.CONV_WIDTH) for i in range(num_cpu)])
-
-        eval_env_id = 'eval_env'
-        eval_env = SubprocVecEnv([make_env(eval_env_id, i, 1, test_df, prices_test,
-                                  self.reward_params, self.CONV_WIDTH, monitor=True) for i in
-                                  range(num_cpu)])
-
         path = dk.data_path
-        stop_train_callback = StopTrainingOnNoModelImprovement(
+
-            max_no_improvement_evals=5,
+        eval_callback = EvalCallback(self.eval_env, best_model_save_path=f"{path}/",
-            min_evals=10,
+                                     log_path=f"{path}/tdqn/logs/", eval_freq=int(eval_freq),
-            verbose=2
+                                     deterministic=True, render=False)
-        )
-        callback_on_best = StopTrainingOnRewardThreshold(reward_threshold=-200, verbose=2)
-        eval_callback = EvalCallback(
-            eval_env, best_model_save_path=f"{path}/",
-            log_path=f"{path}/tdqn/logs/",
-            eval_freq=int(eval_freq),
-            deterministic=True,
-            render=True,
-            callback_after_eval=stop_train_callback,
-            callback_on_new_best=callback_on_best,
-            verbose=2
-        )
         # model arch
         policy_kwargs = dict(activation_fn=th.nn.ReLU,
                              net_arch=[512, 512, 512])
 
-        model = TDQN('TMultiInputPolicy', train_env,
+        model = TDQN('TMultiInputPolicy', self.train_env,
                      policy_kwargs=policy_kwargs,
                      tensorboard_log=f"{path}/tdqn/tensorboard/",
                      replay_buffer_class=ReplayBuffer,
@@ -100,12 +76,40 @@ class ReinforcementLearningTDQN_multiproc(BaseReinforcementLearningModel):
 
         best_model = DQN.load(dk.data_path / "best_model.zip")
         print('Training finished!')
-        eval_env.close()
 
         return best_model
 
+    def set_train_and_eval_environments(self, data_dictionary, prices_train, prices_test):
+        """
+        User overrides this in their prediction model if they are custom a MyRLEnv. Othwerwise
+        leaving this will default to Base5ActEnv
+        """
+        train_df = data_dictionary["train_features"]
+        test_df = data_dictionary["test_features"]
+
+        # environments
+        if not self.train_env:
+            env_id = "train_env"
+            num_cpu = int(self.freqai_info["data_kitchen_thread_count"] / 2)
+            self.train_env = SubprocVecEnv([make_env(env_id, i, 1, train_df, prices_train,
+                                            self.reward_params, self.CONV_WIDTH) for i
+                                            in range(num_cpu)])
+
+            eval_env_id = 'eval_env'
+            self.eval_env = SubprocVecEnv([make_env(eval_env_id, i, 1, test_df, prices_test,
+                                           self.reward_params, self.CONV_WIDTH, monitor=True) for i
+                                           in range(num_cpu)])
+        else:
+            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.train_env.env_method('reset')
+            self.eval_env.env_method('reset')
 
 # User can inherit and customize 5 action environment
+
+
 class MyRLEnv(Base5ActionRLEnv):
     """
     User can override any function in BaseRLEnv and gym.Env. Here the user