def forward_selection(max_features, X_train, y_train): ordered_features = [] ordered_scores = [] # Start with no features. selected_features = [] learner = TemporalRegressionAlgorithms() eval = RegressionEvaluation() prev_best_perf = sys.float_info.max # Select the appropriate number of features. for i in range(0, max_features): print i #Determine the features left to select. features_left = list(set(X_train.columns) - set(selected_features)) best_perf = sys.float_info.max best_feature = '' # For all features we can still select... for f in features_left: temp_selected_features = copy.deepcopy(selected_features) temp_selected_features.append(f) # Determine the mse of a decision tree learner if we were to add # the feature. regr_train_y, regr_test_y = learner.recurrent_neural_network( X_train, y_train, X_train, y_train, gridsearch=False) perf, std_tr = eval.mean_squared_error_with_std( train_y.ix[10:, ], regr_train_y.ix[10:, ]) # If the performance is better than what we have seen so far (we aim for low mse) # we set the current feature to the best feature and the same for the best performance. if perf < best_perf: best_perf = perf best_feature = f # We select the feature with the best performance. selected_features.append(best_feature) prev_best_perf = best_perf ordered_features.append(best_feature) ordered_scores.append(best_perf) return selected_features, ordered_features, ordered_scores
gridsearch=True, per_time_step=False) mean_tr, std_tr = eval.mean_squared_error_with_std( train_y.ix[washout_time:, ], regr_train_y.ix[washout_time:, ]) mean_te, std_te = eval.mean_squared_error_with_std( test_y.ix[washout_time:, ], regr_test_y.ix[washout_time:, ]) performance_tr_res += mean_tr performance_tr_res_std += std_tr performance_te_res += mean_te performance_te_res_std += std_te regr_train_y, regr_test_y = learner.recurrent_neural_network( selected_train_X, train_y, selected_test_X, test_y, gridsearch=True) mean_tr, std_tr = eval.mean_squared_error_with_std( train_y.ix[washout_time:, ], regr_train_y.ix[washout_time:, ]) mean_te, std_te = eval.mean_squared_error_with_std( test_y.ix[washout_time:, ], regr_test_y.ix[washout_time:, ]) performance_tr_rnn += mean_tr performance_tr_rnn_std += std_tr performance_te_rnn += mean_te performance_te_rnn_std += std_te # We only apply the time series in case of the basis features. if (feature_names[i] == 'initial set'):
def main(): # Read the result from the previous chapter and convert the index to datetime try: dataset = pd.read_csv(DATA_PATH / DATASET_FNAME, index_col=0) dataset.index = pd.to_datetime(dataset.index) except IOError as e: print( 'File not found, try to run previous crowdsignals scripts first!') raise e # Create an instance of visualization class to plot the results DataViz = VisualizeDataset(__file__) # Consider the second task, namely the prediction of the heart rate. Therefore create a dataset with the heart # rate as target and split using timestamps, because this is considered as a temporal task print('\n- - - Loading dataset - - -') prepare = PrepareDatasetForLearning() train_X, test_X, train_y, test_y = prepare.split_single_dataset_regression_by_time( dataset, 'hr_watch_rate', '2016-02-08 18:29:56', '2016-02-08 19:34:07', '2016-02-08 20:07:50') print('Training set length is: ', len(train_X.index)) print('Test set length is: ', len(test_X.index)) # Select subsets of the features print('\n- - - Selecting subsets - - -') basic_features = [ 'acc_phone_x', 'acc_phone_y', 'acc_phone_z', 'acc_watch_x', 'acc_watch_y', 'acc_watch_z', 'gyr_phone_x', 'gyr_phone_y', 'gyr_phone_z', 'gyr_watch_x', 'gyr_watch_y', 'gyr_watch_z', 'labelOnTable', 'labelSitting', 'labelWashingHands', 'labelWalking', 'labelStanding', 'labelDriving', 'labelEating', 'labelRunning', 'light_phone_lux', 'mag_phone_x', 'mag_phone_y', 'mag_phone_z', 'mag_watch_x', 'mag_watch_y', 'mag_watch_z', 'press_phone_pressure' ] pca_features = [ 'pca_1', 'pca_2', 'pca_3', 'pca_4', 'pca_5', 'pca_6', 'pca_7' ] time_features = [ name for name in dataset.columns if ('temp_' in name and 'hr_watch' not in name) ] freq_features = [ name for name in dataset.columns if (('_freq' in name) or ('_pse' in name)) ] cluster_features = ['cluster'] print('#basic features: ', len(basic_features)) print('#PCA features: ', len(pca_features)) print('#time features: ', len(time_features)) print('#frequency features: ', len(freq_features)) print('#cluster features: ', len(cluster_features)) features_after_chapter_3 = list(set().union(basic_features, pca_features)) features_after_chapter_4 = list(set().union(features_after_chapter_3, time_features, freq_features)) features_after_chapter_5 = list(set().union(features_after_chapter_4, cluster_features)) selected_features = [ 'temp_pattern_labelOnTable', 'labelOnTable', 'temp_pattern_labelOnTable(b)labelOnTable', 'cluster', 'pca_1_temp_mean_ws_120', 'pca_2_temp_mean_ws_120', 'pca_2', 'acc_watch_y_temp_mean_ws_120', 'gyr_watch_y_pse', 'gyr_watch_x_pse' ] possible_feature_sets = [ basic_features, features_after_chapter_3, features_after_chapter_4, features_after_chapter_5, selected_features ] feature_names = [ 'initial set', 'Chapter 3', 'Chapter 4', 'Chapter 5', 'Selected features' ] if FLAGS.mode == 'correlation' or FLAGS.mode == 'all': # First study whether the time series is stationary and what the autocorrelations are adfuller(dataset['hr_watch_rate'], autolag='AIC') plt.Figure() autocorrelation_plot(dataset['hr_watch_rate']) DataViz.save(plt) plt.show() # Now focus on the learning part learner = TemporalRegressionAlgorithms() evaluate = RegressionEvaluation() if FLAGS.mode == 'overall' or FLAGS.mode == 'all': # Repeat the experiment a number of times to get a bit more robust data as the initialization of e.g. the NN is # random repeats = FLAGS.repeats # Set a washout time to give the NN's the time to stabilize (so don't compute the error during the washout time) washout_time = FLAGS.washout scores_over_all_algs = [] for i in range(0, len(possible_feature_sets)): print(f'Evaluating for features {possible_feature_sets[i]}') selected_train_X = train_X[possible_feature_sets[i]] selected_test_X = test_X[possible_feature_sets[i]] # First run non deterministic classifiers a number of times to average their score performance_tr_res, performance_tr_res_std = 0, 0 performance_te_res, performance_te_res_std = 0, 0 performance_tr_rnn, performance_tr_rnn_std = 0, 0 performance_te_rnn, performance_te_rnn_std = 0, 0 for repeat in range(0, repeats): print(f'--- run {repeat} ---') regr_train_y, regr_test_y = learner.reservoir_computing( selected_train_X, train_y, selected_test_X, test_y, gridsearch=True, per_time_step=False) mean_tr, std_tr = evaluate.mean_squared_error_with_std( train_y.iloc[washout_time:, ], regr_train_y.iloc[washout_time:, ]) mean_te, std_te = evaluate.mean_squared_error_with_std( test_y.iloc[washout_time:, ], regr_test_y.iloc[washout_time:, ]) performance_tr_res += mean_tr performance_tr_res_std += std_tr performance_te_res += mean_te performance_te_res_std += std_te regr_train_y, regr_test_y = learner.recurrent_neural_network( selected_train_X, train_y, selected_test_X, test_y, gridsearch=True) mean_tr, std_tr = evaluate.mean_squared_error_with_std( train_y.iloc[washout_time:, ], regr_train_y.iloc[washout_time:, ]) mean_te, std_te = evaluate.mean_squared_error_with_std( test_y.iloc[washout_time:, ], regr_test_y.iloc[washout_time:, ]) performance_tr_rnn += mean_tr performance_tr_rnn_std += std_tr performance_te_rnn += mean_te performance_te_rnn_std += std_te # Only apply the time series in case of the basis features if feature_names[i] == 'initial set': regr_train_y, regr_test_y = learner.time_series( selected_train_X, train_y, selected_test_X, test_y, gridsearch=True) mean_tr, std_tr = evaluate.mean_squared_error_with_std( train_y.iloc[washout_time:, ], regr_train_y.iloc[washout_time:, ]) mean_te, std_te = evaluate.mean_squared_error_with_std( test_y.iloc[washout_time:, ], regr_test_y.iloc[washout_time:, ]) overall_performance_tr_ts = mean_tr overall_performance_tr_ts_std = std_tr overall_performance_te_ts = mean_te overall_performance_te_ts_std = std_te else: overall_performance_tr_ts = 0 overall_performance_tr_ts_std = 0 overall_performance_te_ts = 0 overall_performance_te_ts_std = 0 overall_performance_tr_res = performance_tr_res / repeats overall_performance_tr_res_std = performance_tr_res_std / repeats overall_performance_te_res = performance_te_res / repeats overall_performance_te_res_std = performance_te_res_std / repeats overall_performance_tr_rnn = performance_tr_rnn / repeats overall_performance_tr_rnn_std = performance_tr_rnn_std / repeats overall_performance_te_rnn = performance_te_rnn / repeats overall_performance_te_rnn_std = performance_te_rnn_std / repeats scores_with_sd = [ (overall_performance_tr_res, overall_performance_tr_res_std, overall_performance_te_res, overall_performance_te_res_std), (overall_performance_tr_rnn, overall_performance_tr_rnn_std, overall_performance_te_rnn, overall_performance_te_rnn_std), (overall_performance_tr_ts, overall_performance_tr_ts_std, overall_performance_te_ts, overall_performance_te_ts_std) ] util.print_table_row_performances_regression( feature_names[i], scores_with_sd) scores_over_all_algs.append(scores_with_sd) DataViz.plot_performances_regression( ['Reservoir', 'RNN', 'Time series'], feature_names, scores_over_all_algs) if FLAGS.mode == 'detail' or FLAGS.mode == 'all': regr_train_y, regr_test_y = learner.reservoir_computing( train_X[features_after_chapter_5], train_y, test_X[features_after_chapter_5], test_y, gridsearch=False) DataViz.plot_numerical_prediction_versus_real( train_X.index, train_y, regr_train_y['hr_watch_rate'], test_X.index, test_y, regr_test_y['hr_watch_rate'], 'heart rate') regr_train_y, regr_test_y = learner.recurrent_neural_network( train_X[basic_features], train_y, test_X[basic_features], test_y, gridsearch=True) DataViz.plot_numerical_prediction_versus_real( train_X.index, train_y, regr_train_y['hr_watch_rate'], test_X.index, test_y, regr_test_y['hr_watch_rate'], 'heart rate') regr_train_y, regr_test_y = learner.time_series( train_X[basic_features], train_y, test_X[basic_features], test_y, gridsearch=True) DataViz.plot_numerical_prediction_versus_real( train_X.index, train_y, regr_train_y['hr_watch_rate'], test_X.index, test_y, regr_test_y['hr_watch_rate'], 'heart rate') if FLAGS.mode == 'dynamical' or FLAGS.mode == 'all': # And now some example code for using the dynamical systems model with parameter tuning (note: focus on # predicting accelerometer data): train_X, test_X, train_y, test_y = prepare.split_single_dataset_regression( copy.deepcopy(dataset), ['acc_phone_x', 'acc_phone_y'], 0.9, filter_data=False, temporal=True) output_sets = learner. \ dynamical_systems_model_nsga_2(train_X, train_y, test_X, test_y, ['self.acc_phone_x', 'self.acc_phone_y', 'self.acc_phone_z'], ['self.a * self.acc_phone_x + self.b * self.acc_phone_y', 'self.c * self.acc_phone_y + self.d * self.acc_phone_z', 'self.e * self.acc_phone_x + self.f * self.acc_phone_z'], ['self.acc_phone_x', 'self.acc_phone_y'], ['self.a', 'self.b', 'self.c', 'self.d', 'self.e', 'self.f'], pop_size=10, max_generations=10, per_time_step=True) DataViz.plot_pareto_front(output_sets) DataViz.plot_numerical_prediction_versus_real_dynsys_mo( train_X.index, train_y, test_X.index, test_y, output_sets, 0, 'acc_phone_x') regr_train_y, regr_test_y = learner. \ dynamical_systems_model_ga(train_X, train_y, test_X, test_y, ['self.acc_phone_x', 'self.acc_phone_y', 'self.acc_phone_z'], ['self.a * self.acc_phone_x + self.b * self.acc_phone_y', 'self.c * self.acc_phone_y + self.d * self.acc_phone_z', 'self.e * self.acc_phone_x + self.f * self.acc_phone_z'], ['self.acc_phone_x', 'self.acc_phone_y'], ['self.a', 'self.b', 'self.c', 'self.d', 'self.e', 'self.f'], pop_size=5, max_generations=10, per_time_step=True) DataViz.plot_numerical_prediction_versus_real( train_X.index, train_y['acc_phone_x'], regr_train_y['acc_phone_x'], test_X.index, test_y['acc_phone_x'], regr_test_y['acc_phone_x'], 'acc_phone_x') regr_train_y, regr_test_y = learner. \ dynamical_systems_model_sa(train_X, train_y, test_X, test_y, ['self.acc_phone_x', 'self.acc_phone_y', 'self.acc_phone_z'], ['self.a * self.acc_phone_x + self.b * self.acc_phone_y', 'self.c * self.acc_phone_y + self.d * self.acc_phone_z', 'self.e * self.acc_phone_x + self.f * self.acc_phone_z'], ['self.acc_phone_x', 'self.acc_phone_y'], ['self.a', 'self.b', 'self.c', 'self.d', 'self.e', 'self.f'], max_generations=10, per_time_step=True) DataViz.plot_numerical_prediction_versus_real( train_X.index, train_y['acc_phone_x'], regr_train_y['acc_phone_x'], test_X.index, test_y['acc_phone_x'], regr_test_y['acc_phone_x'], 'acc_phone_x')