Merge bb62b0fc5a into ae41ab101a

freqtrade/freqai/RL/Base5ActionRLEnv.py

@@ -22,12 +22,18 @@ class Base5ActionRLEnv(BaseEnvironment):
     Base class for a 5 action environment
     """
 
-    def __init__(self, **kwargs):
-        super().__init__(**kwargs)
+    def __init__(self, *args, action_space_type: str = "Discrete", **kwargs):
+        super().__init__(*args, **kwargs)
+        self.action_space_type = action_space_type
         self.actions = Actions
 
     def set_action_space(self):
-        self.action_space = spaces.Discrete(len(Actions))
+            if self.action_space_type == "Discrete":
+                self.action_space = spaces.Discrete(len(Actions))
+            elif self.action_space_type == "Box":
+                self.action_space = spaces.Box(low=-1, high=1, shape=(1,))
+            else:
+                raise ValueError(f"Unknown action space type: {self.action_space_type}")
 
     def step(self, action: int):
         """

freqtrade/freqai/RL/BaseEnvironment.py

@@ -60,6 +60,7 @@ class BaseEnvironment(gym.Env):
         can_short: bool = False,
         pair: str = "",
         df_raw: DataFrame = DataFrame(),
+        action_space_type: str = "Discrete"
     ):
         """
         Initializes the training/eval environment.
@@ -93,6 +94,7 @@ class BaseEnvironment(gym.Env):
         self.tensorboard_metrics: dict = {}
         self.can_short: bool = can_short
         self.live: bool = live
+        self.action_space_type: str = action_space_type
         if not self.live and self.add_state_info:
             raise OperationalException(
                 "`add_state_info` is not available in backtesting. Change "

freqtrade/freqai/RL/BaseReinforcementLearningModel.py

@@ -32,7 +32,7 @@ logger = logging.getLogger(__name__)
 
 torch.multiprocessing.set_sharing_strategy("file_system")
 
-SB3_MODELS = ["PPO", "A2C", "DQN"]
+SB3_MODELS = ["PPO", "A2C", "DQN", "DDPG", "TD3"]
 SB3_CONTRIB_MODELS = ["TRPO", "ARS", "RecurrentPPO", "MaskablePPO", "QRDQN"]
 
 

freqtrade/freqai/prediction_models/ReinforcementLearner_DDPG_TD3.py

@@ -0,0 +1,556 @@
+import copy
+import logging
+import gc
+from pathlib import Path
+from typing import Any, Dict, Type, Callable, List, Optional, Union
+
+import numpy as np
+import torch as th
+import pandas as pd
+from pandas import DataFrame
+from gymnasium import spaces
+import matplotlib
+import matplotlib.transforms as mtransforms
+import matplotlib.pyplot as plt
+from stable_baselines3.common.callbacks import BaseCallback, EvalCallback
+from stable_baselines3.common.logger import HParam, Figure
+
+from freqtrade.freqai.data_kitchen import FreqaiDataKitchen
+from freqtrade.freqai.RL.Base5ActionRLEnv import Actions, Base5ActionRLEnv, Positions
+from freqtrade.freqai.RL.BaseEnvironment import BaseEnvironment, BaseActions
+from freqtrade.freqai.RL.BaseReinforcementLearningModel import BaseReinforcementLearningModel
+from freqtrade.freqai.tensorboard.TensorboardCallback import TensorboardCallback
+
+
+logger = logging.getLogger(__name__)
+
+
+class ReinforcementLearner_DDPG_TD3(BaseReinforcementLearningModel):
+    """
+    Reinforcement Learning Model prediction model for DDPG and TD3.
+
+    Users can inherit from this class to make their own RL model with custom
+    environment/training controls. Define the file as follows:
+
+    ```
+    from freqtrade.freqai.prediction_models.ReinforcementLearner import ReinforcementLearner
+
+    class MyCoolRLModel(ReinforcementLearner):
+    ```
+
+    Save the file to `user_data/freqaimodels`, then run it with:
+
+    freqtrade trade --freqaimodel MyCoolRLModel --config config.json --strategy SomeCoolStrat
+
+    Here the users can override any of the functions
+    available in the `IFreqaiModel` inheritance tree. Most importantly for RL, this
+    is where the user overrides `MyRLEnv` (see below), to define custom
+    `calculate_reward()` function, or to override any other parts of the environment.
+
+    This class also allows users to override any other part of the IFreqaiModel tree.
+    For example, the user can override `def fit()` or `def train()` or `def predict()`
+    to take fine-tuned control over these processes.
+
+    Another common override may be `def data_cleaning_predict()` where the user can
+    take fine-tuned control over the data handling pipeline.
+    """
+
+    def __init__(self, **kwargs) -> None:
+        """
+        Model specific config
+        """
+        super().__init__(**kwargs)
+
+        # Enable learning rate linear schedule
+        self.lr_schedule: bool = self.rl_config.get("lr_schedule", False)
+
+        # Enable tensorboard logging
+        self.activate_tensorboard: bool = self.rl_config.get("activate_tensorboard", True)
+        # TENSORBOARD CALLBACK DOES NOT RECOMMENDED TO USE WITH MULTIPLE ENVS,
+        # IT WILL RETURN FALSE INFORMATIONS, NEVERTHLESS NOT THREAD SAFE WITH SB3!!!
+
+        # Enable tensorboard rollout plot
+        self.tensorboard_plot: bool = self.rl_config.get("tensorboard_plot", False)
+
+    def get_model_params(self):
+        """
+        Get the model specific parameters
+        """
+        model_params = copy.deepcopy(self.freqai_info["model_training_parameters"])
+        
+        if self.lr_schedule:
+            _lr = model_params.get('learning_rate', 0.0003)
+            model_params["learning_rate"] = linear_schedule(_lr)
+            logger.info(f"Learning rate linear schedule enabled, initial value: {_lr}")
+
+        model_params["policy_kwargs"] = dict(
+        net_arch=dict(vf=self.net_arch, pi=self.net_arch),
+        activation_fn=th.nn.ReLU,
+        optimizer_class=th.optim.Adam
+
+        return model_params
+
+    def get_callbacks(self, eval_freq, data_path) -> list:
+        """
+        Get the model specific callbacks
+        """
+        callbacks = []
+        callbacks.append(self.eval_callback)
+        if self.activate_tensorboard:
+            callbacks.append(CustomTensorboardCallback())
+        if self.tensorboard_plot:
+            callbacks.append(FigureRecorderCallback())
+        return callbacks
+
+    def fit(self, data_dictionary: Dict[str, Any], dk: FreqaiDataKitchen, **kwargs):
+        """
+        User customizable fit method
+        :param data_dictionary: dict = common data dictionary containing all train/test
+            features/labels/weights.
+        :param dk: FreqaiDatakitchen = data kitchen for current pair.
+        :return:
+        model Any = trained model to be used for inference in dry/live/backtesting
+        """
+        train_df = data_dictionary["train_features"]
+        total_timesteps = self.freqai_info["rl_config"]["train_cycles"] * len(train_df)
+
+        policy_kwargs = dict(activation_fn=th.nn.ReLU,
+                             net_arch=self.net_arch)
+
+        if self.activate_tensorboard:
+            tb_path = Path(dk.full_path / "tensorboard" / dk.pair.split('/')[0])
+        else:
+            tb_path = None
+
+        model_params = self.get_model_params()
+        logger.info(f"Params: {model_params}")
+
+        if dk.pair not in self.dd.model_dictionary or not self.continual_learning:
+            model = self.MODELCLASS(self.policy_type, self.train_env,
+                                    tensorboard_log=tb_path,
+                                    **model_params)
+        else:
+            logger.info("Continual training activated - starting training from previously "
+                        "trained agent.")
+            model = self.dd.model_dictionary[dk.pair]
+            model.set_env(self.train_env)
+
+        model.learn(
+            total_timesteps=int(total_timesteps),
+            #callback=[self.eval_callback, self.tensorboard_callback],
+            callback=self.get_callbacks(len(train_df), str(dk.data_path)),
+            progress_bar=self.rl_config.get("progress_bar", False)
+        )
+
+        if Path(dk.data_path / "best_model.zip").is_file():
+            logger.info("Callback found a best model.")
+            best_model = self.MODELCLASS.load(dk.data_path / "best_model")
+            return best_model
+
+        logger.info("Couldnt find best model, using final model instead.")
+
+        return model
+
+    MyRLEnv: Type[BaseEnvironment]
+
+    class MyRLEnv(Base5ActionRLEnv):  # type: ignore[no-redef]
+        """
+        User can override any function in BaseRLEnv and gym.Env. Here the user
+        sets a custom reward based on profit and trade duration.
+        """
+        def __init__(self, df, prices, reward_kwargs, window_size=10, starting_point=True, id="boxenv-1", seed=1, config={}, live=False, fee=0.0015, can_short=False, pair="", df_raw=None, action_space_type="Box"):
+            super().__init__(df, prices, reward_kwargs, window_size, starting_point, id, seed, config, live, fee, can_short, pair, df_raw)
+            
+            # Define the action space as a continuous space between -1 and 1 for a single action dimension
+            self.action_space = spaces.Box(low=-1, high=1, shape=(1,), dtype=np.float32)
+            
+            # Define the observation space as before
+            self.observation_space = spaces.Box(
+                low=-np.inf,
+                high=np.inf,
+                shape=(window_size, self.total_features),
+                dtype=np.float32
+            )
+
+        def calculate_reward(self, action: int) -> float:
+            """
+            An example reward function. This is the one function that users will likely
+            wish to inject their own creativity into.
+
+                        Warning!
+            This is function is a showcase of functionality designed to show as many possible
+            environment control features as possible. It is also designed to run quickly
+            on small computers. This is a benchmark, it is *not* for live production.
+
+            :param action: int = The action made by the agent for the current candle.
+            :return:
+            float = the reward to give to the agent for current step (used for optimization
+                of weights in NN)
+            """
+            # first, penalize if the action is not valid
+            if not self._is_valid(action):
+                self.tensorboard_log("invalid", category="actions")
+                return -2
+
+            pnl = self.get_unrealized_profit()
+            factor = 100.
+
+            # reward agent for entering trades
+            if (action == Actions.Long_enter.value
+                    and self._position == Positions.Neutral):
+                return 25
+            if (action == Actions.Short_enter.value
+                    and self._position == Positions.Neutral):
+                return 25
+            # discourage agent from not entering trades
+            if action == Actions.Neutral.value and self._position == Positions.Neutral:
+                return -1
+
+            max_trade_duration = self.rl_config.get('max_trade_duration_candles', 300)
+            trade_duration = self._current_tick - self._last_trade_tick  # type: ignore
+
+            if trade_duration <= max_trade_duration:
+                factor *= 1.5
+            elif trade_duration > max_trade_duration:
+                factor *= 0.5
+
+            # discourage sitting in position
+            if (self._position in (Positions.Short, Positions.Long) and
+                    action == Actions.Neutral.value):
+                return -1 * trade_duration / max_trade_duration
+
+            # close long
+            if action == Actions.Long_exit.value and self._position == Positions.Long:
+                if pnl > self.profit_aim * self.rr:
+                    factor *= self.rl_config["model_reward_parameters"].get("win_reward_factor", 2)
+                return float(pnl * factor)
+
+            # close short
+            if action == Actions.Short_exit.value and self._position == Positions.Short:
+                if pnl > self.profit_aim * self.rr:
+                    factor *= self.rl_config["model_reward_parameters"].get("win_reward_factor", 2)
+                return float(pnl * factor)
+
+            return 0.
+
+        def step(self, action):
+            """
+            Logic for a single step (incrementing one candle in time)
+            by the agent
+            :param: action: int = the action type that the agent plans
+                to take for the current step.
+            :returns:
+                observation = current state of environment
+                step_reward = the reward from `calculate_reward()`
+                _done = if the agent "died" or if the candles finished
+                info = dict passed back to openai gym lib
+            """
+            
+            # Ensure action is within the range [-1, 1]
+            action = np.clip(action, -1, 1)
+            
+            # Apply noise for exploration
+            self.noise_std = 0.3  # Standard deviation for exploration noise
+            noise = np.random.normal(0, self.noise_std, size=action.shape)
+            action = np.tanh(action + noise)  # Ensure action is within -1 to 1
+    
+            # Map the continuous action to one of the five discrete actions
+            discrete_action = self._map_continuous_to_discrete(action)
+            
+            #print(f"{self._current_tick} Action!!!: {action}")
+            #print(f"{self._current_tick} Discrete Action!!!: {discrete_action}")
+
+            self._done = False
+            self._current_tick += 1
+    
+            if self._current_tick == self._end_tick:
+                self._done = True
+    
+            self._update_unrealized_total_profit()
+            step_reward = self.calculate_reward(discrete_action)
+            self.total_reward += step_reward
+            
+            self.tensorboard_log(self.actions._member_names_[discrete_action], category="actions")
+    
+            trade_type = None
+            if self.is_tradesignal(discrete_action):
+    
+                if discrete_action == Actions.Neutral.value:
+                    self._position = Positions.Neutral
+                    trade_type = "neutral"
+                    self._last_trade_tick = None
+                elif discrete_action == Actions.Long_enter.value:
+                    self._position = Positions.Long
+                    trade_type = "enter_long"
+                    self._last_trade_tick = self._current_tick
+                elif discrete_action == Actions.Short_enter.value:
+                    self._position = Positions.Short
+                    trade_type = "enter_short"
+                    self._last_trade_tick = self._current_tick
+                elif discrete_action == Actions.Long_exit.value:
+                    self._update_total_profit()
+                    self._position = Positions.Neutral
+                    trade_type = "exit_long"
+                    self._last_trade_tick = None
+                elif discrete_action == Actions.Short_exit.value:
+                    self._update_total_profit()
+                    self._position = Positions.Neutral
+                    trade_type = "exit_short"
+                    self._last_trade_tick = None
+                else:
+                    print("case not defined")
+    
+                if trade_type is not None:
+                    self.trade_history.append(
+                        {"price": self.current_price(), "index": self._current_tick,
+                         "type": trade_type, "profit": self.get_unrealized_profit()})
+    
+            if (self._total_profit < self.max_drawdown or
+                    self._total_unrealized_profit < self.max_drawdown):
+                self._done = True
+    
+            self._position_history.append(self._position)
+    
+            info = dict(
+                tick=self._current_tick,
+                action=discrete_action,
+                total_reward=self.total_reward,
+                total_profit=self._total_profit,
+                position=self._position.value,
+                trade_duration=self.get_trade_duration(),
+                current_profit_pct=self.get_unrealized_profit()
+            )
+    
+            observation = self._get_observation()
+            # user can play with time if they want
+            truncated = False
+    
+            self._update_history(info)
+    
+            return observation, step_reward, self._done, truncated, info
+    
+        def _map_continuous_to_discrete(self, action):
+            """
+            Map the continuous action (a value between -1 and 1) to one of the discrete actions.
+            """
+            action_value = action[0]  # Extract the single continuous action value
+            
+            # Define the number of discrete actions
+            num_discrete_actions = 5
+            
+            # Calculate the step size for each interval
+            step_size = 2 / num_discrete_actions  # (2 because range is from -1 to 1)
+            
+            # Generate the boundaries dynamically
+            boundaries = th.linspace(-1 + step_size, 1 - step_size, steps=num_discrete_actions - 1)
+            
+            # Find the bucket index for the action value
+            bucket_index = th.bucketize(th.tensor(action_value), boundaries, right=True)
+            
+            # Map the bucket index to discrete actions
+            discrete_actions = [
+                BaseActions.Neutral,
+                BaseActions.Long_enter,
+                BaseActions.Long_exit,
+                BaseActions.Short_enter,
+                BaseActions.Short_exit
+            ]
+            
+            return discrete_actions[bucket_index].value
+
+        def get_rollout_history(self) -> DataFrame:
+            """
+            Get environment data from the first to the last trade
+            """
+            _history_df = pd.DataFrame.from_dict(self.history)
+            _trade_history_df = pd.DataFrame.from_dict(self.trade_history)
+            _rollout_history = _history_df.merge(_trade_history_df, left_on="tick", right_on="index", how="left")
+            
+            _price_history = self.prices.iloc[_rollout_history.tick].copy().reset_index()
+
+            history = pd.merge(
+                _rollout_history,
+                _price_history,
+                left_index=True, right_index=True
+            )
+            return history
+
+        def get_rollout_plot(self):
+            """
+            Plot trades and environment data
+            """
+            def transform_y_offset(ax, offset):
+                return mtransforms.offset_copy(ax.transData, fig=fig, x=0, y=offset, units="inches")
+
+            def plot_markers(ax, ticks, marker, color, size, offset):
+                ax.plot(ticks, marker=marker, color=color, markersize=size, fillstyle="full",
+                        transform=transform_y_offset(ax, offset), linestyle="none")
+
+            plt.style.use("dark_background")
+            fig, axs = plt.subplots(
+                nrows=5, ncols=1,
+                figsize=(16, 9),
+                height_ratios=[6, 1, 1, 1, 1],
+                sharex=True
+            )
+
+            # Return empty fig if no trades
+            if len(self.trade_history) == 0:
+                return fig
+
+            history = self.get_rollout_history()
+            enter_long_prices = history.loc[history["type"] == "enter_long"]["price"]
+            enter_short_prices = history.loc[history["type"] == "enter_short"]["price"]
+            exit_long_prices = history.loc[history["type"] == "exit_long"]["price"]
+            exit_short_prices = history.loc[history["type"] == "exit_short"]["price"]
+
+            axs[0].plot(history["open"], linewidth=1, color="#c28ce3")
+            plot_markers(axs[0], enter_long_prices, "^", "#4ae747", 5, -0.05)
+            plot_markers(axs[0], enter_short_prices, "v", "#f53580", 5, 0.05)
+            plot_markers(axs[0], exit_long_prices, "o", "#4ae747", 3, 0)
+            plot_markers(axs[0], exit_short_prices, "o", "#f53580", 3, 0)
+
+            axs[1].set_ylabel("pnl")
+            axs[1].plot(history["current_profit_pct"], linewidth=1, color="#a29db9")
+            axs[1].axhline(y=0, label='0', alpha=0.33)
+            axs[2].set_ylabel("duration")
+            axs[2].plot(history["trade_duration"], linewidth=1, color="#a29db9")
+            axs[3].set_ylabel("total_reward")
+            axs[3].plot(history["total_reward"], linewidth=1, color="#a29db9")
+            axs[3].axhline(y=0, label='0', alpha=0.33)
+            axs[4].set_ylabel("total_profit")
+            axs[4].set_xlabel("tick")
+            axs[4].plot(history["total_profit"], linewidth=1, color="#a29db9")
+            axs[4].axhline(y=1, label='1', alpha=0.33)
+
+            for _ax in axs:
+                for _border in ["top", "right", "bottom", "left"]:
+                    _ax.spines[_border].set_color("#5b5e4b")
+
+            fig.suptitle(
+                "Total Reward: %.6f" % self.total_reward + " ~ " +
+                "Total Profit: %.6f" % self._total_profit
+            )
+            fig.tight_layout()
+
+            return fig
+
+        def close(self) -> None:
+            gc.collect()
+            th.cuda.empty_cache()
+
+def linear_schedule(initial_value: float) -> Callable[[float], float]:
+    def func(progress_remaining: float) -> float:
+        return progress_remaining * initial_value
+    return func
+
+class CustomTensorboardCallback(TensorboardCallback):
+    """
+    Tensorboard callback
+    """
+
+    def _on_training_start(self) -> None:
+        _lr = self.model.learning_rate
+        
+        if self.model.__class__.__name__ == "DDPG":
+            hparam_dict = {
+                "algorithm": self.model.__class__.__name__,
+                "buffer_size": self.model.buffer_size,
+                "learning_rate": _lr if isinstance(_lr, float) else "lr_schedule",
+                "learning_starts": self.model.learning_starts,
+                "batch_size": self.model.batch_size,
+                "tau": self.model.tau,
+                "gamma": self.model.gamma,
+                "train_freq": self.model.train_freq,
+                "gradient_steps": self.model.gradient_steps,
+            }
+
+        elif self.model.__class__.__name__ == "TD3":
+            hparam_dict = {
+                "algorithm": self.model.__class__.__name__,
+                "learning_rate": _lr if isinstance(_lr, float) else "lr_schedule",
+                "buffer_size": self.model.buffer_size,
+                "learning_starts": self.model.learning_starts,
+                "batch_size": self.model.batch_size,
+                "tau": self.model.tau,
+                "gamma": self.model.gamma,
+                "train_freq": self.model.train_freq,
+                "gradient_steps": self.model.gradient_steps,
+                "policy_delay": self.model.policy_delay,
+                "target_policy_noise": self.model.target_policy_noise,
+                "target_noise_clip": self.model.target_noise_clip,
+            }
+            
+        else:
+            hparam_dict = {
+                "algorithm": self.model.__class__.__name__,
+                "learning_rate": _lr if isinstance(_lr, float) else "lr_schedule",
+                "gamma": self.model.gamma,
+                "gae_lambda": self.model.gae_lambda,
+                "n_steps": self.model.n_steps,
+                "batch_size": self.model.batch_size,
+            }
+
+        # Convert hparam_dict values to str if they are not of type int, float, str, bool, or torch.Tensor
+        hparam_dict = {k: (str(v) if not isinstance(v, (int, float, str, bool, th.Tensor)) else v) for k, v in hparam_dict.items()}
+        
+        metric_dict = {
+            "eval/mean_reward": 0,
+            "rollout/ep_rew_mean": 0,
+            "rollout/ep_len_mean": 0,
+            "info/total_profit": 1,
+            "info/trades_count": 0,
+            "info/trade_duration": 0,
+        }
+
+        self.logger.record(
+            "hparams",
+            HParam(hparam_dict, metric_dict),
+            exclude=("stdout", "log", "json", "csv"),
+        )   
+    
+    def _on_step(self) -> bool:
+
+        local_info = self.locals["infos"][0]
+        if self.training_env is None:
+            return True
+
+        tensorboard_metrics = self.training_env.env_method("get_wrapper_attr", "tensorboard_metrics")[0]
+
+        for metric in local_info:
+            if metric not in ["episode", "terminal_observation", "TimeLimit.truncated"]:
+                self.logger.record(f"info/{metric}", local_info[metric])
+
+        for category in tensorboard_metrics:
+            for metric in tensorboard_metrics[category]:
+                self.logger.record(f"{category}/{metric}", tensorboard_metrics[category][metric])
+
+        return True
+    
+class FigureRecorderCallback(BaseCallback):
+    """
+    Tensorboard figures callback
+    """
+
+    def __init__(self, verbose=0):
+        super().__init__(verbose)
+
+    def _on_step(self) -> bool:
+        return True
+
+    def _on_rollout_end(self):
+        try:
+            # Access the rollout plot directly from the base environment
+            figures = [env.unwrapped.get_rollout_plot() for env in self.training_env.envs]
+        except AttributeError:
+            # If the above fails, try getting it from the wrappers
+            figures = self.training_env.env_method("get_wrapper_attr", "get_rollout_plot")
+
+        for i, fig in enumerate(figures):
+            self.logger.record(
+                f"rollout/env_{i}",
+                Figure(fig, close=True),
+                exclude=("stdout", "log", "json", "csv")
+            )
+            plt.close(fig)
+        return True