reduce unnecessary verbosity, fix error on first training sweep, add LightGBMPredictionModel

This commit is contained in:
robcaulk 2022-06-26 19:02:17 +02:00
parent 852706cd6b
commit 051b99791d
4 changed files with 163 additions and 14 deletions

View File

@ -134,6 +134,7 @@ class FreqaiDataDrawer:
model_filename = self.pair_dict[pair]['model_filename'] = '' model_filename = self.pair_dict[pair]['model_filename'] = ''
coin_first = self.pair_dict[pair]['first'] = True coin_first = self.pair_dict[pair]['first'] = True
trained_timestamp = self.pair_dict[pair]['trained_timestamp'] = 0 trained_timestamp = self.pair_dict[pair]['trained_timestamp'] = 0
self.pair_dict[pair]['priority'] = len(self.pair_dict)
if not data_path_set and self.follow_mode: if not data_path_set and self.follow_mode:
logger.warning(f'Follower could not find current pair {pair} in ' logger.warning(f'Follower could not find current pair {pair} in '

View File

@ -317,6 +317,7 @@ class FreqaiDataKitchen:
# that was based on a single NaN is ultimately protected from buys with do_predict # that was based on a single NaN is ultimately protected from buys with do_predict
drop_index = ~drop_index drop_index = ~drop_index
self.do_predict = np.array(drop_index.replace(True, 1).replace(False, 0)) self.do_predict = np.array(drop_index.replace(True, 1).replace(False, 0))
if (len(self.do_predict) - self.do_predict.sum()) > 0:
logger.info( logger.info(
"dropped %s of %s prediction data points due to NaNs.", "dropped %s of %s prediction data points due to NaNs.",
len(self.do_predict) - self.do_predict.sum(), len(self.do_predict) - self.do_predict.sum(),
@ -562,6 +563,7 @@ class FreqaiDataKitchen:
y_pred = self.svm_model.predict(self.data_dictionary["prediction_features"]) y_pred = self.svm_model.predict(self.data_dictionary["prediction_features"])
do_predict = np.where(y_pred == -1, 0, y_pred) do_predict = np.where(y_pred == -1, 0, y_pred)
if (len(do_predict) - do_predict.sum()) > 0:
logger.info( logger.info(
f'svm_remove_outliers() tossed {len(do_predict) - do_predict.sum()} predictions' f'svm_remove_outliers() tossed {len(do_predict) - do_predict.sum()} predictions'
) )
@ -642,6 +644,7 @@ class FreqaiDataKitchen:
0, 0,
) )
if (len(do_predict) - do_predict.sum()) > 0:
logger.info( logger.info(
f'DI tossed {len(do_predict) - do_predict.sum():.2f} predictions for ' f'DI tossed {len(do_predict) - do_predict.sum():.2f} predictions for '
'being too far from training data' 'being too far from training data'
@ -908,7 +911,7 @@ class FreqaiDataKitchen:
ignore_index=True, axis=0 ignore_index=True, axis=0
) )
logger.info(f'Length of history data {len(history_data[pair][tf])}') # logger.info(f'Length of history data {len(history_data[pair][tf])}')
def set_all_pairs(self) -> None: def set_all_pairs(self) -> None:

View File

@ -0,0 +1,145 @@
import logging
from typing import Any, Dict, Tuple
from lightgbm import LGBMRegressor
from pandas import DataFrame
from freqtrade.freqai.data_kitchen import FreqaiDataKitchen
from freqtrade.freqai.freqai_interface import IFreqaiModel
logger = logging.getLogger(__name__)
class LightGBMPredictionModel(IFreqaiModel):
"""
User created prediction model. The class needs to override three necessary
functions, predict(), train(), fit(). The class inherits ModelHandler which
has its own DataHandler where data is held, saved, loaded, and managed.
"""
def return_values(self, dataframe: DataFrame, dh: FreqaiDataKitchen) -> DataFrame:
dataframe["prediction"] = dh.full_predictions
dataframe["do_predict"] = dh.full_do_predict
dataframe["target_mean"] = dh.full_target_mean
dataframe["target_std"] = dh.full_target_std
if self.freqai_info.get('feature_parameters', {}).get('DI_threshold', 0) > 0:
dataframe["DI"] = dh.full_DI_values
return dataframe
def make_labels(self, dataframe: DataFrame, dh: FreqaiDataKitchen) -> DataFrame:
"""
User defines the labels here (target values).
:params:
:dataframe: the full dataframe for the present training period
"""
dataframe["s"] = (
dataframe["close"]
.shift(-self.feature_parameters["period"])
.rolling(self.feature_parameters["period"])
.mean()
/ dataframe["close"]
- 1
)
return dataframe["s"]
def train(self, unfiltered_dataframe: DataFrame,
pair: str, dh: FreqaiDataKitchen) -> Tuple[DataFrame, DataFrame]:
"""
Filter the training data and train a model to it. Train makes heavy use of the datahkitchen
for storing, saving, loading, and analyzing the data.
:params:
:unfiltered_dataframe: Full dataframe for the current training period
:metadata: pair metadata from strategy.
:returns:
:model: Trained model which can be used to inference (self.predict)
"""
logger.info('--------------------Starting training '
f'{pair} --------------------')
# create the full feature list based on user config info
dh.training_features_list = dh.find_features(unfiltered_dataframe)
unfiltered_labels = self.make_labels(unfiltered_dataframe, dh)
# filter the features requested by user in the configuration file and elegantly handle NaNs
features_filtered, labels_filtered = dh.filter_features(
unfiltered_dataframe,
dh.training_features_list,
unfiltered_labels,
training_filter=True,
)
# split data into train/test data.
data_dictionary = dh.make_train_test_datasets(features_filtered, labels_filtered)
dh.fit_labels() # fit labels to a cauchy distribution so we know what to expect in strategy
# normalize all data based on train_dataset only
data_dictionary = dh.normalize_data(data_dictionary)
# optional additional data cleaning/analysis
self.data_cleaning_train(dh)
logger.info(f'Training model on {len(dh.data_dictionary["train_features"].columns)}'
' features')
logger.info(f'Training model on {len(data_dictionary["train_features"])} data points')
model = self.fit(data_dictionary)
logger.info(f'--------------------done training {pair}--------------------')
return model
def fit(self, data_dictionary: Dict) -> 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.
:params:
:data_dictionary: the dictionary constructed by DataHandler to hold
all the training and test data/labels.
"""
eval_set = (data_dictionary["test_features"], data_dictionary["test_labels"])
X = data_dictionary["train_features"]
y = data_dictionary["train_labels"]
model = LGBMRegressor(seed=42, n_estimators=2000, verbosity=1, force_col_wise=True)
model.fit(X=X, y=y, eval_set=eval_set)
return model
def predict(self, unfiltered_dataframe: DataFrame,
dh: FreqaiDataKitchen) -> Tuple[DataFrame, DataFrame]:
"""
Filter the prediction features data and predict with it.
:param: unfiltered_dataframe: Full dataframe for the current backtest period.
: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 (PCA and DI index)
"""
# logger.info("--------------------Starting prediction--------------------")
original_feature_list = dh.find_features(unfiltered_dataframe)
filtered_dataframe, _ = dh.filter_features(
unfiltered_dataframe, original_feature_list, training_filter=False
)
filtered_dataframe = dh.normalize_data_from_metadata(filtered_dataframe)
dh.data_dictionary["prediction_features"] = filtered_dataframe
# optional additional data cleaning/analysis
self.data_cleaning_predict(dh, filtered_dataframe)
predictions = self.model.predict(dh.data_dictionary["prediction_features"])
# compute the non-normalized predictions
dh.predictions = (predictions + 1) * (dh.data["labels_max"] -
dh.data["labels_min"]) / 2 + dh.data["labels_min"]
# logger.info("--------------------Finished prediction--------------------")
return (dh.predictions, dh.do_predict)

View File

@ -6,4 +6,4 @@ scikit-learn==1.0.2
scikit-optimize==0.9.0 scikit-optimize==0.9.0
joblib==1.1.0 joblib==1.1.0
catboost==1.0.4 catboost==1.0.4
#lightgbm==3.3.2 lightgbm==3.3.2