Python metrics.log_loss函数代码示例

def xgboostcv(max_depth,
              eta,
              num_rounds,
              gamma,
              min_child_weight,
              max_delta_step,
              subsample,
              colsample_bytree,
              silent=True,
              seed=1234):
    
    print ('\nRunning XGBOOST on the cluster')
    
    # Call xgboost in distributed mode (CLI input for params)
    xgb_run = ['max_depth=%s' % int(max_depth),
               'eta=%s' % eta,
               'silent=%s' % silent,
               'gamma=%s' % gamma,
               'min_child_weight=%s' % int(min_child_weight),
               'max_delta_step=%s' % max_delta_step,
               'subsample=%s' % subsample,
               'eval_metric=logloss',
               'colsample_bytree=%s' % colsample_bytree,
               'seed=%s' % seed,
               'objective=binary:logistic',
               'eval[eval_set]=%s' % deval,
               'eval[train_set]=%s' % dtrain,
               'num_round=%s' % int(num_rounds),
               'data=%s' % dtrain,
               'model_out=%s' % model_ouput]
    argv = ['wormhole/repo/dmlc-core/tracker/dmlc_yarn.py', # Where your instance is found!!
            '-n',
            '16',
            'wormhole/bin/xgboost.dmlc', # Where your instance is found!!
            './examples/xgboost-avazu.txt'] + xgb_run
    print(' '.join(argv))
    # Cluster specific ENV VARS.
    Popen(argv,
          env = {'JAVA_HOME': '/usr/lib/jvm/java-1.8.0-openjdk-1.8.0.45-28.b13.el6_6.x86_64/',
                 'HADOOP_HOME': '/usr/',
                 'HADOOP_HDFS_HOME': '/usr/lib/hadoop-hdfs/',
                 'PATH': os.getenv('PATH')}).communicate()

    # Export model to local filesystem
    try:
      os.remove("avazu.model")
    except OSError:
      pass
    Popen(["hadoop","fs","-copyToLocal","/tmp/avazu.model", "."]).communicate()
    # Delete stored model.
    Popen(["hadoop","fs","-rm","/tmp/avazu.model", "."]).communicate()
    
    # Load Model file
    bst = xgb.Booster(model_file='avazu.model')
    preds = bst.predict(dtest)
    y_pred = bst.predict(dtest)
    y_valid = dtest.get_label()
    print('logloss = ', log_loss(y_valid, y_pred))
    # We are maximizing the function.
    return -log_loss(y_valid, y_pred)

def learn(learning_rate, X_train, y_train, X_test, y_test):
	model = GradientBoostingClassifier(
		n_estimators=250,
		verbose=True,
		random_state=241,
		learning_rate=learning_rate
		)
	model.fit(X_train, y_train)
	
	# plot scores
	test_score = list(range(250))
	train_score = list(range(250))

	for i, predictions in enumerate(model.staged_decision_function(X_test)):
		predictions = [x[0] for x in predictions.tolist()] # unpack this stupid format
		predictions = [1/(1 + math.exp(-x)) for x in predictions]
		test_score[i] = log_loss(y_test, predictions)

	for i, predictions in enumerate(model.staged_decision_function(X_train)):
		predictions = [x[0] for x in predictions.tolist()] # unpack this stupid format
		predictions = [1/(1 + math.exp(-x)) for x in predictions]
		train_score[i] = log_loss(y_train, predictions)

	plt.figure()
	plt.plot(test_score, 'r', linewidth=2)
	plt.plot(train_score, 'g', linewidth=2)
	plt.legend(['test', 'train'])
	plt.show()
	
	return train_score, test_score

def fit_model_and_test(params):
    crimes = np.load(DATA_FILE)
    features_train = crimes['features_train']
    all_labels = sorted(list(set(np.unique(crimes['labels_train'])) | set(np.unique(crimes['labels_val']))))
    hidden_units = int(params['hidden_units'])
    batch_size = 64

    labels_train = create_labels(crimes['labels_train'], all_labels)
    labels_vals = create_labels(crimes['labels_val'], all_labels)
    labels_full = create_labels(crimes['labels'], all_labels)

    labels_train = np_utils.to_categorical(labels_train)
    labels_vals = np_utils.to_categorical(labels_vals)
    labels_full = np_utils.to_categorical(labels_full)

    model = create_model_and_fit(features_train,labels_train, hidden_units, len(all_labels),  params['layers'],
                                 params['input_dropout'], params['hidden_dropout'],
                                 batch_size, crimes['features_val'], labels_vals)

    loss_train = log_loss(labels_train, model.predict_proba(crimes['features_train']))
    loss_val = log_loss(labels_vals, model.predict_proba(crimes['features_val']))
    loss_all = log_loss(labels_full, model.predict_proba(crimes['features']))

    print 'loss_all: ', loss_all
    print 'loss_train: ', loss_train
    print 'loss_val: ', loss_val
    sys.stdout.flush()
    return loss_val, model, crimes, all_labels

def Test():
    """Testing ConstrainedMultinomialRegression
    
    Compare the results with scikit-learn LogisticRegression v.15
    
    Returns
    -------
    Log Loss for Logistic Regression, ConstrainedMultinomialRegression
    Accuracy for Logistic Regression, ConstrainedMultinomialRegression
    """
    n = 1000; p = 10; k = 3
    X = np.random.randn(n, p)
    beta = np.random.binomial(1, .5, (p, k))
    log_odd = X.dot(beta)
    prob = np.exp(log_odd)/(1 + np.exp(log_odd))
    y = np.array([np.argmax(i) for i in prob])
    lb = LabelBinarizer()
    Y = lb.fit_transform(y)
    w = randn(k,p)
    cut = n/2
    train = np.arange(cut); valid = np.arange(cut,n) # Split Train and Test
    b = [(0,None)]*(p+1)*k # Constraint on Beta
    cl1 = LogisticRegression()
    cl2 = ConstrainedMultinomialClassifier(bounds = b)
    cl1.fit(X[train], y[train])
    cl2.fit(X[train], y[train])
    prob1 = cl1.predict_proba(X[valid])
    prob2 = cl2.predict_proba(X[valid])
    print log_loss(y[valid], prob1)
    print log_loss(y[valid], prob2)
    yhat1 = cl1.predict(X[valid])
    yhat2 = cl2.predict(X[valid])
    print accuracy_score(y[valid], yhat1)
    print accuracy_score(y[valid], yhat2)

def modelfit(alg, dtrain, predictors, dtest=None, dscore=None, useTrainCV=True, cv_folds=5, early_stopping_rounds=50):

    if useTrainCV:
        xgb_param = alg.get_xgb_params()
        xgtrain = xgb.DMatrix(dtrain[predictors].values, label=dtrain[target].values)
        cvresult = xgb.cv(xgb_param, xgtrain, num_boost_round=alg.get_params()['n_estimators'], nfold=cv_folds,
            metrics=['logloss'], early_stopping_rounds=early_stopping_rounds, show_progress=False)
        alg.set_params(n_estimators=cvresult.shape[0])

    #Fit the algorithm on the data
    alg.fit(dtrain[predictors], dtrain['target'], eval_metric='logloss')

    #Predict training set:
    dtrain_predictions = alg.predict(dtrain[predictors])
    dtrain_predprob = alg.predict_proba(dtrain[predictors])[:,1]
    if isinstance(dtest, pd.DataFrame):
        dtest_predprob = alg.predict_proba(dtest[predictors])[:,1]
    if isinstance(dscore, pd.DataFrame):
        dscore_predprob = alg.predict_proba(dscore[predictors])[:,1]
        np.savetxt('XGBoost_pred_raw.csv', dscore_predprob, delimiter=",")

    #Print model report:
    print "\nModel Report"
    print "Accuracy : %.4g" % metrics.accuracy_score(dtrain['target'].values, dtrain_predictions)
    print "Metric Score (Train): %f" % metrics.log_loss(dtrain['target'], dtrain_predprob)
    if isinstance(dtest, pd.DataFrame):
        print "Metric Score (Test): %f" % metrics.log_loss(dtest['target'], dtest_predprob)

    feat_imp = pd.Series(alg.booster().get_fscore()).sort_values(ascending=False)
    feat_imp.plot(kind='bar', title='Feature Importances')
    plt.ylabel('Feature Importance Score')
    plt.show()

def generic_cv_reg(X,y,model,n_folds,random_state) :
    kf = cross_validation.KFold(y.shape[0],n_folds=n_folds, shuffle=True, random_state=random_state)
    trscores, cvscores, times = [], [], []
    i = 0
    stack_train = np.zeros((len(y))) # stacked predictions
    
    threshold = 0.000001
    
    for i, (train_fold, validate) in enumerate(kf) :
        i = i + 1
        t = time()
        trscore = log_loss(y.iloc[train_fold], model.fit(X.iloc[train_fold], y.iloc[train_fold]).predict(X.iloc[train_fold]))
        
        validation_prediction = model.predict(X.iloc[validate])
        
        validation_prediction[validation_prediction>1-threshold] = 1-threshold
        validation_prediction[validation_prediction<threshold] = threshold
        
        cvscore = log_loss(y.iloc[validate], validation_prediction)
        trscores.append(trscore); cvscores.append(cvscore); times.append(time()-t)
        
        stack_train[validate] = validation_prediction
    
    print("TRAIN %.5f | TEST %.5f | TIME %.2fm (1-fold)" % (np.mean(trscores), np.mean(cvscores), np.mean(times)/60))
    print(model.get_params(deep = True))
    print("\n")
    
    return np.mean(cvscores), stack_train

def cross_valid(X, y, params, iterations, n_folds=6, silent=True):
    print 'Running cross validation'
    pprint.pprint(params)
    print 'Iterations:', iterations
    print 'X shape', X.shape

    y_size = len(y)
    if hasattr(X, 'values'):
        X = X.values
    y = np.array(y)

    kf = cross_validation.KFold(y_size, n_folds=n_folds, shuffle=True,
                                random_state=params['seed'])
    y_pred = np.zeros((y_size, 9))

    logs = []
    for train, test in kf:
        X_train, X_test = X[train, :], X[test, :]
        y_train, y_test = y[train], y[test]

        predictions = predict(X_train, y_train, X_test, params, iterations,
                              None if silent else y_test)
        y_pred[test] = predictions

        logs.append(metrics.log_loss(y_test, predictions))
        print 'Current log_loss:', logs[-1]

    print 'Final log_loss: %s (avg: %s, stddev: %s)' % (
                                                metrics.log_loss(y, y_pred),
                                                np.mean(logs),
                                                np.std(logs))

def ctr_gbdt(model='sklearn-clicklog', from_cache=False, train_dataset_length=100000, test_dataset_length=100000):
    TRAIN_FILE, TEST_FILE = create_dataset(model, from_cache, train_dataset_length, test_dataset_length)

    prediction_model = GradientBoostingClassifier(
        loss='deviance',
        learning_rate=0.1,
        n_estimators=30,
        subsample=1.0,
        min_samples_split=2,
        min_samples_leaf=1,
        min_weight_fraction_leaf=0.0,
        max_depth=5,
    )

    x_train, y_train = clean_data(TRAIN_FILE)
    x_test, y_test = clean_data(TEST_FILE)

    with Timer('fit model'):
        prediction_model.fit(x_train, y_train)

    with Timer('evaluate model'):
        y_prediction_train = prediction_model.predict_proba(x_train)
        y_prediction_test = prediction_model.predict_proba(x_test)

        loss_train = log_loss(y_train, y_prediction_train)
        loss_test = log_loss(y_test, y_prediction_test)

    print 'loss_train: %s' % loss_train
    print 'loss_test: %s' % loss_test

def check_lambda(dirnm, datanm_train, datanm_valid, datanm_orig_train, datanm_orig_valid, samples_per_class, Cs, num_classes):
    spct = 10*70
    tdata, tlabels = load_full(dirnm+datanm_train, spct)
    print tdata.shape, tlabels.shape

    spct = 10
    otdata, otlabels = load_full(dirnm+datanm_orig_train, spct)

    spct = 10*30
    vdata, vlabels = load_full(dirnm+datanm_valid, spct)

    spct = 10
    ovdata, ovlabels = load_full(dirnm+datanm_orig_valid, spct)

    # artif
    ans = np.zeros((len(Cs), 4))

    for i, C in enumerate(Cs):
        clf = LogisticRegression(C  =C,     penalty='l2', multi_class = 'ovr',
                                 tol=0.001, n_jobs = -1, verbose = 0, solver = 'newton-cg')
        clf.fit(tdata, tlabels)

        out_train = clf.predict_proba(tdata)
        out_valid = clf.predict_proba(vdata)
        out_train_real = clf.predict_proba(otdata)
        out_valid_real = clf.predict_proba(ovdata)

        ans[i, 0] += log_loss(tlabels, out_train)
        ans[i, 1] += log_loss(vlabels, out_valid)
        ans[i, 2] += log_loss(otlabels, out_train_real)
        ans[i, 3] += log_loss(ovlabels, out_valid_real)

    np.savez("logreg_lambda", ans= ans, Cs = Cs, num_classes = num_classes, samples_per_class = samples_per_class)
    return ans

def check_lambda(datanm, samples_per_class,depv, num_classes, criterion, num_iter = 100):
    data, labels = load_full(datanm, samples_per_class)
    slo = StratifiedShuffleSplit(labels, n_iter=num_iter, test_size=0.3, train_size=0.7, random_state=None)
    ans = np.zeros((len(depv), 4))
    for train_index, test_index in slo:
        train_data = [data[train_index, :], labels[train_index]]
        valid_data = [data[test_index , :], labels[test_index ]]

        for i, d in enumerate(depv):
            clf = DecisionTreeClassifier(criterion=criterion, splitter='best',
                                         max_depth=d, min_samples_split=2,
                                         min_samples_leaf=1, min_weight_fraction_leaf=0.0,
                                         max_features=None, random_state=None,
                                         max_leaf_nodes=None, class_weight=None, presort=False)
            clf.fit(train_data[0], train_data[1])

            out_train = clf.predict_proba(train_data[0])
            out_valid = clf.predict_proba(valid_data[0])

            ans[i, 0] += log_loss(train_data[1], out_train)
            ans[i, 1] += log_loss(valid_data[1], out_valid)
            ans[i, 2] += brier(train_data[1], out_train, num_classes)
            ans[i, 3] += brier(valid_data[1], out_valid, num_classes)

    ans[:, :] /= num_iter

    np.savez("rand_forest_lambda_" + criterion, ans= ans, mdep = mdep, num_iter = num_iter, num_classes = num_classes, samples_per_class = samples_per_class)
    return ans

def svm_model(train_data_features, train_data_split_crossfold_features, test_data_features, labels, labels_cross_validation_classwise, using_cross_validation2, kf, settings):
    if using_cross_validation2:
        C_base = 4.5
        C_step = 0.5#0.005
        C = C_base
        _results = []
        if(len(train_data_cross_validation_classwise_features) > 0):
            """train_all = np.append(train_data_features, train_data_cross_validation_classwise_features, axis=0)
            labels_all = np.append(labels, labels_cross_validation_classwise)
            kf_all = KFold(len(train_all)-1, n_folds=int(settings['Data']['CrossValidation2']), shuffle=True)
            for train, test in kf_all:
                svc = SVC(kernel="linear", C=C, probability=True)
                model = svc.fit(train_all[train], labels_all[train])
                predicted_classes = model.predict(train_all[test])
                predicted_classes_train = model.predict(train_all[train])
                class_probabilities = model.predict_proba(train_all[test])
                print("C: ",C," n points:", len(predicted_classes), " percentage: ",(labels_all[test] != predicted_classes).sum()*100/len(predicted_classes),"% percentage_train: ", (labels_all[train] != predicted_classes_train).sum()*100/len(predicted_classes_train),"%")
                _results.append((labels_all[test] != predicted_classes).sum())
                C += C_step"""
            for c in pl.frange(C_base,9, C_step):
                svc = SVC(kernel="linear", C=c, probability=True)
                model = svc.fit(train_data_features, labels)
                predicted_classes = model.predict(train_data_cross_validation_classwise_features)
                class_probabilities = model.predict_proba(train_data_cross_validation_classwise_features)
                print("C: ",c," N points:", len(predicted_classes), " percentage: ",(labels_cross_validation_classwise != predicted_classes).sum()*100/len(predicted_classes),"%")
                print("Log_loss: ", log_loss(labels_cross_validation_classwise, class_probabilities))
            for c in pl.frange(1,3, 1):
                svc = SVC(kernel="linear", C=c, probability=True)
                model = svc.fit(train_data_features, labels)
                predicted_classes = model.predict(train_data_cross_validation_classwise_features)
                class_probabilities = model.predict_proba(train_data_cross_validation_classwise_features)
                print("C: ",c," N points:", len(predicted_classes), " percentage: ",(labels_cross_validation_classwise != predicted_classes).sum()*100/len(predicted_classes),"%")
                print("Log_loss: ", log_loss(labels_cross_validation_classwise, class_probabilities))
        else:
            for train, test in kf:
                svc = SVC(kernel="linear", C=C, probability=True)
                model = svc.fit(train_data_features[train], labels[train])
                predicted_classes = model.predict(train_data_features[test])
                predicted_classes_train = model.predict(train_data_features[train])
                class_probabilities = model.predict_proba(train_data_features[test])
                print("C: ",C," n points:", len(predicted_classes), " percentage: ",(labels[test] != predicted_classes).sum()*100/len(predicted_classes),"% percentage_train: ", (labels[train] != predicted_classes_train).sum()*100/len(predicted_classes_train),"%")
                _results.append((labels[test] != predicted_classes).sum())
                C += C_step

        C = C_base + C_step * _results.index(min(_results))
        print("C: ", C)
        if(len(train_data_cross_validation_classwise_features) > 0):
            svc = SVC(kernel="linear", C=C, probability=True)
            model = svc.fit(train_data_features, labels)
            predicted_classes = model.predict(train_data_cross_validation_classwise_features)
            class_probabilities = model.predict_proba(train_data_cross_validation_classwise_features)
            print("C: ",C," N points:", len(predicted_classes), " percentage: ",(labels_cross_validation_classwise != predicted_classes).sum()*100/len(predicted_classes),"%")
            print("Log_loss: ", log_loss(labels_cross_validation_classwise, class_probabilities))
        svc = SVC(kernel="linear", C=C, probability=True)
        model = svc.fit(train_data_features, labels)
        return model.predict_proba(test_data_features), model.predict(test_data_features), model
    else:
        svc = SVC(kernel="linear", C=8, probability=True)
        model = svc.fit(train_data_features, labels)
        return model.predict_proba(test_data_features), model.predict(test_data_features), model

def main():
    X, Y, encoder, scale = load_train_data('train.csv')
    estimators = 500
    X_train, X_valid, Y_train, Y_valid = train_test_split(X, Y, test_size=0.2, random_state=0)
    X_train_real, X_test_real, Y_train_real, Y_test_real = train_test_split(X_train, Y_train, test_size=0.2, random_state=42)
    log.info('Loaded training file')
    X_test, _ = load_csv_file('test.csv', cut_end=False)
    log.info('Loaded test file')

    #Classifier Setup
    tree_clf = ExtraTreesClassifier(n_estimators=estimators, n_jobs=-1,
                                    random_state=42, max_depth=55, min_samples_split=1)

    clf = make_pipeline(TfidfTransformer(), DenseTransformer(), tree_clf)
    log.info('Fitting GradientBoost')
    clf.fit(X_train_real, Y_train_real)
    clf_probs = clf.predict_proba(X_test_real)
    score = log_loss(Y_test_real, clf_probs)
    log.info('Log Loss score un-trained = %f' % score)
    # Calibrate Classifier using ground truth in X,Y_valid
    sig_clf = CalibratedClassifierCV(clf, method="isotonic", cv="prefit")
    log.info('Fitting CalibratedClassifierCV')
    sig_clf.fit(X_valid, Y_valid)
    sig_clf_probs = sig_clf.predict_proba(X_test_real)
    sig_score = log_loss(Y_test_real, sig_clf_probs)
    log.info('Log loss score trained = %f' % sig_score)

    # Ok lets predict the test data with our funky new classifier
    sig_submission_probs = sig_clf.predict_proba(X_test)

    write_out_submission(sig_submission_probs, 'submission.csv')

def xgb_base(train2, y, test2, v, z, xgb_params, N_splits, N_seeds, cname, base_seed=42):
    v[cname], z[cname] = 0, 0
    scores = []

    dtest = xgb.DMatrix(test2)
    for s in range(N_seeds):
        xgb_params['seed'] = s + base_seed
        skf = model_selection.StratifiedKFold(n_splits=N_splits, shuffle=True, random_state=s + base_seed)
        for n, (itrain, ival) in enumerate(skf.split(train2, y)):
            dtrain = xgb.DMatrix(train2.ix[itrain], y[itrain])
            dvalid = xgb.DMatrix(train2.ix[ival], y[ival])
            watch = [(dtrain, 'train'), (dvalid, 'valid')]
            clf = xgb.train(xgb_params, dtrain, 10000, watch, early_stopping_rounds=100, verbose_eval=False)

            p = clf.predict(dvalid)
            v.loc[ival, cname] += pconvert(p)
            score = metrics.log_loss(y[ival], p)
            z[cname]  += pconvert(clf.predict(dtest))
            print(cname, 'seed %d step %d of %d: '%(xgb_params['seed'], n+1, skf.n_splits), score, now())
            scores.append(score)

    z[cname] /= N_splits * N_seeds
    v[cname] /= N_seeds
    print('validation loss: ', metrics.log_loss(y, prestore(v[cname])))
    cv=np.array(scores)
    print(cv, cv.mean(), cv.std())

def check_vb(datanm, samples_per_class, Cs, num_classes, num_iter = 100):
    data, labels = load_full(datanm, samples_per_class)
    slo = StratifiedShuffleSplit(labels, n_iter=num_iter, test_size=0.5, train_size=0.5, random_state=None)
    ans = np.zeros((len(Cs), samples_per_class/2, 2))
    for train_index, test_index in slo:
        train_data = [data[train_index, :], labels[train_index]]
        valid_data = [data[test_index , :], labels[test_index ]]

        for l in xrange(samples_per_class/2):
            ind_train = []
            ind_valid = []
            for k in xrange(num_classes):
                ind_train = ind_train + np.where(train_data[1] == k)[0].tolist()[:l+1]
                ind_valid = ind_valid + np.where(valid_data[1] == k)[0].tolist()[:l+1]

            ctrain_data = [ train_data[0][ind_train], train_data[1][ind_train] ]
            cvalid_data = [ valid_data[0][ind_valid], valid_data[1][ind_valid] ]

            for i, C in enumerate(Cs):
                clf = LogisticRegression(C  =C   , penalty='l2', multi_class = 'ovr',
                                         tol=0.001, n_jobs = -1 , verbose = 0)#, solver = 'newton-cg')
                clf.fit(ctrain_data[0], ctrain_data[1])

                out_train = clf.predict_proba(ctrain_data[0])
                out_valid = clf.predict_proba(cvalid_data[0])

                ans[i, l, 0] += log_loss(ctrain_data[1], out_train)
                ans[i, l, 1] += log_loss(cvalid_data[1], out_valid)

    ans /= num_iter

    np.savez("logreg_bv", ans= ans, Cs = Cs, num_iter = num_iter, num_classes = num_classes, samples_per_class = samples_per_class)
    return ans

def generic_cv_np(X,y,model,n_folds,random_state) :
    kf = cross_validation.KFold(y.shape[0],n_folds=n_folds, shuffle=True, random_state=random_state)
    trscores, cvscores, times = [], [], []
    i = 0
    stack_train = np.zeros((len(y))) # stacked predictions
    for i, (train_fold, validate) in enumerate(kf) :
        i = i + 1
        t = time()
        
        model.fit(X[train_fold,], y[train_fold])
        
        trscore = log_loss(y[train_fold], model.predict_proba(X[train_fold,]))
        
        validation_prediction = model.predict_proba(X[validate,])
        
        cvscore = log_loss(y[validate], validation_prediction)
        trscores.append(trscore); cvscores.append(cvscore); times.append(time()-t)
        
        stack_train[validate] = validation_prediction
    
    print("TRAIN %.5f | TEST %.5f | TIME %.2fm (1-fold)" % (np.mean(trscores), np.mean(cvscores), np.mean(times)/60))
    print(model.get_params())
    print("\n")
    
    return np.mean(cvscores), stack_train