predictops/learn/learning.py

   1 from configparser import ConfigParser
   2 from logging import getLogger
   3 from logging.config import fileConfig
   4 from math import sqrt
   5 from pathlib import Path
   6 from sklearn.metrics import mean_squared_error, mean_absolute_error
   7 from sklearn.model_selection import train_test_split
   8 from statistics import mean, stdev
   9
  10 import lightgbm as lgb
  11 import matplotlib
  12 import os
  13 import pandas as pd
  14 import pylab as P
  15 import xgboost
  16
  17 fileConfig((Path.cwd() / 'config') / 'logging.cfg')
  18 logger = getLogger()
  19
  20
  21 class Learning:
  22
  23     def __init__(self, config_file=None, file_name=None,
  24                  X=None, y=None, horizon=0):
  25         self._config = ConfigParser()
  26         self._config.read(config_file)
  27         self._file_name = file_name
  28         logger.info("Dealing with the horizon of prediction")
  29         if horizon:
  30             self._X = X[:-horizon]
  31             self._y = y[horizon:]
  32         else:
  33             self._X = X
  34             self._y = y
  35         self._learn()
  36         self._evaluate()
  37
  38     def _learn(self):
  39         logger.info("Generation of learning sets")
  40         self._df = self._X
  41         self._df['cible'] = self._y
  42         train_val_set, test_set = train_test_split(self._df, test_size=0.2, random_state=42)
  43         train_set, val_set = train_test_split(train_val_set, test_size=0.2, random_state=42)
  44
  45         self._X_test = test_set.drop('cible', axis=1)
  46         self._y_test = test_set['cible'].copy()
  47
  48         X_train = train_set.drop('cible', axis=1)
  49         y_train = train_set['cible'].copy()
  50         X_val = val_set.drop('cible', axis=1)
  51         y_val = val_set['cible'].copy()
  52
  53         logger.info("Start learning")
  54         if self._config['MODEL']['method'] == 'xgboost':
  55             logger.info("Using xgboost regressor")
  56             self._reg = xgboost.XGBRegressor(learning_rate=self._config['HYPERPARAMETERS'].getfloat('learning_rate'),
  57                                              max_depth=self._config['HYPERPARAMETERS'].getint('max_depth'),
  58                                              random_state=self._config['HYPERPARAMETERS'].getint('random_state'),
  59                                              n_estimators=self._config['HYPERPARAMETERS'].getint('n_estimators'),
  60                                              n_jobs=self._config['HYPERPARAMETERS'].getint('n_jobs'),
  61                                              objective='count:poisson')
  62
  63             self._reg.fit(X_train, y_train,
  64                           eval_set=[(X_val, y_val)],
  65                           early_stopping_rounds=10)
  66         elif self._config['MODEL']['method'] == 'lightgbm':
  67             train_data = lgb.Dataset(X_train, label=y_train)
  68             val_data = lgb.Dataset(X_val, label=y_val)
  69             num_round = self._config['HYPERPARAMETERS'].getint('num_round')
  70             param = {
  71                 'learning_rate': self._config['HYPERPARAMETERS'].getfloat('learning_rate'),
  72                 'metric': self._config['HYPERPARAMETERS'].get('metric'),
  73                 'num_iterations': self._config['HYPERPARAMETERS'].getint('num_iterations'),
  74                 'num_leaves': self._config['HYPERPARAMETERS'].getint('num_leaves'),
  75                 'objective': self._config['HYPERPARAMETERS'].get('objective')
  76             }
  77             self._reg = lgb.train(param, train_data, num_round, valid_sets=[val_data])
  78
  79     def _evaluate(self):
  80         logger.info("Evaluation of the learner")
  81         y_test_pred = self._reg.predict(self._X_test)
  82         txt = f"Average interventions per time unit: {mean(self._df.cible)}\n"
  83         txt += f"Standard deviation: {stdev(self._df.cible)}\n\n"
  84
  85         txt += f"Mean absolute error: {mean_absolute_error(y_test_pred, self._y_test)}\n"
  86         txt += f"Root mean squared error: {sqrt(mean_squared_error(y_test_pred, self._y_test))}\n\n"
  87
  88         for k in range(10):
  89             txt += f"Percentage of errors lower than {k}: {[abs(int(u-v))<=k for u,v in zip(self._y_test.values, y_test_pred)].count(True)/len(self._y_test)*100}\n"
  90
  91         rep = (Path.cwd() / self._file_name)
  92         rep.mkdir()
  93         self._filename = str(self._file_name / os.path.basename(self._file_name))
  94
  95         print(txt)
  96         with open(self._filename + ".result", 'w') as f:
  97             f.write(txt)
  98
  99         y_true = self._df[self._df.year == self._df.year.max()].cible
 100         x_true = self._df[self._df.year == self._df.year.max()].drop('cible', axis=1)
 101
 102         yy_test_pred = self._reg.predict(x_true)
 103         P.figure(figsize=(36, 16))
 104         P.plot(list(y_true)[:300], color='blue', label='actual')
 105         P.plot(yy_test_pred[:300], color='red', label='predicted')
 106         P.title('Predictions for 2018')
 107         P.xlabel('Hour in the year')
 108         P.ylabel('Number of cumulated interventions')
 109         P.legend()
 110         P.savefig(self._filename + ".png")
 111
 112         yy_test_pred = self._reg.predict(self._X_test)
 113         P.figure(figsize=(36, 16))
 114         P.plot(list(self._y_test)[:300], color='blue', label='actual')
 115         P.plot(yy_test_pred[:300], color='red', label='predicted')
 116         P.title('Predictions for test set')
 117         P.xlabel('Hour in the year')
 118         P.ylabel('Number of cumulated interventions')
 119         P.legend()
 120         P.savefig(self._filename + "-test.png")
 121
 122         if self._config['MODEL']['method'] == 'xgboost':
 123             xgboost.plot_importance(self._reg)
 124             fig = matplotlib.pyplot.gcf()
 125             fig.set_size_inches(15, 130)
 126             fig.savefig(self._filename + '-feat_importance.pdf')