def best_ratio(yProba, yValidation, weightsValidation, pas = 0.01): ratio_s = np.arange(0., 0.99, pas) best_ams = 0. for ratio in ratio_s: yPredicted = get_yPredicted_ratio(yProba, ratio) s, b = submission.get_s_b(yPredicted, yValidation, weightsValidation) if b >= 0. and s >= 0.: ams = hbc.AMS(s,b) if ams >= best_ams: best_ratio = ratio best_ams = ams else: if b < 0.: print ("WARNING: For a ratio of %f, b < 0 (b= %f).") %(ratio, b) print ("This ratio has been ignored.") else: print ("WARNING: For a ratio of %f, s < 0 (s= %f).") %(ratio, s) print ("This ratio has been ignored.") ams = hbc.AMS(s,b) if ams >= best_ams: best_ratio = ratio best_ams = ams return best_ams, best_ratio
def best_treshold(yProba, yValidation, weightsValidation, pas = 0.01): """ Returns the treshold that maximises the AMS for the vectors of proba given yProba : vectors of the probabilities computed with a classifier yValid : vectors of the true label of the data yWeights : vectors of the weights pas : size of the interval between two probabilities tested The weights must be balanced ! """ treshold_s = np.arange(0., 1.0, pas) best_ams = 0. for treshold in treshold_s: yPredicted_prov = get_yPredicted_treshold(yProba, treshold) # if we work with multi-class: if len(yPredicted_prov.shape) == 2: if yPredicted_prov.shape[1] == 5: # Reduce multiclass to binary yPredicted = np.ones(yPredicted_prov.shape[0]) yPredicted[yPredicted_prov[:,4] == 0] = 0 else: print "Error: in best_treshold() the shape of the input isn't correct" else: yPredicted = yPredicted_prov s, b = submission.get_s_b(yPredicted, yValidation, weightsValidation) ams = hbc.AMS(s,b) if ams >= best_ams: best_treshold = treshold best_ams = ams return best_ams, best_treshold
def evaluate_AMS(final_prediction, valid_s): # Get s and b for each group (s_s, b_s) and the final final_s and # final_b: if type(final_prediction[2]) == list: y_predicted_s = list(zip(*final_prediction)[2]) final_s, final_b, s_s, b_s = submission.get_s_b(y_predicted_s, valid_s[2], valid_s[3]) else: y_predicted_s = final_prediction[2] final_s, final_b = submission.get_s_b(y_predicted_s, valid_s[2], valid_s[3]) # Balance the s and b final_s *= 250000/25000 final_b *= 250000/25000 # AMS final: AMS = hbc.AMS(final_s , final_b) print ("Expected AMS score for the 'on-top' classifier : %f") %AMS #AMS by group if type(y_predicted_s) == list: AMS_s = [] for i, (s,b) in enumerate(zip(s_s, b_s)): s *= 250000/y_predicted_s[i].shape[0] b *= 250000/y_predicted_s[i].shape[0] score = hbc.AMS(s,b) AMS_s.append(score) print("Expected AMS score for the 'on-top' classifer: group %i : %f" \ %(i, score)) print(" ") else: AMS_s = AMS return final_s, final_b, AMS, AMS_s
def best_ratio_combinaison(yProba_s, yValidation_s, weightsValidation_s, ratio_s): """ returns the best ratio combinaison with the ratios specified in ratio_s for each group ratio_s : List of the list of the ratios to test for each group the size of each list should not exceed 4 for computationnal time issues """ best_ratio_comb = [0.,0.,0.,0.,0.,0.,0.,0.] AMS_max = 0. """ ratio_1_s = [0.06, 0.08,0.10,0.12] ratio_2_s = [0.15,0.16,0.17,0.18] ratio_3_s = [0.36,0.38,0.40,0.42] ratio_4_s = [0.16,0.18,0.2,0.22] ratio_5_s = [0.007,0.008,0.009,0.01] ratio_6_s = [0.003,0.004,0.005,0.006] ratio_7_s = [0.003,0.004,0.005,0.006] ratio_8_s = [0.007,0.008,0.009,0.01] """ g_combinaisons = itertools.product(ratio_s[0], ratio_s[1], ratio_s[2], ratio_s[3], ratio_s[4], ratio_s[5], ratio_s[6], ratio_s[7]) # if we work with multi-class: if len(yProba_s[0].shape) == 2: if yProba_s[0].shape[1] == 5: for i,subset in enumerate(yProba_s): yProba_s[i] = preTreatment.multiclass2binary(subset) compteur = 0 for combinaison in g_combinaisons: #if compteur%10000==0: # print "number of iterations : %i" %compteur compteur +=1 L = list(combinaison) yPredicted_s, yPredicted_conca = get_yPredicted_ratio_8(yProba_s, L) finals, finalb, s_s, b_s = submission.get_s_b(yPredicted_s, yValidation_s, weightsValidation_s) AMS = hbc.AMS(finals, finalb) if AMS > AMS_max: AMS_max = AMS best_ratio_comb = L return AMS_max, best_ratio_comb
def main(): ############### ### IMPORT #### ############### # Importation parameters: split= True normalize = True noise_var = 0. ratio_train = 0.9 # Import the training data: print("Extracting the data sets...") start = time.clock() train_s, valid_s, test_s = tokenizer.extract_data(split= split, \ normalize= normalize, \ noise_variance= noise_var, \ ratio_train= ratio_train) yValid_conca = preTreatment.concatenate_vectors(valid_s[2]) weights_conca = preTreatment.concatenate_vectors(valid_s[3]) stop = time.clock() print ("Extraction time: %i s") %(stop-start) print(" ") print(" ") # Create the elected vectors for each group (best AMS score) best_yPredicted_s = [np.zeros(valid_s[2][i].shape[0]) for i in range(8)] best_yProba_s = [np.zeros(valid_s[2][i].shape[0]) for i in range(8)] best_AMS_s = [0. for i in range(8)] best_method_s = [0 for i in range(8)] best_ratio_s = [0 for i in range(8)] best_AMS_1_method = 0. best_method = "methode" best_ratio = "0." ###################### ### PRE-TREATMENT #### ###################### print("------------------------- Pre-treatment --------------------------") ### Average number of signal per subset: print("Train subsets signal average:") train_s_average = preTreatment.ratio_sig_per_dataset(train_s[2]) print(" ") print("Valid subsets signal average:") valid_s_average = preTreatment.ratio_sig_per_dataset(valid_s[2]) print(" ") print(" ") ############ # ANALYSES # ############ # Dictionnary that will contain all the data for each methods. In the end # we'll have a dict of dict # Keys of the methods : {naiveBayes, svm, kNeighbors, lda, qda, adaBoost, # randomForest, gradientBoosting} dMethods ={} # NAIVE BAYES: kwargs_bayes = {} dMethods['naiveBayes'] = analyse.analyse(train_s, valid_s, 'naiveBayes', kwargs_bayes) kwargs_bayes = {} dMethods['naiveBayes'] = analyse.analyse(train_s, valid_s, 'naiveBayes', kwargs_bayes) # SVM kwargs_svm ={} dMethods['svm'] = analyse.analyse(train_s, valid_s,'svm', kwargs_svm) # K NEIGHBORS kwargs_tuning_kn = {'n_neighbors': [20,50]} dTuning = tuningModel.parameters_grid_search(train_s, valid_s, 'kNeighbors', kwargs_tuning_kn) dMethods['kNeighbors'] = combineClassifiers.select_best_classifiers(dTuning, valid_s) # LDA kwargs_lda = {} dMethods['lda'] = analyse.analyse(train_s, valid_s, 'lda', kwargs_lda) # QDA kwargs_qda= {} dMethods['qda'] = analyse.analyse(train_s, valid_s, 'qda', kwargs_qda) # ADABOOST kwargs_ada= { 'n_estimators': 50, 'learning_rate': 1., 'algorithm': 'SAMME.R', 'random_state':None} dMethods['adaBoost'] = analyse.analyse(train_s, valid_s, 'adaBoost', kwargs_ada) # RANDOM FOREST: kwargs_tuning_rdf = {'n_estimators': [10,50,100]} dTuning = tuningModel.parameters_grid_search(train_s, valid_s, 'randomForest', kwargs_tuning_rdf) dMethods['randomForest'] = combineClassifiers.select_best_classifiers(dTuning, valid_s) # GRADIENT BOOSTING kwargs_gradB = {} dMethods['gradientBoosting'] = analyse.analyse(train_s, valid_s, 'gradientBoosting', kwargs_gradB) kwargs_tuning_gradB = {'loss': ['deviance'], 'learning_rate': [0.1], 'n_estimators': [100], 'subsample': [1.0], 'min_samples_split': [2], 'min_samples_leaf': [1], 'max_depth': [10], 'init': [None], 'random_state': [None], 'max_features': [None], 'verbose': [0]} dTuning = tuningModel.parameters_grid_search(train_s, valid_s, 'gradientBoosting', kwargs_tuning_gradB) dMethods['gradientBoosting'] = combineClassifiers.select_best_classifiers( dTuning, valid_s) print(" ") ################## # POST-TREATMENT # ################## print("-------------------- Best overall combination --------------------") dCombine = combineClassifiers.select_best_classifiers(dMethods, valid_s) print("-------------------------- Thresholding --------------------------") # COMBINED CLASSIFIERS: f = open("Tests/test_treshold_combined.txt","w") yProba_s = dCombine['yProba_s'] yPredicted_s = dCombine['yPredicted_s'] #Let's concatenate the vectors yProba_conca = preTreatment.concatenate_vectors(yProba_s) yPredicted_conca = preTreatment.concatenate_vectors(yPredicted_s) # Best treshold global best_treshold = tresholding.best_treshold(yProba_conca, yValid_conca, weights_conca) yPredicted_treshold = tresholding.get_yPredicted_treshold(yProba_conca, best_treshold) s, b = submission.get_s_b(yPredicted_treshold, yValid_conca, weights_conca) s *= 10 b *= 10 ams = hbc.AMS(s,b) if ams > best_AMS_1_method: best_AMS_1_method = ams best_method = dCombine['method'][i] best_ratio = best_treshold # Best treshold group by group for i in range(8): best_treshold = tresholding.best_treshold(yProba_s[i], valid_s[2][i], valid_s[3][i]) yPredicted_s[i] = tresholding.get_yPredicted_treshold(yProba_s[i], best_treshold) s, b = submission.get_s_b(yPredicted_s[i], valid_s[2][i], valid_s[3][i]) s *= 250000/yPredicted_s[i].shape[0] b *= 250000/yPredicted_s[i].shape[0] ams = hbc.AMS(s,b) if ams > best_AMS_s[i]: best_yPredicted_s[i] = yPredicted_s[i] best_yProba_s[i] = yProba_s[i] best_AMS_s[i] = ams best_method_s[i] = dCombine['method'][i] best_ratio_s[i] = best_treshold # FOR EACH METHOD: for method in dMethods: yProba_s = dMethods[method]['yProba_s'] yPredicted_s = dMethods[method]['yPredicted_s'] #Let's concatenate the vectors yProba_conca = preTreatment.concatenate_vectors(yProba_s) yPredicted_conca = preTreatment.concatenate_vectors(yPredicted_s) # Best treshold global best_treshold = tresholding.best_treshold(yProba_conca, yValid_conca, weights_conca) yPredicted_treshold = tresholding.get_yPredicted_treshold(yProba_conca, best_treshold) s, b = submission.get_s_b(yPredicted_treshold, yValid_conca, weights_conca) s *= 10 b *= 10 ams = hbc.AMS(s,b) if ams > best_AMS_1_method: best_AMS_1_method = ams best_method = str(method) best_ratio = best_treshold # Best treshold group by group for i in range(8): best_treshold = tresholding.best_treshold(yProba_s[i], valid_s[2][i], valid_s[3][i]) yPredicted_s[i] = tresholding.get_yPredicted_treshold(yProba_s[i], best_treshold) s, b = submission.get_s_b(yPredicted_s[i], valid_s[2][i], valid_s[3][i]) s *= 250000/yPredicted_s[i].shape[0] b *= 250000/yPredicted_s[i].shape[0] ams = hbc.AMS(s,b) if ams > best_AMS_s[i]: best_yPredicted_s[i] = yPredicted_s[i] best_yProba_s[i] = yProba_s[i] best_AMS_s[i] = ams best_method_s[i] = str(method) best_ratio_s[i] = best_treshold # Let's concatenate the 8 vectors which performs the best on each on # each of the sub group and tresholding it best_yPredicted_conca = preTreatment.concatenate_vectors(best_yPredicted_s) best_treshold_conca = tresholding.best_treshold(best_yPredicted_conca, yValid_conca, weights_conca) best_yPredicted_conca_treshold = tresholding.get_yPredicted_treshold(best_yPredicted_conca, best_treshold_conca) best_final_s, best_final_b, best_s_s, best_b_s = submission.get_s_b_8(best_yPredicted_s, valid_s[2], valid_s[3]) best_s_treshold, best_b_treshold = submission.get_s_b(best_yPredicted_conca_treshold, yValid_conca, weights_conca) best_final_s *= 10 best_final_b *= 10 best_s_treshold *= 10 best_b_treshold *= 10 best_AMS = hbc.AMS(best_final_s, best_final_b) best_AMS_treshold = hbc.AMS(best_s_treshold, best_b_treshold) print "Best AMS using one of the methods : %f" %best_AMS_1_method print " method : %s" %(str(method)) print " ratio : %f" %(best_ratio) print " " print "Best AMS final : %f" %best_AMS print "Best AMS final after final tresholding : %f" %best_AMS_treshold print "best ratio on the concatenated vector : %f" %best_treshold_conca print " " for n in range(8): print "Best AMS group %i: %f - method %s - ratio %f" \ %(n, best_AMS_s[n], best_method_s[n], best_ratio_s[n]) return best_yPredicted_s, valid_s
predProba_Valid2_s = xgBoost.predict_proba(predictor_s, valid_RM_s_2[1]) # Thresholding the predictions: predProba_Valid2 = preTreatment.concatenate_vectors(predProba_Valid2_s) predLabel5_Valid2 = tresholding.get_yPredicted_ratio(predProba_Valid2, best_ratio) # Binarize the prediction: predLabel_Valid2 = preTreatment.multiclass2binary(predLabel5_Valid2) # Concatenate data: yValid2 = preTreatment.concatenate_vectors(valid_RM_s_2[2]) weightsValidation = preTreatment.concatenate_vectors(valid_RM_s_2[3]) # Estimation the AMS: s, b = submission.get_s_b(predLabel_Valid2, yValid2, weightsValidation) s *= 250000/predLabel_Valid2.shape[0] b *= 250000/predLabel_Valid2.shape[0] ams = hbc.AMS(s,b) print "Valid_RM_2 - ratio : %f - best ams : %f" %(best_ratio, ams) print(" ") # Saving the model if it's better: if ams > best_ams: best_ams = ams best_imp_lim = importance_lim best_best_ratio = best_ratio best_n_removeFeatures = n_removeFeatures best_predictor_s = predictor_s
def train(max_depth, n_rounds): ############### ### IMPORT #### ############### # Importation parameters: split= True normalize = True noise_var = 0. train_size = 200000 train_size2 = 25000 valid_size = 25000 remove_999 = False # Import the training data: print("Extracting the data sets...") start = time.clock() train_s, train2_s, valid_s, test_s = tokenizer.extract_data(split= split, \ normalize= normalize, \ remove_999 = remove_999, \ noise_variance= noise_var, \ n_classes = "multiclass", \ train_size = train_size, \ train_size2 = train_size2, \ valid_size = valid_size) #RANDOM FOREST: #kwargs_grad = {} #kwargs_rdf = {'n_estimators': 100} print "Training on the train set ..." #predictor_s = randomForest.train_classifier(train_s[1], train_s[2], kwargs_rdf) #XGBOOST kwargs_xgb = {'bst_parameters': \ {'booster_type': 0, \ #'objective': 'binary:logitraw', 'objective': 'multi:softprob', 'num_class': 5, 'bst:eta': 0.1, # the bigger the more conservative 'bst:subsample': 1, # prevent over fitting if <1 'bst:max_depth': max_depth, 'eval_metric': 'auc', 'silent': 1, 'nthread': 8 }, \ 'n_rounds': n_rounds} predictor_s = xgBoost.train_classifier(train_s[1], train_s[2], train_s[3], 550000, kwargs_xgb) #TEST / SUBMISSION """ yProbaTest_s = [] yProbaTestBinary_s = [] print "Classifying the test set..." for i in range(8): yProbaTest = xgBoost.predict_proba(predictor_s[i], test_s[1][i]) yProbaTest_s.append(yProbaTest) print "Making the binary proba vector..." for i in range(8): yProbaTestBinary_s.append(np.zeros(yProbaTest_s[i].shape[0])) for i in range(8): for j in range(yProbaTest_s[i].shape[0]): yProbaTestBinary_s[i][j] = 1 - yProbaTest_s[i][j][0] print "Concatenating the vectors..." yProbaTestBinary = preTreatment.concatenate_vectors(yProbaTestBinary_s) IDs = preTreatment.concatenate_vectors(test_s[0]) yProbaTestBinaryRanked = submission.rank_signals(yProbaTestBinary) yPredictedTest = tresholding.get_yPredicted_ratio(yProbaTestBinary, 0.15) s = submission.print_submission(IDs, yProbaTestBinaryRanked, yPredictedTest, "newAMSmesure") """ # TRAIN AND VALID yPredictedTrain2_s = [] yProbaTrain2_s = [] yProbaTrain2Binary_s = [] yPredictedValid_s = [] yProbaValid_s = [] yProbaValidBinary_s = [] print "Classifying the train2 set..." for i in range(8): yProbaTrain2 = xgBoost.predict_proba(predictor_s[i], train2_s[1][i]) yProbaTrain2_s.append(yProbaTrain2) print "Classifying the valid set..." for i in range(8): yProbaValid = xgBoost.predict_proba(predictor_s[i], valid_s[1][i]) yProbaValid_s.append(yProbaValid) print "Making the binary proba vector..." for i in range(8): yProbaTrain2Binary_s.append(np.zeros(yProbaTrain2_s[i].shape[0])) yProbaValidBinary_s.append(np.zeros(yProbaValid_s[i].shape[0])) for i in range(8): for j in range(yProbaTrain2_s[i].shape[0]): yProbaTrain2Binary_s[i][j] = 1 - yProbaTrain2_s[i][j][0] for j in range(yProbaValid_s[i].shape[0]): yProbaValidBinary_s[i][j] = 1 - yProbaValid_s[i][j][0] print "Concatenating the vectors..." yProbaTrain2Binary = preTreatment.concatenate_vectors(yProbaTrain2Binary_s) yProbaValidBinary = preTreatment.concatenate_vectors(yProbaValidBinary_s) yTrain2 = preTreatment.concatenate_vectors(train2_s[2]) yValid = preTreatment.concatenate_vectors(valid_s[2]) weightsTrain2 = preTreatment.concatenate_vectors(train2_s[3]) weightsValid = preTreatment.concatenate_vectors(valid_s[3]) print "Putting all the real labels to 1" yTrain2 = preTreatment.multiclass2binary(yTrain2) yValid = preTreatment.multiclass2binary(yValid) print "Getting the best ratios..." best_ams_train2_global, best_ratio_global = tresholding.best_ratio(yProbaTrain2Binary, yTrain2, weightsTrain2) #best_ams_train2_combinaison, best_ratio_combinaison = tresholding.best_ratio_combinaison_global(yProbaTrain2Binary_s, train2_s[2], train2_s[3], 1) yPredictedValid = tresholding.get_yPredicted_ratio(yProbaValidBinary, 0.15) yPredictedValid_best_ratio_global = tresholding.get_yPredicted_ratio(yProbaValidBinary, best_ratio_global) #yPredictedValid_best_ratio_combinaison_s, yPredictedValid_best_ratio_combinaison = tresholding.get_yPredicted_ratio_8(yProbaTrain2Binary_s, best_ratio_combinaison) #Let's compute the predicted AMS s, b = submission.get_s_b(yPredictedValid, yValid, weightsValid) AMS = hbc.AMS(s,b) #s_best_ratio_combinaison, b_best_ratio_combinaison = submission.get_s_b(yPredictedValid_best_ratio_combinaison, yValid, weightsValid) #AMS_best_ratio_combinaison = hbc.AMS(s_best_ratio_combinaison, b_best_ratio_combinaison) s_best_ratio_global, b_best_ratio_global = submission.get_s_b(yPredictedValid_best_ratio_global, yValid, weightsValid) AMS_best_ratio_global = hbc.AMS(s_best_ratio_global, b_best_ratio_global) print "AMS 0.15 = %f" %AMS print " " #print "AMS best ratio combi= %f" %AMS_best_ratio_combinaison #print "best AMS train2 ratio combinaison= %f" %best_ams_train2_combinaison #print "best ratio combinaison train 2 = %s" %str(best_ratio_combinaison) print " " print "best AMS valid ratio global= %f" %AMS_best_ratio_global print "best AMS train2 ratio global= %f" %best_ams_train2_global print "best ratio global train2 = %f" %best_ratio_global return AMS
def analyse(train_s, train2_s, valid_s, method_name, kwargs={}): """ methode name = string, name of the method (eg :"naiveBayes") kwargs = dictionnary of the paraters of the method train_s = training set for the classifier(s) train2_s = training set for the meta parameters (eg : the best treshold) valid_s : validation set None of the set must be empty ! """ # Prediction on the validation set: print("------------------- Analyse: %s -----------------------") \ %(method_name) classifier_s = eval(method_name).train_classifier(train_s[1], train_s[2], kwargs) yProbaTrain2_s = eval(method_name).predict_proba(classifier_s, train2_s[1]) yProbaValid_s = eval(method_name).predict_proba(classifier_s, valid_s[1]) # Convert the validations vectors four 's' classes into one single s # classe if type(valid_s[2]) == list: for i in range(len(valid_s[2])): for j in range(valid_s[2][i].shape[0]): if valid_s[2][i][j] >=1: valid_s[2][i][j] = 1 # Convert the train2 vectors four 's' classes into one single s # classe if type(train2_s[2]) == list: for i in range(len(train2_s[2])): for j in range(train2_s[2][i].shape[0]): if train2_s[2][i][j] >=1: train2_s[2][i][j] = 1 # Let's define the vectors of probabilities of being 's' # Train2 set if type(yProbaTrain2_s) == list: yProbaTrain2Binary_s = [] for i in range(8): yProbaTrain2Binary_s.append(np.zeros(len(yProbaTrain2_s[i][:,1]))) for i in range(8): for j in range(len(yProbaTrain2_s[i][:,1])): yProbaTrain2Binary_s[i][j] = 1 - yProbaTrain2_s[i][j][0] else: yProbaTrain2Binary_s = np.zeros(len(yProbaTrain2_s[i][:,1])) for j in range(len(yProbaTrain2_s[i][:,1])): yProbaTrain2Binary_s[j] = 1 - yProbaTrain2_s[j][0] # Validation set if type(yProbaValid_s) == list: yProbaValidBinary_s = [] for i in range(8): yProbaValidBinary_s.append(np.zeros(len(yProbaValid_s[i][:,1]))) for i in range(8): for j in range(len(yProbaValid_s[i][:,1])): yProbaValidBinary_s[i][j] = 1 - yProbaValid_s[i][j][0] else: yProbaValidBinary_s = np.zeros(len(yProbaValid_s[i][:,1])) for j in range(len(yProbaValid_s[i][:,1])): yProbaValidBinary_s[j] = 1 - yProbaValid_s[j][0] # If we work with lists, let's get the concatenated vectors: # TRAIN SET if type(train_s[3]) ==list: weightsTrain_conca = preTreatment.concatenate_vectors(train_s[3]) else: weightsTrain_conca = train_s[3] # VALID SET # Validation Vectors if type(valid_s[2]) == list: yValid_conca = preTreatment.concatenate_vectors(valid_s[2]) else: yValid_conca = valid_s[2] # Weights Vectors if type(valid_s[3]) == list: weightsValid_conca = preTreatment.concatenate_vectors(valid_s[3]) else: weightsValid_conca = valid_s[3] # Binary Proba Vectors if type(yProbaValidBinary_s) == list: yProbaValidBinary_conca = preTreatment.concatenate_vectors( yProbaValidBinary_s) else: yProbaValidBinary_conca = yProbaValidBinary_s # All Proba Vectors if type(yProbaValid_s) == list: yProbaValid_conca = preTreatment.concatenate_vectors(yProbaValid_s) else: yProbaValid_conca = yProbaValid_s #TRAIN2 SET # Validation Vectors if type(train2_s[2]) == list: yTrain2_conca = preTreatment.concatenate_vectors(train2_s[2]) else: yTrain2_conca = train2_s[2] # Weights Vectors if type(train2_s[3]) == list: weightsTrain2_conca = preTreatment.concatenate_vectors(train2_s[3]) else: weightsTrain2_conca = train2_s[3] # Binary Proba Vectors if type(yProbaTrain2Binary_s) == list: yProbaTrain2Binary_conca = preTreatment.concatenate_vectors( yProbaTrain2Binary_s) else: yProbaTrain2Binary_conca = yProbaTrain2Binary_s # All Proba Vectors if type(yProbaTrain2_s) == list: yProbaTrain2_conca = preTreatment.concatenate_vectors(yProbaTrain2_s) else: yProbaTrain2_conca = yProbaTrain2_s # Let's rebalance the weight so their sum is equal to the total sum # of the train set sumWeightsTotal = sum(weightsTrain_conca)+sum(weightsTrain2_conca)+sum(weightsValid_conca) weightsTrain2_conca *= sumWeightsTotal/sum(weightsTrain2_conca) weightsValid_conca *= sumWeightsTotal/sum(weightsValid_conca) for i in range(8): train2_s[3][i] *= sumWeightsTotal/sum(weightsTrain2_conca) valid_s[3][i] *= sumWeightsTotal/sum(weightsValid_conca) # Let's get the best global treshold on the train2 set AMS_treshold_train2, best_treshold_global = tresholding.\ best_treshold(yProbaTrain2Binary_conca, yTrain2_conca, weightsTrain2_conca) yPredictedValid_conca_treshold = tresholding.get_yPredicted_treshold( yProbaValidBinary_conca, best_treshold_global) # Let's get the best ratio treshold on the train2 set AMS_ratio_global_train2, best_ratio_global = tresholding.\ best_ratio(yProbaTrain2Binary_conca, yTrain2_conca, weightsTrain2_conca) yPredictedValid_conca_ratio_global = tresholding.get_yPredicted_ratio( yProbaValidBinary_conca, best_ratio_global) # Let's get the best ratios combinaison if type(train_s[2]) == list: AMS_ratio_combinaison_train2, best_ratio_combinaison = tresholding.\ best_ratio_combinaison_global( yProbaTrain2Binary_s, train2_s[2], train2_s[3], 30) yPredictedValid_ratio_comb_s, yPredictedValid_conca_ratio_combinaison =\ tresholding.get_yPredicted_ratio_8( yProbaValidBinary_s, best_ratio_combinaison) # Let's compute the final s and b for each method s_treshold, b_treshold = submission.get_s_b( yPredictedValid_conca_treshold, yValid_conca, weightsValid_conca) s_ratio_global, b_ratio_global = submission.get_s_b( yPredictedValid_conca_ratio_global, yValid_conca, weightsValid_conca) if type(train_s[2]) == list: s_ratio_combinaison, b_ratio_combinaison = submission.get_s_b( yPredictedValid_conca_ratio_combinaison, yValid_conca, weightsValid_conca) # AMS final: AMS_treshold_valid = hbc.AMS(s_treshold, b_treshold) AMS_ratio_global_valid = hbc.AMS(s_ratio_global, b_ratio_global) if type(train_s[2]) == list: AMS_ratio_combinaison_valid = hbc.AMS(s_ratio_combinaison, b_ratio_combinaison) """ #AMS by group: if type(train_s[2]) == list: AMS_s = [] for i, (s,b) in enumerate(zip(s_s, b_s)): s *= 250000/yPredictedValid_s[i].shape[0] b *= 250000/yPredictedValid_s[i].shape[0] score = hbc.AMS(s,b) AMS_s.append(score) """ # Classification error: classif_succ_treshold = eval(method_name).get_classification_error( yPredictedValid_conca_treshold, yValid_conca, normalize= True) classif_succ_ratio_global = eval(method_name).get_classification_error( yPredictedValid_conca_ratio_global, yValid_conca, normalize= True) classif_succ_ratio_combinaison = eval(method_name).get_classification_error( yPredictedValid_conca_ratio_combinaison, yValid_conca, normalize= True) # Numerical score: """ if type(yProbaValid_s) == list: sum_s_treshold_s = [] sum_b_treshold_s = [] sum_s_ratio_global_s = [] sum_b_ratio_global_s = [] sum_s_ratio_combinaison_s = [] sum_b_ratio_combinaison_s = [] for i in range(len(yPredictedValid_s)): # treshold sum_s_treshold, sum_b_treshold = submission.get_numerical_score(yPredictedValid_conca_treshold_s[i], valid_s[2][i]) sum_s_treshold_s.append(sum_s) sum_b_treshold_s.append(sum_b) # ratio global sum_s_ratio_global, sum_b_ratio_global = submission.get_numerical_score(yPredictedValid_conca_ratio_global_s[i], valid_s[2][i]) sum_s_ratio_global_s.append(sum_s_ratio_global) sum_b_ratio_global_s.append(sum_b_ratio_global) # ratio combinaison sum_s_ratio_combinaison, sum_b_ratio_combinaison = submission.get_numerical_score(yPredictedValid_conca_ratio_combinaison_s[i], valid_s[2][i]) sum_s_ratio_combinaison_s.append(sum_s_ratio_combinaison) sum_b_ratio_combinaison_s.append(sum_b_ratio_combinaison) else: sum_s, sum_b = submission.get_numerical_score(yPredictedValid_s, valid_s[2]) """ d = {'classifier_s':classifier_s, 'yPredictedValid_conca_treshold': yPredictedValid_conca_treshold, 'yPredictedValid_conca_ratio_global' : \ yPredictedValid_conca_ratio_global, 'yProbaTrain2_s': yProbaTrain2_s, 'yProbaTrain2Binary_s': yProbaTrain2Binary_s, 'yProbaTrain2_conca': yProbaTrain2_conca, 'yProbaTrain2Binary_conca': yProbaTrain2Binary_conca, 'yProbaValid_s':yProbaValid_s, 'yProbaValidBinary_s':yProbaValidBinary_s, 'yProbaValid_conca':yProbaValid_conca, 'yProbaValidBinary_conca': yProbaValidBinary_conca, 'AMS_treshold_train2':AMS_treshold_train2, 'AMS_ratio_global_train2':AMS_ratio_global_train2, 'AMS_treshold_valid':AMS_treshold_valid, 'AMS_ratio_global_valid':AMS_ratio_global_valid, 'best_treshold_global' : best_treshold_global, 'best_ratio_global':best_ratio_global, 'classif_succ_treshold': classif_succ_treshold, 'classif_succ_ratio_global': classif_succ_ratio_global, 'method': method_name, 'parameters': kwargs} if type(train_s[2])==list: d['yPredictedValid_conca_ratio_combinaison'] = yPredictedValid_conca_ratio_combinaison d['AMS_ratio_combinaison_train2'] = AMS_ratio_combinaison_train2 d['AMS_ratio_combinaison_valid'] = AMS_ratio_combinaison_valid, d['best_ratio_combinaison'] = best_ratio_combinaison, d['classif_succ_ratio_combinaison'] = classif_succ_ratio_combinaison return d
def main(): ############### ### IMPORT #### ############### # Importation parameters: split= True normalize = True noise_var = 0. n_classes = "binary" train_size = 200000 train_size2 = 25000 valid_size = 25000 # Import the training data: print("Extracting the data sets...") start = time.clock() train_s, train2_s, valid_s, test_s = tokenizer.extract_data(split= split, normalize= normalize, noise_variance= noise_var, n_classes = n_classes, train_size = train_size, train_size2 = train_size2, valid_size = valid_size) # Remerging the y and weights of the validation if necessary: if type(valid_s[2]) == list: yValid_conca = preTreatment.concatenate_vectors(valid_s[2]) weights_conca = preTreatment.concatenate_vectors(valid_s[3]) stop = time.clock() print ("Extraction time: %i s") %(stop-start) print(" ") print(" ") ###################### ### PRE-TREATMENT #### ###################### print("------------------------- Pre-treatment --------------------------") ### Average number of signal per subset: print("Train subsets signal average:") train_s_average = preTreatment.ratio_sig_per_dataset(train_s[2]) print(" ") print("Valid subsets signal average:") valid_s_average = preTreatment.ratio_sig_per_dataset(valid_s[2]) print(" ") print(" ") ############ # ANALYSES # ############ # Dictionnary that will contain all the data for each methods. In the end # we'll have a dict of dict # Keys of the methods : {naiveBayes, svm, kNeighbors, lda, qda, adaBoost, # randomForest} dMethods ={} # NAIVE BAYES: kwargs_bayes = {} dMethods['naiveBayes'] = analyse.analyse(train_s= train_s, train2_s= train2_s, valid_s= valid_s, method_name = 'naiveBayes', kwargs = kwargs_bayes) # SVM """ kwargs_svm ={} dMethods['svm'] = analyse.analyse(train_s, valid_s,'svm', kwargs_svm) """ """ # K NEIGHBORS kwargs_kn = {'n_neighbors':50} dMethods['kNeighbors'] = analyse.analyse(train_s, valid_s, 'kNeighbors', kwargs_kn) """ # LDA kwargs_lda = {} dMethods['lda'] = analyse.analyse(train_s= train_s, train2_s= train2_s, valid_s= valid_s, method_name = 'lda', kwargs = kwargs_lda) # QDA kwargs_qda= {} dMethods['qda'] = analyse.analyse(train_s= train_s, train2_s= train2_s, valid_s= valid_s, method_name = 'qda', kwargs = kwargs_qda) """ # ADABOOST kwargs_ada= { 'n_estimators': 50, 'learning_rate': 1., 'algorithm': 'SAMME.R', 'random_state':None} dMethods['adaBoost'] = analyse.analyse(train_s, valid_s, 'adaBoost', kwargs_ada) """ # RANDOM FOREST: kwargs_randomForest= {'n_estimators': 10} dMethods['randomForest'] = analyse.analyse(train_s= train_s, train2_s= train2_s, valid_s= valid_s, method_name = 'randomForest', kwargs = kwargs_randomForest) # RANDOM FOREST 2: kwargs_randomForest= {'n_estimators': 100} dMethods['randomForest2'] = analyse.analyse(train_s= train_s, train2_s= train2_s, valid_s= valid_s, method_name = 'randomForest', kwargs = kwargs_randomForest) """ # ADABOOST2 kwargs_ada= { 'n_estimators': 100, 'learning_rate': .5, 'algorithm': 'SAMME.R', 'random_state':None} dMethods['adaBoost2'] = analyse.analyse(train_s, valid_s, 'adaBoost', kwargs_ada) # RANDOM FOREST 3: kwargs_randomForest= {'n_estimators': 100} dMethods['randomForest3'] = analyse.analyse(train_s= train_s, train2_s= train2_s, valid_s= valid_s, method_name = 'randomForest', kwargs = kwargs_randomForest) # RANDOM FOREST 4: kwargs_randomForest= {'n_estimators': 100} dMethods['randomForest4'] = analyse.analyse(train_s= train_s, train2_s= train2_s, valid_s= valid_s, method_name = 'randomForest', kwargs = kwargs_randomForest) # RANDOM FOREST 5: kwargs_randomForest= {'n_estimators': 100} dMethods['randomForest5'] = analyse.analyse(train_s= train_s, train2_s= train2_s, valid_s= valid_s, method_name = 'randomForest', kwargs = kwargs_randomForest) # GRADIENT BOOSTING: kwargs_gradB = {'loss': 'deviance', 'learning_rate': 0.1, 'n_estimators': 100, 'subsample': 1.0, 'min_samples_split': 2, 'min_samples_leaf': 200, 'max_depth': 10, 'init': None, 'random_state': None, 'max_features': None, 'verbose': 0} dMethods['gradientBoosting'] = analyse.analyse(train_s, valid_s, 'gradientBoosting', kwargs_gradB) """ print(" ") ################## # POST-TREATMENT # ################## print("------------------------ Feaure importance: -----------------------") if type(dMethods['randomForest2']['predictor_s']) == list: for i,predictor_s in enumerate(dMethods['randomForest2']['predictor_s']): print "Subset %i:" %i print predictor_s.feature_importances_ else: print "Dataset: " print dMethods['randomForest2']['predictor_s'].feature_importances_ print("------------------------ On-top predictor -----------------------") # Classifiers to be ignored: #ignore = ['randomForest2', 'randomForest'] ignore = [] clf_onTop = 'randomForest' parameters = {}#{'C': 0.5, 'kernel': 'rbf', 'degree': 3, 'gamma': 0.0, # 'coef0': 0.0, 'shrinking':True, 'probability':True, # 'tol': 0.001, 'cache_size': 200, 'class_weight': None} print ("We will use an 'on-top' predictor on %i classifiers using a %s.") \ %(len(dMethods.keys())-len(ignore), clf_onTop) final_prediction_s, dOnTop = onTopClassifier.SL_classification(dMethods, valid_s, train_s, ignore = ignore, method= clf_onTop, parameters= parameters) print("-------------------------- Tresholding -------------------------") ### ON THE 'ON-TOP' CLASSIFIER: # Create the elected vectors for each group (best AMS score) OT_best_yPredicted_s = [np.zeros(valid_s[2][i].shape[0]) for i in range(8)] OT_best_yProba_s = [np.zeros(valid_s[2][i].shape[0]) for i in range(8)] OT_best_AMS_s = [0. for i in range(8)] OT_best_method_s = [0 for i in range(8)] OT_best_ratio_s = [0 for i in range(8)] OT_best_sum_s_s = [0 for i in range(8)] OT_best_sum_b_s = [0 for i in range(8)] OT_best_method = "On-top" OT_yProba_s = dOnTop['yProba_s'] OT_yPredicted_s = dOnTop['yPredicted_s'] #Let's concatenate the vectors OT_yProba_conca = preTreatment.concatenate_vectors(OT_yProba_s) OT_yPredicted_conca = preTreatment.concatenate_vectors(OT_yPredicted_s) # Best treshold global OT_best_ratio = tresholding.best_treshold(OT_yProba_conca, yValid_conca, weights_conca) OT_yPredicted_treshold = tresholding.get_yPredicted_treshold(OT_yProba_conca, OT_best_ratio) OT_s, OT_b = submission.get_s_b(OT_yPredicted_treshold, yValid_conca, weights_conca) OT_s *= 10 OT_b *= 10 OT_best_AMS = hbc.AMS(OT_s,OT_b) # COMPARISON BEST TRESHOLD IN DMETHOD # FOR EACH METHOD: best_yPredicted_s = [np.zeros(valid_s[2][i].shape[0]) for i in range(8)] best_yProba_s = [np.zeros(valid_s[2][i].shape[0]) for i in range(8)] best_AMS_s = [0. for i in range(8)] best_method_s = [0 for i in range(8)] best_ratio_s = [0 for i in range(8)] best_AMS_1_method = 0. best_method = "methode" best_ratio = "0." for method in dMethods: yProba_s = dMethods[method]['yProba_s'] yPredicted_s = dMethods[method]['yPredicted_s'] #Let's concatenate the vectors yProba_conca = preTreatment.concatenate_vectors(yProba_s) yPredicted_conca = preTreatment.concatenate_vectors(yPredicted_s) # Best treshold global best_treshold = tresholding.best_treshold(yProba_conca, yValid_conca, weights_conca) yPredicted_treshold = tresholding.get_yPredicted_treshold(yProba_conca, best_treshold) s, b = submission.get_s_b(yPredicted_treshold, yValid_conca, weights_conca) s *= 10 b *= 10 ams = hbc.AMS(s,b) if ams > best_AMS_1_method: best_AMS_1_method = ams best_method = str(method) best_ratio = best_treshold # Let's concatenate the 8 vectors which performs the best on each on # each of the sub group and tresholding it best_yPredicted_conca = preTreatment.concatenate_vectors(best_yPredicted_s) best_treshold_conca = tresholding.best_treshold(best_yPredicted_conca, yValid_conca, weights_conca) best_yPredicted_conca_treshold = tresholding.get_yPredicted_treshold(best_yPredicted_conca, best_treshold_conca) best_final_s, best_final_b, best_s_s, best_b_s = submission.get_s_b_8(best_yPredicted_s, valid_s[2], valid_s[3]) best_s_treshold, best_b_treshold = submission.get_s_b(best_yPredicted_conca_treshold, yValid_conca, weights_conca) best_final_s *= 10 best_final_b *= 10 best_s_treshold *= 10 best_b_treshold *= 10 best_AMS = hbc.AMS(best_final_s, best_final_b) best_AMS_treshold = hbc.AMS(best_s_treshold, best_b_treshold) print "Best AMS using one of the methods : %f" %best_AMS_1_method print " method : %s" %(str(method)) print " ratio : %f" %(best_ratio) print " " print "Best AMS concatenate: %f" %best_AMS print "Best AMS concatenate after final tresholding : %f" %best_AMS_treshold print "best ratio on the concatenated vector : %f" %best_treshold_conca print " " print "Best AMS on-top : %f" %OT_best_AMS print "Best ratio on the concatenated vector : %f" %OT_best_ratio print " " """ # Best treshold group by group for i in range(8): OT_best_treshold_s = tresholding.best_treshold(OT_yProba_s[i], valid_s[2][i], valid_s[3][i]) OT_yPredicted_s[i] = tresholding.get_yPredicted_treshold(OT_yProba_s[i], OT_best_treshold_s) s, b = submission.get_s_b(OT_yPredicted_s[i], valid_s[2][i], valid_s[3][i]) s *= 250000/yPredicted_s[i].shape[0] b *= 250000/yPredicted_s[i].shape[0] ams = hbc.AMS(s,b) if ams > best_AMS_s[i]: best_yPredicted_s[i] = yPredicted_s[i] best_yProba_s[i] = yProba_s[i] best_AMS_s[i] = ams best_method_s[i] = dOnTop['method'] best_ratio_s[i] = best_treshold best_sum_s_s[i] = s best_sum_b_s[i] = b for n in range(8): print "Best AMS group %i: %f - method %s - ratio %f" \ %(n, best_AMS_s[n], best_method_s[n], best_ratio_s[n]) print "Best AMS : %f" %best_AMS_1_method print " ratio : %f" %(best_ratio) print " " """ """ ############## # SUBMISSION # ############## print("-------------------------- Submission ---------------------------") # Prediction on the test set: # method used for the submission # TODO : Verifier que le nom de la method a bien la bonne forme( # creer une liste de noms de methodes) #method = "randomForest" #test_prediction_s, test_proba_s = eval(method).get_test_prediction( # dMethods[method]['predictor_s'], # test_s[1]) test_prediction_s, test_proba_s = onTopClassifier.get_SL_test_prediction( dMethods, dOnTop, test_s[1]) print("Test subsets signal average:") test_s_average = preTreatment.ratio_sig_per_dataset(test_prediction_s) print(" ") #RankOrder = np.arange(1,550001) if type(test_prediction_s) == list: test_prediction_s = np.concatenate(test_prediction_s) test_proba_s = np.concatenate(test_proba_s) RankOrder = onTopClassifier.rank_signals(test_proba_s) ID = np.concatenate(test_s[0]) else: ID = test_s[0] # Create a submission file: sub = submission.print_submission(ID, RankOrder , test_prediction_s) return sub """ return 0