import collections import importlib import logging import re import shutil import threading from datetime import datetime, timedelta, timezone from pathlib import Path from typing import Any, Dict, Tuple, TypedDict import numpy as np import pandas as pd import psutil import rapidjson from joblib import dump, load from joblib.externals import cloudpickle from numpy.typing import NDArray from pandas import DataFrame from freqtrade.configuration import TimeRange from freqtrade.constants import Config from freqtrade.data.history import load_pair_history from freqtrade.exceptions import OperationalException from freqtrade.freqai.data_kitchen import FreqaiDataKitchen from freqtrade.strategy.interface import IStrategy logger = logging.getLogger(__name__) class pair_info(TypedDict): model_filename: str trained_timestamp: int data_path: str extras: dict class FreqaiDataDrawer: """ Class aimed at holding all pair models/info in memory for better inferencing/retrainig/saving /loading to/from disk. This object remains persistent throughout live/dry. Record of contribution: FreqAI was developed by a group of individuals who all contributed specific skillsets to the project. Conception and software development: Robert Caulk @robcaulk Theoretical brainstorming: Elin Törnquist @th0rntwig Code review, software architecture brainstorming: @xmatthias Beta testing and bug reporting: @bloodhunter4rc, Salah Lamkadem @ikonx, @ken11o2, @longyu, @paranoidandy, @smidelis, @smarm Juha Nykänen @suikula, Wagner Costa @wagnercosta, Johan Vlugt @Jooopieeert """ def __init__(self, full_path: Path, config: Config): self.config = config self.freqai_info = config.get("freqai", {}) # dictionary holding all pair metadata necessary to load in from disk self.pair_dict: Dict[str, pair_info] = {} # dictionary holding all actively inferenced models in memory given a model filename self.model_dictionary: Dict[str, Any] = {} # all additional metadata that we want to keep in ram self.meta_data_dictionary: Dict[str, Dict[str, Any]] = {} self.model_return_values: Dict[str, DataFrame] = {} self.historic_data: Dict[str, Dict[str, DataFrame]] = {} self.historic_predictions: Dict[str, DataFrame] = {} self.full_path = full_path self.historic_predictions_path = Path(self.full_path / "historic_predictions.pkl") self.historic_predictions_path_parquet = Path( self.full_path / "historic_predictions.parquet") self.historic_predictions_bkp_path_parquet = Path( self.full_path / "historic_predictions.backup.parquet") self.pair_dictionary_path = Path(self.full_path / "pair_dictionary.json") self.global_metadata_path = Path(self.full_path / "global_metadata.json") self.metric_tracker_path = Path(self.full_path / "metric_tracker.json") self.load_drawer_from_disk() self.load_historic_predictions_from_disk() self.metric_tracker: Dict[str, Dict[str, Dict[str, list]]] = {} self.load_metric_tracker_from_disk() self.training_queue: Dict[str, int] = {} self.history_lock = threading.Lock() self.save_lock = threading.Lock() self.pair_dict_lock = threading.Lock() self.metric_tracker_lock = threading.Lock() self.old_DBSCAN_eps: Dict[str, float] = {} self.empty_pair_dict: pair_info = { "model_filename": "", "trained_timestamp": 0, "data_path": "", "extras": {}} self.model_type = self.freqai_info.get('model_save_type', 'joblib') def update_metric_tracker(self, metric: str, value: float, pair: str) -> None: """ General utility for adding and updating custom metrics. Typically used for adding training performance, train timings, inferenc timings, cpu loads etc. """ with self.metric_tracker_lock: if pair not in self.metric_tracker: self.metric_tracker[pair] = {} if metric not in self.metric_tracker[pair]: self.metric_tracker[pair][metric] = {'timestamp': [], 'value': []} timestamp = int(datetime.now(timezone.utc).timestamp()) self.metric_tracker[pair][metric]['value'].append(value) self.metric_tracker[pair][metric]['timestamp'].append(timestamp) def collect_metrics(self, time_spent: float, pair: str): """ Add metrics to the metric tracker dictionary """ load1, load5, load15 = psutil.getloadavg() cpus = psutil.cpu_count() self.update_metric_tracker('train_time', time_spent, pair) self.update_metric_tracker('cpu_load1min', load1 / cpus, pair) self.update_metric_tracker('cpu_load5min', load5 / cpus, pair) self.update_metric_tracker('cpu_load15min', load15 / cpus, pair) def load_global_metadata_from_disk(self): """ Locate and load a previously saved global metadata in present model folder. """ exists = self.global_metadata_path.is_file() if exists: with self.global_metadata_path.open("r") as fp: metatada_dict = rapidjson.load(fp, number_mode=rapidjson.NM_NATIVE) return metatada_dict return {} def load_drawer_from_disk(self): """ Locate and load a previously saved data drawer full of all pair model metadata in present model folder. Load any existing metric tracker that may be present. """ exists = self.pair_dictionary_path.is_file() if exists: with self.pair_dictionary_path.open("r") as fp: self.pair_dict = rapidjson.load(fp, number_mode=rapidjson.NM_NATIVE) else: logger.info("Could not find existing datadrawer, starting from scratch") def load_metric_tracker_from_disk(self): """ Tries to load an existing metrics dictionary if the user wants to collect metrics. """ if self.freqai_info.get('write_metrics_to_disk', False): exists = self.metric_tracker_path.is_file() if exists: with self.metric_tracker_path.open("r") as fp: self.metric_tracker = rapidjson.load(fp, number_mode=rapidjson.NM_NATIVE) logger.info("Loading existing metric tracker from disk.") else: logger.info("Could not find existing metric tracker, starting from scratch") def load_historic_predictions_from_disk(self): """ Locate and load a previously saved historic predictions. :return: bool - whether or not the drawer was located """ exists = self.historic_predictions_path_parquet.is_file() convert = self.historic_predictions_path.is_file() if exists: try: self.historic_predictions = pd.read_parquet(self.historic_predictions_path_parquet) logger.info( f"Found existing historic predictions at {self.full_path}, but beware " "that statistics may be inaccurate if the bot has been offline for " "an extended period of time." ) except EOFError: logger.warning( 'Historical prediction file was corrupted. Trying to load backup file.') self.historic_predictions = pd.read_parquet( self.historic_predictions_bkp_path_parquet) logger.warning('FreqAI successfully loaded the backup historical predictions file.') elif not exists and convert: logger.info("Converting your historic predictions pkl to parquet" "to improve performance.") with Path.open(self.historic_predictions_path, "rb") as fp: self.historic_predictions = cloudpickle.load(fp) self.historic_predictions.to_parquet(self.historic_predictions_path_parquet) exists = True else: logger.warning( f"Follower could not find historic predictions at {self.full_path} " "sending null values back to strategy" ) return exists def save_historic_predictions_to_disk(self): """ Save historic predictions pickle to disk """ self.historic_predictions.to_parquet(self.historic_predictions_path_parquet) # create a backup shutil.copy(self.historic_predictions_path, self.historic_predictions_bkp_path_parquet) def save_metric_tracker_to_disk(self): """ Save metric tracker of all pair metrics collected. """ with self.save_lock: with self.metric_tracker_path.open('w') as fp: rapidjson.dump(self.metric_tracker, fp, default=self.np_encoder, number_mode=rapidjson.NM_NATIVE) def save_drawer_to_disk(self) -> None: """ Save data drawer full of all pair model metadata in present model folder. """ with self.save_lock: with self.pair_dictionary_path.open('w') as fp: rapidjson.dump(self.pair_dict, fp, default=self.np_encoder, number_mode=rapidjson.NM_NATIVE) def save_global_metadata_to_disk(self, metadata: Dict[str, Any]): """ Save global metadata json to disk """ with self.save_lock: with self.global_metadata_path.open('w') as fp: rapidjson.dump(metadata, fp, default=self.np_encoder, number_mode=rapidjson.NM_NATIVE) def np_encoder(self, object): if isinstance(object, np.generic): return object.item() def get_pair_dict_info(self, pair: str) -> Tuple[str, int]: """ Locate and load existing model metadata from persistent storage. If not located, create a new one and append the current pair to it and prepare it for its first training :param pair: str: pair to lookup :return: model_filename: str = unique filename used for loading persistent objects from disk trained_timestamp: int = the last time the coin was trained """ pair_dict = self.pair_dict.get(pair) if pair_dict: model_filename = pair_dict["model_filename"] trained_timestamp = pair_dict["trained_timestamp"] else: self.pair_dict[pair] = self.empty_pair_dict.copy() model_filename = "" trained_timestamp = 0 return model_filename, trained_timestamp def set_pair_dict_info(self, metadata: dict) -> None: pair_in_dict = self.pair_dict.get(metadata["pair"]) if pair_in_dict: return else: self.pair_dict[metadata["pair"]] = self.empty_pair_dict.copy() return def set_initial_return_values(self, pair: str, pred_df: DataFrame) -> None: """ Set the initial return values to the historical predictions dataframe. This avoids needing to repredict on historical candles, and also stores historical predictions despite retrainings (so stored predictions are true predictions, not just inferencing on trained data) """ hist_df = self.historic_predictions len_diff = len(hist_df[pair].index) - len(pred_df.index) if len_diff < 0: df_concat = pd.concat([pred_df.iloc[:abs(len_diff)], hist_df[pair]], ignore_index=True, keys=hist_df[pair].keys()) else: df_concat = hist_df[pair].tail(len(pred_df.index)).reset_index(drop=True) df_concat = df_concat.fillna(0) self.model_return_values[pair] = df_concat def append_model_predictions(self, pair: str, predictions: DataFrame, do_preds: NDArray[np.int_], dk: FreqaiDataKitchen, strat_df: DataFrame) -> None: """ Append model predictions to historic predictions dataframe, then set the strategy return dataframe to the tail of the historic predictions. The length of the tail is equivalent to the length of the dataframe that entered FreqAI from the strategy originally. Doing this allows FreqUI to always display the correct historic predictions. """ len_df = len(strat_df) index = self.historic_predictions[pair].index[-1:] columns = self.historic_predictions[pair].columns nan_df = pd.DataFrame(np.nan, index=index, columns=columns) self.historic_predictions[pair] = pd.concat( [self.historic_predictions[pair], nan_df], ignore_index=True, axis=0) df = self.historic_predictions[pair] # model outputs and associated statistics for label in predictions.columns: df[label].iloc[-1] = predictions[label].iloc[-1] if df[label].dtype == object: continue df[f"{label}_mean"].iloc[-1] = dk.data["labels_mean"][label] df[f"{label}_std"].iloc[-1] = dk.data["labels_std"][label] # outlier indicators df["do_predict"].iloc[-1] = do_preds[-1] if self.freqai_info["feature_parameters"].get("DI_threshold", 0) > 0: df["DI_values"].iloc[-1] = dk.DI_values[-1] # extra values the user added within custom prediction model if dk.data['extra_returns_per_train']: rets = dk.data['extra_returns_per_train'] for return_str in rets: df[return_str].iloc[-1] = rets[return_str] # this logic carries users between version without needing to # change their identifier if 'close_price' not in df.columns: df['close_price'] = np.nan df['date_pred'] = np.nan df['close_price'].iloc[-1] = strat_df['close'].iloc[-1] df['date_pred'].iloc[-1] = strat_df['date'].iloc[-1] self.model_return_values[pair] = df.tail(len_df).reset_index(drop=True) def attach_return_values_to_return_dataframe( self, pair: str, dataframe: DataFrame) -> DataFrame: """ Attach the return values to the strat dataframe :param dataframe: DataFrame = strategy dataframe :return: DataFrame = strat dataframe with return values attached """ df = self.model_return_values[pair] to_keep = [col for col in dataframe.columns if not col.startswith("&")] dataframe = pd.concat([dataframe[to_keep], df], axis=1) return dataframe def return_null_values_to_strategy(self, dataframe: DataFrame, dk: FreqaiDataKitchen) -> None: """ Build 0 filled dataframe to return to strategy """ dk.find_features(dataframe) dk.find_labels(dataframe) full_labels = dk.label_list + dk.unique_class_list for label in full_labels: dataframe[label] = 0 dataframe[f"{label}_mean"] = 0 dataframe[f"{label}_std"] = 0 dataframe["do_predict"] = 0 if self.freqai_info["feature_parameters"].get("DI_threshold", 0) > 0: dataframe["DI_values"] = 0 if dk.data['extra_returns_per_train']: rets = dk.data['extra_returns_per_train'] for return_str in rets: dataframe[return_str] = 0 dk.return_dataframe = dataframe def purge_old_models(self) -> None: num_keep = self.freqai_info["purge_old_models"] if not num_keep: return elif type(num_keep) == bool: num_keep = 2 model_folders = [x for x in self.full_path.iterdir() if x.is_dir()] pattern = re.compile(r"sub-train-(\w+)_(\d{10})") delete_dict: Dict[str, Any] = {} for dir in model_folders: result = pattern.match(str(dir.name)) if result is None: continue coin = result.group(1) timestamp = result.group(2) if coin not in delete_dict: delete_dict[coin] = {} delete_dict[coin]["num_folders"] = 1 delete_dict[coin]["timestamps"] = {int(timestamp): dir} else: delete_dict[coin]["num_folders"] += 1 delete_dict[coin]["timestamps"][int(timestamp)] = dir for coin in delete_dict: if delete_dict[coin]["num_folders"] > num_keep: sorted_dict = collections.OrderedDict( sorted(delete_dict[coin]["timestamps"].items()) ) num_delete = len(sorted_dict) - num_keep deleted = 0 for k, v in sorted_dict.items(): if deleted >= num_delete: break logger.info(f"Freqai purging old model file {v}") shutil.rmtree(v) deleted += 1 def save_metadata(self, dk: FreqaiDataKitchen) -> None: """ Saves only metadata for backtesting studies if user prefers not to save model data. This saves tremendous amounts of space for users generating huge studies. This is only active when `save_backtest_models`: false (not default) """ if not dk.data_path.is_dir(): dk.data_path.mkdir(parents=True, exist_ok=True) save_path = Path(dk.data_path) dk.data["data_path"] = str(dk.data_path) dk.data["model_filename"] = str(dk.model_filename) dk.data["training_features_list"] = list(dk.data_dictionary["train_features"].columns) dk.data["label_list"] = dk.label_list with (save_path / f"{dk.model_filename}_metadata.json").open("w") as fp: rapidjson.dump(dk.data, fp, default=self.np_encoder, number_mode=rapidjson.NM_NATIVE) return def save_data(self, model: Any, coin: str, dk: FreqaiDataKitchen) -> None: """ Saves all data associated with a model for a single sub-train time range :param model: User trained model which can be reused for inferencing to generate predictions """ if not dk.data_path.is_dir(): dk.data_path.mkdir(parents=True, exist_ok=True) save_path = Path(dk.data_path) # Save the trained model if self.model_type == 'joblib': dump(model, save_path / f"{dk.model_filename}_model.joblib") elif self.model_type == 'keras': model.save(save_path / f"{dk.model_filename}_model.h5") elif 'stable_baselines' in self.model_type or 'sb3_contrib' == self.model_type: model.save(save_path / f"{dk.model_filename}_model.zip") if dk.svm_model is not None: dump(dk.svm_model, save_path / f"{dk.model_filename}_svm_model.joblib") dk.data["data_path"] = str(dk.data_path) dk.data["model_filename"] = str(dk.model_filename) dk.data["training_features_list"] = dk.training_features_list dk.data["label_list"] = dk.label_list # store the metadata with (save_path / f"{dk.model_filename}_metadata.json").open("w") as fp: rapidjson.dump(dk.data, fp, default=self.np_encoder, number_mode=rapidjson.NM_NATIVE) # save the train data to file so we can check preds for area of applicability later dk.data_dictionary["train_features"].to_pickle( save_path / f"{dk.model_filename}_trained_df.pkl" ) dk.data_dictionary["train_dates"].to_pickle( save_path / f"{dk.model_filename}_trained_dates_df.pkl" ) if self.freqai_info["feature_parameters"].get("principal_component_analysis"): cloudpickle.dump( dk.pca, (dk.data_path / f"{dk.model_filename}_pca_object.pkl").open("wb") ) self.model_dictionary[coin] = model self.pair_dict[coin]["model_filename"] = dk.model_filename self.pair_dict[coin]["data_path"] = str(dk.data_path) if coin not in self.meta_data_dictionary: self.meta_data_dictionary[coin] = {} self.meta_data_dictionary[coin]["train_df"] = dk.data_dictionary["train_features"] self.meta_data_dictionary[coin]["meta_data"] = dk.data self.save_drawer_to_disk() return def load_metadata(self, dk: FreqaiDataKitchen) -> None: """ Load only metadata into datakitchen to increase performance during presaved backtesting (prediction file loading). """ with (dk.data_path / f"{dk.model_filename}_metadata.json").open("r") as fp: dk.data = rapidjson.load(fp, number_mode=rapidjson.NM_NATIVE) dk.training_features_list = dk.data["training_features_list"] dk.label_list = dk.data["label_list"] def load_data(self, coin: str, dk: FreqaiDataKitchen) -> Any: """ loads all data required to make a prediction on a sub-train time range :returns: :model: User trained model which can be inferenced for new predictions """ if not self.pair_dict[coin]["model_filename"]: return None if dk.live: dk.model_filename = self.pair_dict[coin]["model_filename"] dk.data_path = Path(self.pair_dict[coin]["data_path"]) if coin in self.meta_data_dictionary: dk.data = self.meta_data_dictionary[coin]["meta_data"] dk.data_dictionary["train_features"] = self.meta_data_dictionary[coin]["train_df"] else: with (dk.data_path / f"{dk.model_filename}_metadata.json").open("r") as fp: dk.data = rapidjson.load(fp, number_mode=rapidjson.NM_NATIVE) dk.data_dictionary["train_features"] = pd.read_pickle( dk.data_path / f"{dk.model_filename}_trained_df.pkl" ) dk.training_features_list = dk.data["training_features_list"] dk.label_list = dk.data["label_list"] # try to access model in memory instead of loading object from disk to save time if dk.live and coin in self.model_dictionary: model = self.model_dictionary[coin] elif self.model_type == 'joblib': model = load(dk.data_path / f"{dk.model_filename}_model.joblib") elif self.model_type == 'keras': from tensorflow import keras model = keras.models.load_model(dk.data_path / f"{dk.model_filename}_model.h5") elif 'stable_baselines' in self.model_type or 'sb3_contrib' == self.model_type: mod = importlib.import_module( self.model_type, self.freqai_info['rl_config']['model_type']) MODELCLASS = getattr(mod, self.freqai_info['rl_config']['model_type']) model = MODELCLASS.load(dk.data_path / f"{dk.model_filename}_model") if Path(dk.data_path / f"{dk.model_filename}_svm_model.joblib").is_file(): dk.svm_model = load(dk.data_path / f"{dk.model_filename}_svm_model.joblib") if not model: raise OperationalException( f"Unable to load model, ensure model exists at " f"{dk.data_path} " ) # load it into ram if it was loaded from disk if coin not in self.model_dictionary: self.model_dictionary[coin] = model if self.config["freqai"]["feature_parameters"]["principal_component_analysis"]: dk.pca = cloudpickle.load( (dk.data_path / f"{dk.model_filename}_pca_object.pkl").open("rb") ) return model def update_historic_data(self, strategy: IStrategy, dk: FreqaiDataKitchen) -> None: """ Append new candles to our stores historic data (in memory) so that we do not need to load candle history from disk and we dont need to pinging exchange multiple times for the same candle. :param dataframe: DataFrame = strategy provided dataframe """ feat_params = self.freqai_info["feature_parameters"] with self.history_lock: history_data = self.historic_data for pair in dk.all_pairs: for tf in feat_params.get("include_timeframes"): hist_df = history_data[pair][tf] # check if newest candle is already appended df_dp = strategy.dp.get_pair_dataframe(pair, tf) if len(df_dp.index) == 0: continue if str(hist_df.iloc[-1]["date"]) == str( df_dp.iloc[-1:]["date"].iloc[-1] ): continue try: index = ( df_dp.loc[ df_dp["date"] == hist_df.iloc[-1]["date"] ].index[0] + 1 ) except IndexError: if hist_df.iloc[-1]['date'] < df_dp['date'].iloc[0]: raise OperationalException("In memory historical data is older than " f"oldest DataProvider candle for {pair} on " f"timeframe {tf}") else: index = -1 logger.warning( f"No common dates in historical data and dataprovider for {pair}. " f"Appending latest dataprovider candle to historical data " "but please be aware that there is likely a gap in the historical " "data. \n" f"Historical data ends at {hist_df.iloc[-1]['date']} " f"while dataprovider starts at {df_dp['date'].iloc[0]} and" f"ends at {df_dp['date'].iloc[0]}." ) history_data[pair][tf] = pd.concat( [ hist_df, df_dp.iloc[index:], ], ignore_index=True, axis=0, ) self.current_candle = history_data[dk.pair][self.config['timeframe']].iloc[-1]['date'] def load_all_pair_histories(self, timerange: TimeRange, dk: FreqaiDataKitchen) -> None: """ Load pair histories for all whitelist and corr_pairlist pairs. Only called once upon startup of bot. :param timerange: TimeRange = full timerange required to populate all indicators for training according to user defined train_period_days """ history_data = self.historic_data for pair in dk.all_pairs: if pair not in history_data: history_data[pair] = {} for tf in self.freqai_info["feature_parameters"].get("include_timeframes"): history_data[pair][tf] = load_pair_history( datadir=self.config["datadir"], timeframe=tf, pair=pair, timerange=timerange, data_format=self.config.get("dataformat_ohlcv", "json"), candle_type=self.config.get("trading_mode", "spot"), ) def get_base_and_corr_dataframes( self, timerange: TimeRange, pair: str, dk: FreqaiDataKitchen ) -> Tuple[Dict[Any, Any], Dict[Any, Any]]: """ Searches through our historic_data in memory and returns the dataframes relevant to the present pair. :param timerange: TimeRange = full timerange required to populate all indicators for training according to user defined train_period_days :param metadata: dict = strategy furnished pair metadata """ with self.history_lock: corr_dataframes: Dict[Any, Any] = {} base_dataframes: Dict[Any, Any] = {} historic_data = self.historic_data pairs = self.freqai_info["feature_parameters"].get( "include_corr_pairlist", [] ) for tf in self.freqai_info["feature_parameters"].get("include_timeframes"): base_dataframes[tf] = dk.slice_dataframe( timerange, historic_data[pair][tf]).reset_index(drop=True) if pairs: for p in pairs: if pair in p: continue # dont repeat anything from whitelist if p not in corr_dataframes: corr_dataframes[p] = {} corr_dataframes[p][tf] = dk.slice_dataframe( timerange, historic_data[p][tf] ).reset_index(drop=True) return corr_dataframes, base_dataframes def get_timerange_from_live_historic_predictions(self) -> TimeRange: """ Returns timerange information based on historic predictions file :return: timerange calculated from saved live data """ if not self.historic_predictions_path.is_file(): raise OperationalException( 'Historic predictions not found. Historic predictions data is required ' 'to run backtest with the freqai-backtest-live-models option ' ) self.load_historic_predictions_from_disk() all_pairs_end_dates = [] for pair in self.historic_predictions: pair_historic_data = self.historic_predictions[pair] all_pairs_end_dates.append(pair_historic_data.date_pred.max()) global_metadata = self.load_global_metadata_from_disk() start_date = datetime.fromtimestamp(int(global_metadata["start_dry_live_date"])) end_date = max(all_pairs_end_dates) # add 1 day to string timerange to ensure BT module will load all dataframe data end_date = end_date + timedelta(days=1) backtesting_timerange = TimeRange( 'date', 'date', int(start_date.timestamp()), int(end_date.timestamp()) ) return backtesting_timerange