from logging.config import fileConfig
from os import listdir
from pathlib import Path
+from sklearn import preprocessing
import numpy as np
import pandas as pd
fileConfig((Path.cwd() / 'config') / 'logging.cfg')
logger = getLogger()
+
class Preprocessing:
'''
Generate a pandas dataframe from a dictionary of features per datetime, which
- NaN values are then filled with last known values.
'''
- def __init__(self, config_file = None, dict_features = None, features = None):
+ def __init__(self, config_file=None,
+ dict_features=None, dict_target=None):
'''
Constructor that defines all needed attributes and collects features.
'''
- self._config = ConfigParser()
- self._config.read(config_file)
+ self._config = config_file
self._start = datetime.strptime(self._config['DATETIME']['start'],
'%m/%d/%Y %H:%M:%S')
self._end = datetime.strptime(self._config['DATETIME']['end'],
- '%m/%d/%Y %H:%M:%S')
- self._timestep = timedelta(hours =
- self._config['DATETIME'].getfloat('hourStep'))
+ '%m/%d/%Y %H:%M:%S')
+ self._timestep = timedelta(hours=self._config['DATETIME'].getfloat('hourStep'))
self._dict_features = dict_features
+ self._dict_target = dict_target
+
self._full_dict = None
self._dataframe = None
self._datetimes = []
- # If features are not provided to the constructor, then we collect
- # any existing feature in the dictionary
- if features != None:
- self._features = features
- else:
- self._features = set(chain.from_iterable([tuple(u.keys())
- for u in [*dict_features.values()]]))
- feature_files = Path.cwd() / 'config' / 'features'
- self._features = {feat : {'numerical': False} for feat in self._features}
- for feature_file in listdir(feature_files):
- if feature_file.endswith('csv'):
- with open(feature_files / feature_file , "r") as f:
- reader = DictReader(f, delimiter=',')
- typed_names = {row['name']: row['type'] for row in reader}
- for feature in self._features:
- if feature.split('_')[0] in typed_names:
- self._features[feature]['type'] = int(typed_names[feature.split('_')[0]])
- elif feature_file.endswith('cfg'):
+
+ self._features = set(chain.from_iterable([tuple(u.keys())
+ for u in [*dict_features.values()]]))
+
+ #feature_files = Path.cwd() / 'config' / 'features'
+ self._features = {feat: {'numerical': False, 'categorical': False}
+ for feat in self._features}
+
+ for feature in self._config['FEATURES']:
+ if self._config['FEATURES'][feature]:
+ feature_file = self._config['FEATURE_CONFIG'][feature]
config = ConfigParser()
- config.read(feature_files / feature_file)
+ config.read(eval(feature_file))
for section in config:
if config.has_option(section, 'numerical'):
- self._features[section]['numerical'] = config[section].getboolean('numerical')
-
+ for feature in self._features:
+ if feature.split('_')[0] == section:
+ self._features[feature]['binary'] = config[section].getboolean('binary')
+ self._features[feature]['categorical'] = config[section].getboolean('categorical')
+ self._features[feature]['numerical'] = config[section].getboolean('numerical')
+ self._binary_columns = [k for k in self._features if self._features[k]['binary']]
+ self._categorical_columns = [k for k in self._features if self._features[k]['categorical']]
+ self._numerical_columns = [k for k in self._features if self._features[k]['numerical']]
@property
def start(self):
def start(self, x):
self._start = x
-
@property
def end(self):
return self._end
def end(self, x):
self._end = x
-
@property
def timestep(self):
return self._timestep
def timestep(self, x):
self._timestep = x
-
def _fill_dict(self):
'''
Add datetime keys in the dated feature dictionary that are missing. The
while current <= self._end:
self._datetimes.append(current)
if current not in self._dict_features:
- self._dict_features[current] = {feature:np.NaN
+ self._dict_features[current] = {feature: np.NaN
for feature in self._features}
else:
- null_dict = {feature:np.NaN
+ null_dict = {feature: np.NaN
for feature in self._features}
null_dict.update(self._dict_features[current])
self._dict_features[current] = null_dict
current += self._timestep
for k in self._dict_features:
- null_dict = {feature:np.NaN
+ null_dict = {feature: np.NaN
for feature in self._features}
null_dict.update(self._dict_features[k])
self._dict_features[k] = null_dict
self._full_dict = {k: self._dict_features[k]
for k in sorted(self._dict_features.keys())}
-
-
@property
def full_dict(self):
'''
self._fill_dict()
return self._full_dict
+ def _fill_nan(self):
+ '''
+ Fill NaN values, either by propagation or by interpolation (linear or splines)
+ '''
+ logger.info("Filling NaN numerical values in the feature dataframe")
+ # We interpolate (linearly or with splines) only numerical columns
+ # The interpolation
+ if self._config['PREPROCESSING']['fill_method'] == 'propagate':
+ self._dataframe[self._numerical_columns] =\
+ self._dataframe[self._numerical_columns].fillna(method='ffill')
+ elif self._config['PREPROCESSING']['fill_method'] == 'linear':
+ self._dataframe[self._numerical_columns] =\
+ self._dataframe[self._numerical_columns].interpolate()
+ elif self._config['PREPROCESSING']['fill_method'] == 'spline':
+ self._dataframe[self._numerical_columns] =\
+ self._dataframe[self._numerical_columns].interpolate(method='spline',
+ order=self._config['PREPROCESSING'].getint('order'))
+
+ # For the categorical columns, NaN values are filled by duplicating
+ # the last known value (forward fill method)
+ logger.info("Filling NaN categorical values in the feature dataframe")
+ self._dataframe[self._categorical_columns] =\
+ self._dataframe[self._categorical_columns].fillna(method='ffill')
+
+ # Uncomment this line to fill NaN values at the beginning of the
+ # dataframe. This may not be a good idea, especially for features
+ # that are available only for recent years, e.g., air quality
+ #self._dataframe = self._dataframe.fillna(method='bfill')
+
+ # Dropping rows that are not related to our datetime window (start/
+ # step / end)
+ logger.info("Dropping rows that are not related to our datetime window")
+ dates = tuple((x.year, x.month, x.day, x.hour) for x in self._datetimes)
+ self._dataframe['row_ok'] =\
+ self._dataframe.apply(lambda x: (int(x.year), int(x.month), int(x.dayInMonth), int(x.hour)) in dates, axis=1)
+ self._dataframe = self._dataframe[self._dataframe['row_ok']]
+ self._dataframe = self._dataframe.drop(['row_ok'], axis=1)
+ logger.info("Rows dropped")
+
+ def _add_history(self):
+ '''
+ Integrating previous nb of interventions as features
+ '''
+ logger.info("Integrating previous nb of interventions as features")
+ nb_lines = eval(self._config['HISTORY_KNOWLEDGE']['nb_lines'])
+ for k in range(1, nb_lines + 1):
+ name = 'history_' + str(nb_lines - k + 1)
+ self._dataframe[name] = [np.NaN] * k + list(self._dict_target.values())[:-k]
+ self._numerical_columns.append(name)
+ self._dataframe = self._dataframe[nb_lines:]
+
+ def _standardize(self):
+ '''
+ Normalizing numerical features
+ '''
+ logger.info("Standardizing numerical values in the feature dataframe")
+ # We operate only on numerical columns
+ self._dataframe[self._numerical_columns] =\
+ preprocessing.scale(self._dataframe[self._numerical_columns])
+ def _one_hot_encoding(self):
+ '''
+ Apply a one hot encoding for category features
+ '''
+ logger.info("One hot encoding for categorical feature")
+ # We store numerical columns
+ df_out = pd.DataFrame()
+ for col in self._numerical_columns:
+ df_out[col] = self._dataframe[col]
+ # Idem for binary features
+ for col in self._binary_columns:
+ df_out[col] = self._dataframe[col]
+ # The one hot encoding
+ for col in self._categorical_columns:
+ pd1 = pd.get_dummies(self._dataframe[col], prefix=col)
+ for col1 in pd1.columns:
+ df_out[col1] = pd1[col1]
+ self._dataframe = df_out
@property
def dataframe(self):
logger.info("Creating feature dataframe from feature dictionary")
self._dataframe = pd.DataFrame.from_dict(self.full_dict,
orient='index')
- logger.info("Filling NaN numerical values in the feature dataframe")
- # We interpolate (linearly or with splines) only numerical columns
- numerical_columns = [k for k in self._features if self._features[k]['type'] == 1
- or (self._features[k]['type'] == 3 and self._features[k]['numerical'])]
- # The interpolation
- if self._config['PREPROCESSING']['fill_method'] == 'propagate':
- self._dataframe[numerical_columns] =\
- self._dataframe[numerical_columns].fillna(method='ffill')
- elif self._config['PREPROCESSING']['fill_method'] == 'linear':
- self._dataframe[numerical_columns] =\
- self._dataframe[numerical_columns].interpolate()
- elif self._config['PREPROCESSING']['fill_method'] == 'spline':
- self._dataframe[numerical_columns] =\
- self._dataframe[numerical_columns].interpolate(method='spline',
- order=self._config['PREPROCESSING'].getint('order'))
-
- # For the categorical columns, NaN values are filled by duplicating
- # the last known value (forward fill method)
- logger.info("Filling NaN categorical values in the feature dataframe")
- categorical_columns = [k for k in self._features if self._features[k]['type'] == 2
- or (self._features[k]['type'] == 3 and not self._features[k]['numerical'])]
- self._dataframe[categorical_columns] =\
- self._dataframe[categorical_columns].fillna(method='ffill')
-
- # Uncomment this line to fill NaN values at the beginning of the
- # dataframe. This may not be a good idea, especially for features
- # that are available only for recent years, e.g., air quality
- #self._dataframe = self._dataframe.fillna(method='bfill')
-
- # Dropping rows that are not related to our datetime window (start/
- # step / end)
- self._dataframe = self._dataframe.drop([k.to_pydatetime()
- for k in self._dataframe.T
- if k not in self._datetimes])
+ # Dealing with NaN values
+ self._fill_nan()
+ # Adding previous (historical) nb_interventions as features
+ self._add_history()
+ # Normalizing numerical values
+ self._standardize()
+ # Dealing with categorical features
+ self._one_hot_encoding()
return self._dataframe
-
@dataframe.setter
def dataframe(self, df):
self._dataframe = df
