stable/freqtrade/freqai/freqai_interface.py

955 lines
41 KiB
Python
Raw Normal View History

2022-05-04 15:53:40 +00:00
import logging
import threading
import time
2022-05-04 15:53:40 +00:00
from abc import ABC, abstractmethod
from collections import deque
from datetime import datetime, timezone
from pathlib import Path
2022-11-12 09:54:34 +00:00
from typing import Any, Dict, List, Literal, Optional, Tuple
2022-08-15 04:53:02 +00:00
import numpy as np
import pandas as pd
2022-09-28 22:10:18 +00:00
import psutil
from numpy.typing import NDArray
from pandas import DataFrame
2022-08-15 04:53:02 +00:00
from freqtrade.configuration import TimeRange
2022-11-10 17:26:14 +00:00
from freqtrade.constants import Config
from freqtrade.data.dataprovider import DataProvider
2022-05-06 14:20:52 +00:00
from freqtrade.enums import RunMode
from freqtrade.exceptions import OperationalException
2022-08-15 04:53:02 +00:00
from freqtrade.exchange import timeframe_to_seconds
from freqtrade.freqai.data_drawer import FreqaiDataDrawer
from freqtrade.freqai.data_kitchen import FreqaiDataKitchen
from freqtrade.freqai.utils import plot_feature_importance, record_params
from freqtrade.strategy.interface import IStrategy
pd.options.mode.chained_assignment = None
2022-05-04 15:53:40 +00:00
logger = logging.getLogger(__name__)
class IFreqaiModel(ABC):
"""
Class containing all tools for training and prediction in the strategy.
Base*PredictionModels inherit from this class.
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
2022-08-14 18:24:29 +00:00
Juha Nykänen @suikula, Wagner Costa @wagnercosta, Johan Vlugt @Jooopieeert
"""
2022-09-18 11:20:36 +00:00
def __init__(self, config: Config) -> None:
self.config = config
self.assert_config(self.config)
self.freqai_info: Dict[str, Any] = config["freqai"]
self.data_split_parameters: Dict[str, Any] = config.get("freqai", {}).get(
"data_split_parameters", {})
self.model_training_parameters: Dict[str, Any] = config.get("freqai", {}).get(
"model_training_parameters", {})
self.identifier: str = self.freqai_info.get("identifier", "no_id_provided")
self.retrain = False
self.first = True
self.set_full_path()
self.follow_mode: bool = self.freqai_info.get("follow_mode", False)
self.save_backtest_models: bool = self.freqai_info.get("save_backtest_models", True)
if self.save_backtest_models:
logger.info('Backtesting module configured to save all models.')
2022-11-12 17:37:23 +00:00
2022-07-02 16:09:38 +00:00
self.dd = FreqaiDataDrawer(Path(self.full_path), self.config, self.follow_mode)
# set current candle to arbitrary historical date
self.current_candle: datetime = datetime.fromtimestamp(637887600, tz=timezone.utc)
self.dd.current_candle = self.current_candle
self.scanning = False
2022-08-18 17:15:29 +00:00
self.ft_params = self.freqai_info["feature_parameters"]
2022-10-30 12:28:01 +00:00
self.corr_pairlist: List[str] = self.ft_params.get("include_corr_pairlist", [])
self.keras: bool = self.freqai_info.get("keras", False)
2022-08-18 17:15:29 +00:00
if self.keras and self.ft_params.get("DI_threshold", 0):
self.ft_params["DI_threshold"] = 0
logger.warning("DI threshold is not configured for Keras models yet. Deactivating.")
2022-11-03 20:35:12 +00:00
self.CONV_WIDTH = self.freqai_info.get('conv_width', 1)
2022-08-18 17:15:29 +00:00
if self.ft_params.get("inlier_metric_window", 0):
self.CONV_WIDTH = self.ft_params.get("inlier_metric_window", 0) * 2
self.pair_it = 0
2022-08-22 11:30:30 +00:00
self.pair_it_train = 0
self.total_pairs = len(self.config.get("exchange", {}).get("pair_whitelist"))
self.train_queue = self._set_train_queue()
self.inference_time: float = 0
2022-08-22 11:30:30 +00:00
self.train_time: float = 0
self.begin_time: float = 0
2022-08-22 11:30:30 +00:00
self.begin_time_train: float = 0
self.base_tf_seconds = timeframe_to_seconds(self.config['timeframe'])
self.continual_learning = self.freqai_info.get('continual_learning', False)
self.plot_features = self.ft_params.get("plot_feature_importances", 0)
self.corr_dataframes: Dict[str, DataFrame] = {}
# get_corr_dataframes is controlling the caching of corr_dataframes
# for improved performance. Careful with this boolean.
self.get_corr_dataframes: bool = True
2022-09-03 19:24:14 +00:00
self._threads: List[threading.Thread] = []
self._stop_event = threading.Event()
2022-11-29 23:53:35 +00:00
self.metadata: Dict[str, Any] = self.dd.load_global_metadata_from_disk()
self.data_provider: Optional[DataProvider] = None
2022-09-29 12:01:22 +00:00
self.max_system_threads = max(int(psutil.cpu_count() * 2 - 2), 1)
2022-09-03 19:24:14 +00:00
record_params(config, self.full_path)
2022-09-05 20:43:28 +00:00
def __getstate__(self):
"""
Return an empty state to be pickled in hyperopt
2022-09-05 20:43:28 +00:00
"""
return ({})
2022-09-05 20:43:28 +00:00
2022-09-18 11:20:36 +00:00
def assert_config(self, config: Config) -> None:
2022-07-03 08:59:38 +00:00
if not config.get("freqai", {}):
raise OperationalException("No freqai parameters found in configuration file.")
def start(self, dataframe: DataFrame, metadata: dict, strategy: IStrategy) -> DataFrame:
"""
Entry point to the FreqaiModel from a specific pair, it will train a new model if
2022-05-15 14:25:08 +00:00
necessary before making the prediction.
2022-07-24 14:54:39 +00:00
:param dataframe: Full dataframe coming from strategy - it contains entire
backtesting timerange + additional historical data necessary to train
the model.
2022-07-24 14:54:39 +00:00
:param metadata: pair metadata coming from strategy.
:param strategy: Strategy to train on
"""
self.live = strategy.dp.runmode in (RunMode.DRY_RUN, RunMode.LIVE)
2022-07-02 16:09:38 +00:00
self.dd.set_pair_dict_info(metadata)
self.data_provider = strategy.dp
if self.live:
self.inference_timer('start')
self.dk = FreqaiDataKitchen(self.config, self.live, metadata["pair"])
2022-07-02 16:09:38 +00:00
dk = self.start_live(dataframe, metadata, strategy, self.dk)
dataframe = dk.remove_features_from_df(dk.return_dataframe)
# For backtesting, each pair enters and then gets trained for each window along the
# sliding window defined by "train_period_days" (training window) and "live_retrain_hours"
# (backtest window, i.e. window immediately following the training window).
# FreqAI slides the window and sequentially builds the backtesting results before returning
# the concatenated results for the full backtesting period back to the strategy.
elif not self.follow_mode:
self.dk = FreqaiDataKitchen(self.config, self.live, metadata["pair"])
dataframe = self.dk.use_strategy_to_populate_indicators(
strategy, prediction_dataframe=dataframe, pair=metadata["pair"]
)
if not self.config.get("freqai_backtest_live_models", False):
2022-11-29 23:53:35 +00:00
logger.info(f"Training {len(self.dk.training_timeranges)} timeranges")
2022-11-12 17:37:23 +00:00
dk = self.start_backtesting(dataframe, metadata, self.dk)
dataframe = dk.remove_features_from_df(dk.return_dataframe)
else:
2022-11-29 23:53:35 +00:00
logger.info(
"Backtesting using historic predictions (live models)")
dk = self.start_backtesting_from_historic_predictions(
2022-11-12 17:37:23 +00:00
dataframe, metadata, self.dk)
dataframe = dk.return_dataframe
2022-08-22 11:30:30 +00:00
self.clean_up()
if self.live:
self.inference_timer('stop', metadata["pair"])
2022-11-12 17:37:23 +00:00
return dataframe
2022-08-22 11:30:30 +00:00
def clean_up(self):
"""
Objects that should be handled by GC already between coins, but
are explicitly shown here to help demonstrate the non-persistence of these
objects.
"""
self.model = None
self.dk = None
2022-09-28 03:06:05 +00:00
def _on_stop(self):
"""
Callback for Subclasses to override to include logic for shutting down resources
when SIGINT is sent.
"""
return
2022-09-03 19:24:14 +00:00
def shutdown(self):
"""
Cleans up threads on Shutdown, set stop event. Join threads to wait
for current training iteration.
"""
logger.info("Stopping FreqAI")
self._stop_event.set()
self.data_provider = None
2022-09-28 03:06:05 +00:00
self._on_stop()
2022-09-03 19:24:14 +00:00
logger.info("Waiting on Training iteration")
for _thread in self._threads:
_thread.join()
def start_scanning(self, *args, **kwargs) -> None:
"""
Start `self._start_scanning` in a separate thread
"""
_thread = threading.Thread(target=self._start_scanning, args=args, kwargs=kwargs)
self._threads.append(_thread)
_thread.start()
def _start_scanning(self, strategy: IStrategy) -> None:
"""
Function designed to constantly scan pairs for retraining on a separate thread (intracandle)
to improve model youth. This function is agnostic to data preparation/collection/storage,
2022-07-02 16:09:38 +00:00
it simply trains on what ever data is available in the self.dd.
:param strategy: IStrategy = The user defined strategy class
"""
2022-09-03 19:24:14 +00:00
while not self._stop_event.is_set():
time.sleep(1)
pair = self.train_queue[0]
# ensure pair is avaialble in dp
if pair not in strategy.dp.current_whitelist():
self.train_queue.popleft()
logger.warning(f'{pair} not in current whitelist, removing from train queue.')
continue
(_, trained_timestamp, _) = self.dd.get_pair_dict_info(pair)
dk = FreqaiDataKitchen(self.config, self.live, pair)
(
retrain,
new_trained_timerange,
data_load_timerange,
) = dk.check_if_new_training_required(trained_timestamp)
if retrain:
self.train_timer('start')
2022-10-11 17:05:46 +00:00
dk.set_paths(pair, new_trained_timerange.stopts)
try:
self.extract_data_and_train_model(
new_trained_timerange, pair, strategy, dk, data_load_timerange
)
except Exception as msg:
2022-10-07 14:05:41 +00:00
logger.warning(f"Training {pair} raised exception {msg.__class__.__name__}. "
f"Message: {msg}, skipping.")
self.train_timer('stop', pair)
# only rotate the queue after the first has been trained.
self.train_queue.rotate(-1)
self.dd.save_historic_predictions_to_disk()
if self.freqai_info.get('write_metrics_to_disk', False):
self.dd.save_metric_tracker_to_disk()
2022-07-03 08:59:38 +00:00
def start_backtesting(
self, dataframe: DataFrame, metadata: dict, dk: FreqaiDataKitchen
) -> FreqaiDataKitchen:
"""
The main broad execution for backtesting. For backtesting, each pair enters and then gets
trained for each window along the sliding window defined by "train_period_days"
(training window) and "backtest_period_days" (backtest window, i.e. window immediately
following the training window). FreqAI slides the window and sequentially builds
the backtesting results before returning the concatenated results for the full
backtesting period back to the strategy.
:param dataframe: DataFrame = strategy passed dataframe
:param metadata: Dict = pair metadata
:param dk: FreqaiDataKitchen = Data management/analysis tool associated to present pair only
:return:
FreqaiDataKitchen = Data management/analysis tool associated to present pair only
"""
self.pair_it += 1
train_it = 0
# Loop enforcing the sliding window training/backtesting paradigm
# tr_train is the training time range e.g. 1 historical month
# tr_backtest is the backtesting time range e.g. the week directly
# following tr_train. Both of these windows slide through the
# entire backtest
for tr_train, tr_backtest in zip(dk.training_timeranges, dk.backtesting_timeranges):
pair = metadata["pair"]
(_, _, _) = self.dd.get_pair_dict_info(pair)
train_it += 1
total_trains = len(dk.backtesting_timeranges)
self.training_timerange = tr_train
dataframe_train = dk.slice_dataframe(tr_train, dataframe)
dataframe_backtest = dk.slice_dataframe(tr_backtest, dataframe)
if not self.ensure_data_exists(dataframe_backtest, tr_backtest, pair):
continue
self.log_backtesting_progress(tr_train, pair, train_it, total_trains)
timestamp_model_id = int(tr_train.stopts)
if dk.backtest_live_models:
timestamp_model_id = int(tr_backtest.startts)
dk.set_paths(pair, timestamp_model_id)
dk.set_new_model_names(pair, timestamp_model_id)
if dk.check_if_backtest_prediction_is_valid(len(dataframe_backtest)):
self.dd.load_metadata(dk)
dk.find_features(dataframe_train)
self.check_if_feature_list_matches_strategy(dk)
append_df = dk.get_backtesting_prediction()
2022-08-31 14:23:48 +00:00
dk.append_predictions(append_df)
else:
if not self.model_exists(dk):
2022-08-31 14:23:48 +00:00
dk.find_features(dataframe_train)
dk.find_labels(dataframe_train)
2022-09-24 08:34:14 +00:00
self.model = self.train(dataframe_train, pair, dk)
self.dd.pair_dict[pair]["trained_timestamp"] = int(
tr_train.stopts)
if self.plot_features:
plot_feature_importance(self.model, pair, dk, self.plot_features)
if self.save_backtest_models:
logger.info('Saving backtest model to disk.')
2022-09-24 08:34:14 +00:00
self.dd.save_data(self.model, pair, dk)
else:
logger.info('Saving metadata to disk.')
2022-09-25 18:22:19 +00:00
self.dd.save_metadata(dk)
2022-08-31 14:23:48 +00:00
else:
2022-09-24 08:34:14 +00:00
self.model = self.dd.load_data(pair, dk)
2022-08-31 14:23:48 +00:00
pred_df, do_preds = self.predict(dataframe_backtest, dk)
append_df = dk.get_predictions_to_append(pred_df, do_preds, dataframe_backtest)
2022-08-31 14:23:48 +00:00
dk.append_predictions(append_df)
dk.save_backtesting_prediction(append_df)
self.backtesting_fit_live_predictions(dk)
dk.fill_predictions(dataframe)
2022-07-02 16:09:38 +00:00
return dk
2022-07-03 08:59:38 +00:00
def start_live(
self, dataframe: DataFrame, metadata: dict, strategy: IStrategy, dk: FreqaiDataKitchen
) -> FreqaiDataKitchen:
"""
The main broad execution for dry/live. This function will check if a retraining should be
performed, and if so, retrain and reset the model.
:param dataframe: DataFrame = strategy passed dataframe
:param metadata: Dict = pair metadata
:param strategy: IStrategy = currently employed strategy
dk: FreqaiDataKitchen = Data management/analysis tool associated to present pair only
:returns:
dk: FreqaiDataKitchen = Data management/analysis tool associated to present pair only
"""
# update follower
if self.follow_mode:
2022-07-02 16:09:38 +00:00
self.dd.update_follower_metadata()
# get the model metadata associated with the current pair
(_, trained_timestamp, return_null_array) = self.dd.get_pair_dict_info(metadata["pair"])
# if the metadata doesn't exist, the follower returns null arrays to strategy
if self.follow_mode and return_null_array:
2022-07-03 08:59:38 +00:00
logger.info("Returning null array from follower to strategy")
2022-07-02 16:09:38 +00:00
self.dd.return_null_values_to_strategy(dataframe, dk)
return dk
# append the historic data once per round
2022-07-02 16:09:38 +00:00
if self.dd.historic_data:
self.dd.update_historic_data(strategy, dk)
logger.debug(f'Updating historic data on pair {metadata["pair"]}')
self.track_current_candle()
if not self.follow_mode:
2022-07-03 08:59:38 +00:00
(_, new_trained_timerange, data_load_timerange) = dk.check_if_new_training_required(
trained_timestamp
)
dk.set_paths(metadata["pair"], new_trained_timerange.stopts)
2022-08-17 13:22:48 +00:00
# load candle history into memory if it is not yet.
2022-07-02 16:09:38 +00:00
if not self.dd.historic_data:
self.dd.load_all_pair_histories(data_load_timerange, dk)
2022-06-15 22:21:15 +00:00
if not self.scanning:
self.scanning = True
self.start_scanning(strategy)
elif self.follow_mode:
2022-07-03 08:59:38 +00:00
dk.set_paths(metadata["pair"], trained_timestamp)
logger.info(
2022-07-25 07:24:40 +00:00
"FreqAI instance set to follow_mode, finding existing pair "
2022-07-03 08:59:38 +00:00
f"using { self.identifier }"
)
# load the model and associated data into the data kitchen
self.model = self.dd.load_data(metadata["pair"], dk)
dataframe = dk.use_strategy_to_populate_indicators(
strategy, prediction_dataframe=dataframe, pair=metadata["pair"],
do_corr_pairs=self.get_corr_dataframes
)
2022-06-15 22:21:15 +00:00
if not self.model:
logger.warning(
f"No model ready for {metadata['pair']}, returning null values to strategy."
)
2022-07-02 16:09:38 +00:00
self.dd.return_null_values_to_strategy(dataframe, dk)
return dk
if self.corr_pairlist:
dataframe = self.cache_corr_pairlist_dfs(dataframe, dk)
dk.find_labels(dataframe)
2022-07-03 08:59:38 +00:00
self.build_strategy_return_arrays(dataframe, dk, metadata["pair"], trained_timestamp)
2022-07-02 16:09:38 +00:00
return dk
2022-07-03 08:59:38 +00:00
def build_strategy_return_arrays(
self, dataframe: DataFrame, dk: FreqaiDataKitchen, pair: str, trained_timestamp: int
) -> None:
# hold the historical predictions in memory so we are sending back
# correct array to strategy
2022-07-02 16:09:38 +00:00
if pair not in self.dd.model_return_values:
# first predictions are made on entire historical candle set coming from strategy. This
# allows FreqUI to show full return values.
2022-07-02 16:09:38 +00:00
pred_df, do_preds = self.predict(dataframe, dk)
2022-08-10 13:16:50 +00:00
if pair not in self.dd.historic_predictions:
self.set_initial_historic_predictions(pred_df, dk, pair, dataframe)
2022-08-12 14:12:28 +00:00
self.dd.set_initial_return_values(pair, pred_df)
2022-07-02 16:09:38 +00:00
dk.return_dataframe = self.dd.attach_return_values_to_return_dataframe(pair, dataframe)
return
2022-07-02 16:09:38 +00:00
elif self.dk.check_if_model_expired(trained_timestamp):
pred_df = DataFrame(np.zeros((2, len(dk.label_list))), columns=dk.label_list)
do_preds = np.ones(2, dtype=np.int_) * 2
dk.DI_values = np.zeros(2)
2022-07-03 08:59:38 +00:00
logger.warning(
f"Model expired for {pair}, returning null values to strategy. Strategy "
2022-07-03 08:59:38 +00:00
"construction should take care to consider this event with "
"prediction == 0 and do_predict == 2"
)
else:
# remaining predictions are made only on the most recent candles for performance and
# historical accuracy reasons.
2022-07-02 16:09:38 +00:00
pred_df, do_preds = self.predict(dataframe.iloc[-self.CONV_WIDTH:], dk, first=False)
2022-08-10 13:16:50 +00:00
if self.freqai_info.get('fit_live_predictions_candles', 0) and self.live:
self.fit_live_predictions(dk, pair)
self.dd.append_model_predictions(pair, pred_df, do_preds, dk, dataframe)
2022-07-02 16:09:38 +00:00
dk.return_dataframe = self.dd.attach_return_values_to_return_dataframe(pair, dataframe)
return
2022-07-03 08:59:38 +00:00
def check_if_feature_list_matches_strategy(
self, dk: FreqaiDataKitchen
2022-07-03 08:59:38 +00:00
) -> None:
"""
Ensure user is passing the proper feature set if they are reusing an `identifier` pointing
to a folder holding existing models.
:param dataframe: DataFrame = strategy provided dataframe
:param dk: FreqaiDataKitchen = non-persistent data container/analyzer for
current coin/bot loop
"""
2022-07-03 08:59:38 +00:00
if "training_features_list_raw" in dk.data:
feature_list = dk.data["training_features_list_raw"]
else:
feature_list = dk.data['training_features_list']
2022-09-30 22:22:05 +00:00
2022-07-02 16:09:38 +00:00
if dk.training_features_list != feature_list:
2022-07-03 08:59:38 +00:00
raise OperationalException(
"Trying to access pretrained model with `identifier` "
"but found different features furnished by current strategy."
"Change `identifier` to train from scratch, or ensure the"
2022-07-03 08:59:38 +00:00
"strategy is furnishing the same features as the pretrained"
"model. In case of --strategy-list, please be aware that FreqAI "
"requires all strategies to maintain identical "
"populate_any_indicator() functions"
2022-07-03 08:59:38 +00:00
)
2022-07-02 16:09:38 +00:00
def data_cleaning_train(self, dk: FreqaiDataKitchen) -> None:
"""
2022-08-19 16:35:24 +00:00
Base data cleaning method for train.
Functions here improve/modify the input data by identifying outliers,
computing additional metrics, adding noise, reducing dimensionality etc.
"""
2022-08-18 17:15:29 +00:00
ft_params = self.freqai_info["feature_parameters"]
if ft_params.get('inlier_metric_window', 0):
dk.compute_inlier_metric(set_='train')
if self.freqai_info["data_split_parameters"]["test_size"] > 0:
dk.compute_inlier_metric(set_='test')
2022-08-18 17:15:29 +00:00
if ft_params.get(
"principal_component_analysis", False
):
2022-07-02 16:09:38 +00:00
dk.principal_component_analysis()
2022-08-18 17:15:29 +00:00
if ft_params.get("use_SVM_to_remove_outliers", False):
2022-07-02 16:09:38 +00:00
dk.use_SVM_to_remove_outliers(predict=False)
2022-08-18 17:15:29 +00:00
if ft_params.get("DI_threshold", 0):
2022-07-02 16:09:38 +00:00
dk.data["avg_mean_dist"] = dk.compute_distances()
2022-08-18 17:15:29 +00:00
if ft_params.get("use_DBSCAN_to_remove_outliers", False):
if dk.pair in self.dd.old_DBSCAN_eps:
eps = self.dd.old_DBSCAN_eps[dk.pair]
else:
eps = None
dk.use_DBSCAN_to_remove_outliers(predict=False, eps=eps)
self.dd.old_DBSCAN_eps[dk.pair] = dk.data['DBSCAN_eps']
2022-08-19 16:35:24 +00:00
if self.freqai_info["feature_parameters"].get('noise_standard_deviation', 0):
dk.add_noise_to_training_features()
2022-10-01 12:18:46 +00:00
def data_cleaning_predict(self, dk: FreqaiDataKitchen) -> None:
"""
Base data cleaning method for predict.
2022-08-19 16:35:24 +00:00
Functions here are complementary to the functions of data_cleaning_train.
"""
2022-08-18 17:15:29 +00:00
ft_params = self.freqai_info["feature_parameters"]
2022-10-01 12:18:46 +00:00
# ensure user is feeding the correct indicators to the model
self.check_if_feature_list_matches_strategy(dk)
2022-08-18 17:15:29 +00:00
if ft_params.get('inlier_metric_window', 0):
dk.compute_inlier_metric(set_='predict')
if ft_params.get(
"principal_component_analysis", False
):
dk.pca_transform(dk.data_dictionary['prediction_features'])
2022-08-18 17:15:29 +00:00
if ft_params.get("use_SVM_to_remove_outliers", False):
2022-07-02 16:09:38 +00:00
dk.use_SVM_to_remove_outliers(predict=True)
2022-08-18 17:15:29 +00:00
if ft_params.get("DI_threshold", 0):
2022-07-02 16:09:38 +00:00
dk.check_if_pred_in_training_spaces()
2022-08-18 17:15:29 +00:00
if ft_params.get("use_DBSCAN_to_remove_outliers", False):
dk.use_DBSCAN_to_remove_outliers(predict=True)
def model_exists(self, dk: FreqaiDataKitchen) -> bool:
"""
Given a pair and path, check if a model already exists
:param pair: pair e.g. BTC/USD
:param path: path to model
:return:
:boolean: whether the model file exists or not.
"""
path_to_modelfile = Path(dk.data_path / f"{dk.model_filename}_model.joblib")
file_exists = path_to_modelfile.is_file()
if file_exists:
2022-07-02 16:09:38 +00:00
logger.info("Found model at %s", dk.data_path / dk.model_filename)
else:
2022-07-02 16:09:38 +00:00
logger.info("Could not find model at %s", dk.data_path / dk.model_filename)
return file_exists
def set_full_path(self) -> None:
"""
Creates and sets the full path for the identifier
"""
2022-07-03 08:59:38 +00:00
self.full_path = Path(
2022-10-22 01:53:33 +00:00
self.config["user_data_dir"] / "models" / f"{self.identifier}"
2022-07-03 08:59:38 +00:00
)
self.full_path.mkdir(parents=True, exist_ok=True)
def extract_data_and_train_model(
2022-07-03 08:59:38 +00:00
self,
new_trained_timerange: TimeRange,
pair: str,
strategy: IStrategy,
dk: FreqaiDataKitchen,
data_load_timerange: TimeRange,
):
"""
2022-08-22 11:30:30 +00:00
Retrieve data and train model.
:param new_trained_timerange: TimeRange = the timerange to train the model on
:param metadata: dict = strategy provided metadata
:param strategy: IStrategy = user defined strategy object
:param dk: FreqaiDataKitchen = non-persistent data container for current coin/loop
:param data_load_timerange: TimeRange = the amount of data to be loaded
for populate_any_indicators
(larger than new_trained_timerange so that
new_trained_timerange does not contain any NaNs)
"""
corr_dataframes, base_dataframes = self.dd.get_base_and_corr_dataframes(
data_load_timerange, pair, dk
2022-07-03 08:59:38 +00:00
)
unfiltered_dataframe = dk.use_strategy_to_populate_indicators(
strategy, corr_dataframes, base_dataframes, pair
)
2022-07-02 16:09:38 +00:00
unfiltered_dataframe = dk.slice_dataframe(new_trained_timerange, unfiltered_dataframe)
# find the features indicated by strategy and store in datakitchen
dk.find_features(unfiltered_dataframe)
dk.find_labels(unfiltered_dataframe)
2022-07-02 16:09:38 +00:00
model = self.train(unfiltered_dataframe, pair, dk)
2022-07-03 08:59:38 +00:00
self.dd.pair_dict[pair]["trained_timestamp"] = new_trained_timerange.stopts
dk.set_new_model_names(pair, new_trained_timerange.stopts)
self.dd.save_data(model, pair, dk)
2022-06-16 14:12:38 +00:00
if self.plot_features:
plot_feature_importance(model, pair, dk, self.plot_features)
2022-09-17 15:53:43 +00:00
2022-07-03 08:59:38 +00:00
if self.freqai_info.get("purge_old_models", False):
2022-07-02 16:09:38 +00:00
self.dd.purge_old_models()
def set_initial_historic_predictions(
self, pred_df: DataFrame, dk: FreqaiDataKitchen, pair: str, strat_df: DataFrame
) -> None:
"""
This function is called only if the datadrawer failed to load an
existing set of historic predictions. In this case, it builds
the structure and sets fake predictions off the first training
data. After that, FreqAI will append new real predictions to the
set of historic predictions.
These values are used to generate live statistics which can be used
in the strategy for adaptive values. E.g. &*_mean/std are quantities
that can computed based on live predictions from the set of historical
predictions. Those values can be used in the user strategy to better
assess prediction rarity, and thus wait for probabilistically favorable
entries relative to the live historical predictions.
If the user reuses an identifier on a subsequent instance,
this function will not be called. In that case, "real" predictions
will be appended to the loaded set of historic predictions.
2022-10-10 12:15:30 +00:00
:param df: DataFrame = the dataframe containing the training feature data
:param model: Any = A model which was `fit` using a common library such as
catboost or lightgbm
:param dk: FreqaiDataKitchen = object containing methods for data analysis
:param pair: str = current pair
"""
self.dd.historic_predictions[pair] = pred_df
hist_preds_df = self.dd.historic_predictions[pair]
self.set_start_dry_live_date(strat_df)
2022-11-29 23:53:35 +00:00
for label in hist_preds_df.columns:
if hist_preds_df[label].dtype == object:
continue
hist_preds_df[f'{label}_mean'] = 0
hist_preds_df[f'{label}_std'] = 0
hist_preds_df['do_predict'] = 0
if self.freqai_info['feature_parameters'].get('DI_threshold', 0) > 0:
hist_preds_df['DI_values'] = 0
for return_str in dk.data['extra_returns_per_train']:
hist_preds_df[return_str] = dk.data['extra_returns_per_train'][return_str]
hist_preds_df['close_price'] = strat_df['close']
hist_preds_df['date_pred'] = strat_df['date']
# # for keras type models, the conv_window needs to be prepended so
# # viewing is correct in frequi
if self.freqai_info.get('keras', False) or self.ft_params.get('inlier_metric_window', 0):
n_lost_points = self.freqai_info.get('conv_width', 2)
zeros_df = DataFrame(np.zeros((n_lost_points, len(hist_preds_df.columns))),
columns=hist_preds_df.columns)
2022-08-12 14:12:28 +00:00
self.dd.historic_predictions[pair] = pd.concat(
[zeros_df, hist_preds_df], axis=0, ignore_index=True)
def fit_live_predictions(self, dk: FreqaiDataKitchen, pair: str) -> None:
"""
Fit the labels with a gaussian distribution
"""
import scipy as spy
2022-08-10 13:16:50 +00:00
# add classes from classifier label types if used
full_labels = dk.label_list + dk.unique_class_list
num_candles = self.freqai_info.get("fit_live_predictions_candles", 100)
dk.data["labels_mean"], dk.data["labels_std"] = {}, {}
2022-08-10 13:16:50 +00:00
for label in full_labels:
if self.dd.historic_predictions[dk.pair][label].dtype == object:
continue
f = spy.stats.norm.fit(
self.dd.historic_predictions[dk.pair][label].tail(num_candles))
dk.data["labels_mean"][label], dk.data["labels_std"][label] = f[0], f[1]
return
def inference_timer(self, do: Literal['start', 'stop'] = 'start', pair: str = ''):
"""
Timer designed to track the cumulative time spent in FreqAI for one pass through
the whitelist. This will check if the time spent is more than 1/4 the time
of a single candle, and if so, it will warn the user of degraded performance
"""
if do == 'start':
self.pair_it += 1
self.begin_time = time.time()
elif do == 'stop':
end = time.time()
time_spent = (end - self.begin_time)
if self.freqai_info.get('write_metrics_to_disk', False):
self.dd.update_metric_tracker('inference_time', time_spent, pair)
self.inference_time += time_spent
if self.pair_it == self.total_pairs:
logger.info(
f'Total time spent inferencing pairlist {self.inference_time:.2f} seconds')
if self.inference_time > 0.25 * self.base_tf_seconds:
2022-09-03 13:01:28 +00:00
logger.warning("Inference took over 25% of the candle time. Reduce pairlist to"
" avoid blinding open trades and degrading performance.")
self.pair_it = 0
self.inference_time = 0
return
def train_timer(self, do: Literal['start', 'stop'] = 'start', pair: str = ''):
2022-08-22 11:30:30 +00:00
"""
Timer designed to track the cumulative time spent training the full pairlist in
FreqAI.
"""
if do == 'start':
self.pair_it_train += 1
self.begin_time_train = time.time()
elif do == 'stop':
end = time.time()
time_spent = (end - self.begin_time_train)
if self.freqai_info.get('write_metrics_to_disk', False):
self.dd.collect_metrics(time_spent, pair)
self.train_time += time_spent
2022-08-22 11:30:30 +00:00
if self.pair_it_train == self.total_pairs:
logger.info(
f'Total time spent training pairlist {self.train_time:.2f} seconds')
self.pair_it_train = 0
self.train_time = 0
return
2022-09-07 16:58:55 +00:00
def get_init_model(self, pair: str) -> Any:
if pair not in self.dd.model_dictionary or not self.continual_learning:
init_model = None
else:
init_model = self.dd.model_dictionary[pair]
return init_model
def _set_train_queue(self):
"""
Sets train queue from existing train timestamps if they exist
otherwise it sets the train queue based on the provided whitelist.
"""
current_pairlist = self.config.get("exchange", {}).get("pair_whitelist")
if not self.dd.pair_dict:
2022-09-19 17:16:32 +00:00
logger.info('Set fresh train queue from whitelist. '
f'Queue: {current_pairlist}')
return deque(current_pairlist)
best_queue = deque()
pair_dict_sorted = sorted(self.dd.pair_dict.items(),
key=lambda k: k[1]['trained_timestamp'])
for pair in pair_dict_sorted:
if pair[0] in current_pairlist:
2022-09-19 10:47:20 +00:00
best_queue.append(pair[0])
for pair in current_pairlist:
if pair not in best_queue:
best_queue.appendleft(pair)
2022-09-19 17:16:32 +00:00
logger.info('Set existing queue from trained timestamps. '
f'Best approximation queue: {best_queue}')
return best_queue
def cache_corr_pairlist_dfs(self, dataframe: DataFrame, dk: FreqaiDataKitchen) -> DataFrame:
"""
Cache the corr_pairlist dfs to speed up performance for subsequent pairs during the
current candle.
:param dataframe: strategy fed dataframe
:param dk: datakitchen object for current asset
:return: dataframe to attach/extract cached corr_pair dfs to/from.
"""
if self.get_corr_dataframes:
self.corr_dataframes = dk.extract_corr_pair_columns_from_populated_indicators(dataframe)
if not self.corr_dataframes:
logger.warning("Couldn't cache corr_pair dataframes for improved performance. "
"Consider ensuring that the full coin/stake, e.g. XYZ/USD, "
"is included in the column names when you are creating features "
"in `populate_any_indicators()`.")
self.get_corr_dataframes = not bool(self.corr_dataframes)
2022-11-02 19:20:35 +00:00
elif self.corr_dataframes:
dataframe = dk.attach_corr_pair_columns(
dataframe, self.corr_dataframes, dk.pair)
return dataframe
def track_current_candle(self):
"""
Checks if the latest candle appended by the datadrawer is
equivalent to the latest candle seen by FreqAI. If not, it
asks to refresh the cached corr_dfs, and resets the pair
counter.
"""
if self.dd.current_candle > self.current_candle:
self.get_corr_dataframes = True
self.pair_it = 1
self.current_candle = self.dd.current_candle
def ensure_data_exists(self, dataframe_backtest: DataFrame,
tr_backtest: TimeRange, pair: str) -> bool:
"""
Check if the dataframe is empty, if not, report useful information to user.
:param dataframe_backtest: the backtesting dataframe, maybe empty.
:param tr_backtest: current backtesting timerange.
:param pair: current pair
:return: if the data exists or not
"""
if self.config.get("freqai_backtest_live_models", False) and len(dataframe_backtest) == 0:
2022-11-10 17:26:14 +00:00
logger.info(f"No data found for pair {pair} from "
f"from { tr_backtest.start_fmt} to {tr_backtest.stop_fmt}. "
"Probably more than one training within the same candle period.")
2022-11-04 15:41:38 +00:00
return False
return True
def log_backtesting_progress(self, tr_train: TimeRange, pair: str,
train_it: int, total_trains: int):
"""
Log the backtesting progress so user knows how many pairs have been trained and
2022-11-07 18:35:28 +00:00
how many more pairs/trains remain.
:param tr_train: the training timerange
:param train_it: the train iteration for the current pair (the sliding window progress)
:param pair: the current pair
:param total_trains: total trains (total number of slides for the sliding window)
"""
if not self.config.get("freqai_backtest_live_models", False):
logger.info(
f"Training {pair}, {self.pair_it}/{self.total_pairs} pairs"
2022-11-10 17:26:14 +00:00
f" from {tr_train.start_fmt} "
f"to {tr_train.stop_fmt}, {train_it}/{total_trains} "
"trains"
)
def backtesting_fit_live_predictions(self, dk: FreqaiDataKitchen):
"""
Apply fit_live_predictions function in backtesting with a dummy historic_predictions
2022-11-09 13:07:24 +00:00
The loop is required to simulate dry/live operation, as it is not possible to predict
the type of logic implemented by the user.
:param dk: datakitchen object
"""
fit_live_predictions_candles = self.freqai_info.get("fit_live_predictions_candles", 0)
if fit_live_predictions_candles:
logger.info("Applying fit_live_predictions in backtesting")
label_columns = [col for col in dk.full_df.columns if (
col.startswith("&") and
not (col.startswith("&") and col.endswith("_mean")) and
not (col.startswith("&") and col.endswith("_std")) and
col not in self.dk.data["extra_returns_per_train"])
]
for index in range(len(dk.full_df)):
if index >= fit_live_predictions_candles:
self.dd.historic_predictions[self.dk.pair] = (
dk.full_df.iloc[index - fit_live_predictions_candles:index])
self.fit_live_predictions(self.dk, self.dk.pair)
for label in label_columns:
if dk.full_df[label].dtype == object:
continue
if "labels_mean" in self.dk.data:
dk.full_df.at[index, f"{label}_mean"] = (
self.dk.data["labels_mean"][label])
if "labels_std" in self.dk.data:
dk.full_df.at[index, f"{label}_std"] = self.dk.data["labels_std"][label]
for extra_col in self.dk.data["extra_returns_per_train"]:
dk.full_df.at[index, f"{extra_col}"] = (
self.dk.data["extra_returns_per_train"][extra_col])
return
def update_metadata(self, metadata: Dict[str, Any]):
"""
Update global metadata and save the updated json file
:param metadata: new global metadata dict
"""
self.dd.save_global_metadata_to_disk(metadata)
self.metadata = metadata
def set_start_dry_live_date(self, live_dataframe: DataFrame):
key_name = "start_dry_live_date"
if key_name not in self.metadata:
metadata = self.metadata
metadata[key_name] = int(
pd.to_datetime(live_dataframe.tail(1)["date"].values[0]).timestamp())
self.update_metadata(metadata)
2022-11-29 23:53:35 +00:00
def start_backtesting_from_historic_predictions(
self, dataframe: DataFrame, metadata: dict, dk: FreqaiDataKitchen
) -> FreqaiDataKitchen:
"""
:param dataframe: DataFrame = strategy passed dataframe
:param metadata: Dict = pair metadata
:param dk: FreqaiDataKitchen = Data management/analysis tool associated to present pair only
:return:
FreqaiDataKitchen = Data management/analysis tool associated to present pair only
"""
pair = metadata["pair"]
dk.return_dataframe = dataframe
saved_dataframe = self.dd.historic_predictions[pair]
columns_to_drop = list(set(saved_dataframe.columns).intersection(
dk.return_dataframe.columns))
dk.return_dataframe = dk.return_dataframe.drop(columns=list(columns_to_drop))
dk.return_dataframe = pd.merge(
dk.return_dataframe, saved_dataframe, how='left', left_on='date', right_on="date_pred")
# dk.return_dataframe = dk.return_dataframe[saved_dataframe.columns].fillna(0)
return dk
# Following methods which are overridden by user made prediction models.
2022-08-06 12:55:46 +00:00
# See freqai/prediction_models/CatboostPredictionModel.py for an example.
@abstractmethod
def train(self, unfiltered_df: DataFrame, pair: str,
2022-09-07 16:58:55 +00:00
dk: FreqaiDataKitchen, **kwargs) -> Any:
"""
Filter the training data and train a model to it. Train makes heavy use of the datahandler
for storing, saving, loading, and analyzing the data.
:param unfiltered_df: Full dataframe for the current training period
:param metadata: pair metadata from strategy.
:return: Trained model which can be used to inference (self.predict)
"""
@abstractmethod
2022-09-07 16:58:55 +00:00
def fit(self, data_dictionary: Dict[str, Any], dk: FreqaiDataKitchen, **kwargs) -> Any:
"""
Most regressors use the same function names and arguments e.g. user
can drop in LGBMRegressor in place of CatBoostRegressor and all data
management will be properly handled by Freqai.
:param data_dictionary: Dict = the dictionary constructed by DataHandler to hold
all the training and test data/labels.
"""
return
@abstractmethod
2022-07-03 08:59:38 +00:00
def predict(
self, unfiltered_df: DataFrame, dk: FreqaiDataKitchen, **kwargs
) -> Tuple[DataFrame, NDArray[np.int_]]:
"""
Filter the prediction features data and predict with it.
:param unfiltered_df: Full dataframe for the current backtest period.
:param dk: FreqaiDataKitchen = Data management/analysis tool associated to present pair only
:param first: boolean = whether this is the first prediction or not.
:return:
:predictions: np.array of predictions
:do_predict: np.array of 1s and 0s to indicate places where freqai needed to remove
data (NaNs) or felt uncertain about data (i.e. SVM and/or DI index)
"""