def gainsChart(self, clf, X_train, y_train):
predictions = clf.predict_proba(X_train)
plot_cumulative_gain(y_train,
X_train,
figsize=(12, 8),
title_fontsize=20,
text_fontsize=18)
def test_ax(self):
np.random.seed(0)
clf = LogisticRegression()
clf.fit(self.X, self.y)
probas = clf.predict_proba(self.X)
fig, ax = plt.subplots(1, 1)
out_ax = plot_cumulative_gain(self.y, probas)
assert ax is not out_ax
out_ax = plot_cumulative_gain(self.y, probas, ax=ax)
assert ax is out_ax
def log_cumulative_gain(y_true,
y_pred,
experiment=None,
channel_name='metric_charts',
prefix=''):
"""Creates cumulative gain chart and logs it to Neptune.
Args:
y_true (array-like, shape (n_samples)): Ground truth (correct) target values.
y_pred (array-like, shape (n_samples, 2)): Predictions for classes 0 and 1 with values from 0 to 1.
experiment(`neptune.experiments.Experiment`): Neptune experiment. Default is None.
channel_name(str): name of the neptune channel. Default is 'metric_charts'.
prefix(str): Prefix that will be added before metric name when logged to Neptune.
Examples:
Train the model and make predictions on test::
from sklearn.datasets import make_classification
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import train_test_split
from sklearn.metrics import classification_report
X, y = make_classification(n_samples=2000)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)
model = RandomForestClassifier()
model.fit(X_train, y_train)
y_test_pred = model.predict_proba(X_test)
Create and log cumulative gain chart to Neptune::
import neptune
from neptunecontrib.monitoring.metrics import log_cumulative_gain
neptune.init()
with neptune.create_experiment():
log_cumulative_gain(y_test, y_test_pred)
Check out this experiment https://ui.neptune.ai/o/neptune-ai/org/binary-classification-metrics/e/BIN-101/logs.
"""
assert len(
y_pred.shape
) == 2, 'y_pred needs to be (n_samples, 2), use expand_prediction helper to format it'
_exp = experiment if experiment else neptune
expect_not_a_run(_exp)
fig, ax = plt.subplots()
plt_metrics.plot_cumulative_gain(y_true, y_pred, ax=ax)
send_figure(fig, channel_name=prefix + channel_name, experiment=_exp)
plt.close()
def lift_gain_curves(self):
y_pred_proba_both_classes = np.column_stack(
[1 - self.y_pred_proba, self.y_pred_proba])
gain = plot_cumulative_gain(self.y_test,
y_pred_proba_both_classes,
title='Cumulative Gains Curve')
plt.show()
lift = plot_lift_curve(self.y_test,
y_pred_proba_both_classes,
title='Lift curve')
plt.show()
def k_NN(self):
"""Classify the patientsusing k-NN, selecting the best number of
neighbours both via a 5-fold and a loo cross-validation procedure.
Plot the two estimated error for each possible value of k. Add to the
plot the corresponding test errors (i.e., the test error you would have
obtained fitting k-NN with the same k) and comment on the results."""
nnbs = KNeighborsClassifier(n_neighbors=10)
modl = nnbs.fit(self.Xtrain, self.ytrain)
score = modl.score(self.Xtest, self.ytest)
ypred = np.array(modl.predict(self.Xtest).tolist())
if self.plot:
yprobas = \
np.append((1-ypred).reshape(-1,1), ypred.reshape(-1,1), axis=1)
plot_cumulative_gain(self.ytest.values, yprobas)
plt.title("Cumulative Gains Curve of kNN prediction with\n"\
+ f"binary accuracy of {100*score:.2f}%")
plt.show()
"""Research which k in kNN is the best:"""
num = 100
klin = np.linspace(1, 100, num)
scorearr = np.zeros(num)
for counter, i in enumerate(klin):
nnbs = KNeighborsClassifier(n_neighbors=int(i))
modl = nnbs.fit(self.Xtrain, self.ytrain)
score = modl.score(self.Xtest, self.ytest)
scorearr[counter] = score
if self.plot:
plt.plot(klin, scorearr)
plt.title("The kNN method analysis for multiple k values.")
plt.grid()
plt.xlabel("k")
plt.ylabel("Prediction scores")
plt.show()
return 1
def plot_analysis(combine,
test_name,
y_true,
y_pred,
y_proba,
labels,
verbose,
library,
save=True,
show=True,
sessionid="testing",
prefix=""):
met_index = 0
plt.rcParams.update({'font.size': 14})
# TODO: Find a way to do this better
pltmetrics.plot_confusion_matrix(y_true, y_pred)
if not combine:
#plt.gcf().set_size_inches(3.65,3.65)
save_show(plt, library + "/" + prefix, sessionid, "confusion_matrix",
show, save, False, True, True, False)
else:
plt.subplot(2, 4, met_index + 1)
met_index += 1
plt.rcParams.update({'font.size': 12})
pltmetrics.plot_roc_curve(y_true, y_proba)
for text in plt.gca().legend_.get_texts():
text.set_text(text.get_text().replace("ROC curve of class", "class"))
text.set_text(text.get_text().replace("area =", "AUC: "))
text.set_text(text.get_text().replace("micro-average ROC curve",
"micro-avg"))
text.set_text(text.get_text().replace("macro-average ROC curve",
"macro-avg"))
if not combine:
#plt.gcf().set_size_inches(3.65,3.65)
save_show(plt, library + "/" + prefix, sessionid, "roc_curves", show,
save, False, True, True, False)
else:
plt.subplot(2, 4, met_index + 1)
met_index += 1
if len(labels) < 3:
pltmetrics.plot_ks_statistic(y_true, y_proba)
if not combine:
#plt.gcf().set_size_inches(3.65,3.65)
save_show(plt, library + "/" + prefix, sessionid, "ks_statistics",
show, save, False, True, True, False)
else:
plt.subplot(2, 4, met_index + 1)
met_index += 1
pltmetrics.plot_precision_recall_curve(y_true, y_proba)
for text in plt.gca().legend_.get_texts():
text.set_text(text.get_text().replace(
"Precision-recall curve of class", "class"))
text.set_text(text.get_text().replace("area =", "AUC: "))
text.set_text(text.get_text().replace(
"micro-average Precision-recall curve", "micro-avg"))
text.set_text(text.get_text().replace("macro-average Precision-recall",
"macro-avg"))
if not combine:
#plt.gcf().set_size_inches(3.65,3.65)
save_show(plt, library + "/" + prefix, sessionid,
"precision_recall_curve", show, save, False, True, True,
False)
else:
plt.subplot(2, 4, met_index + 1)
met_index += 1
if len(labels) < 3:
pltmetrics.plot_cumulative_gain(y_true, y_proba)
if not combine:
#plt.gcf().set_size_inches(3.65,3.65)
save_show(plt, library + "/" + prefix, sessionid,
"cumulative_gain", show, save, False, True, True, False)
else:
plt.subplot(2, 4, met_index + 1)
met_index += 1
if len(labels) < 3:
pltmetrics.plot_lift_curve(y_true, y_proba)
if not combine:
#plt.gcf().set_size_inches(3.65,3.65)
save_show(plt, library + "/" + prefix, sessionid, "lift_curve",
show, save, False, True, True, False)
else:
plt.subplot(2, 4, met_index + 1)
met_index += 1
if combine:
plt.suptitle(test_name)
plt.tight_layout(rect=[0, 0.03, 1, 0.95])
save_show(plt,
library,
sessionid,
figname,
show,
save,
True,
analysis=True)
def test_array_like(self):
plot_cumulative_gain([0, 1], [[0.8, 0.2], [0.2, 0.8]])
plot_cumulative_gain([0, 'a'], [[0.8, 0.2], [0.2, 0.8]])
plot_cumulative_gain(['b', 'a'], [[0.8, 0.2], [0.2, 0.8]])
def test_string_classes(self):
np.random.seed(0)
clf = LogisticRegression()
clf.fit(self.X, convert_labels_into_string(self.y))
probas = clf.predict_proba(self.X)
plot_cumulative_gain(convert_labels_into_string(self.y), probas)
plt.plot(np.arange(n), Train_acc, label="train", color="b")
plt.plot([0, n], [train_acc, train_acc],
label="avg-train",
linestyle=":",
color="b")
plt.plot(np.arange(n), Test_acc, label="test", color="y")
plt.plot([0, n], [test_acc, test_acc],
label="avg-test",
linestyle=":",
color="y")
plt.xlabel("n"), plt.ylabel("accuracy %")
plt.legend(), plt.grid(), plt.show()
#False-negative and false-positive error rates
print("False-neg. rate: %.3f" % false_neg)
print("False-pos. rate: %.3f" % false_pos)
"""
output:
False-neg. rate: 0.305
False-pos. rate: 0.172
"""
#ROC curves of last train_test_split sample
p_test = np.ravel(p_test)
y_probas = np.zeros((len(p_test), 2))
y_probas[:, 0] = 1 - p_test
y_probas[:, 1] = p_test
plot_cumulative_gain(y_true=y_test, y_probas=y_probas)
plt.show()
def plot_lift(y_true, y_pred):
plot_cumulative_gain(y_true, y_pred)
plt.show()
def GridSearch(self, X_test, X_train, y_test, y_train, lmbd_vals, eta_vals,
iterations):
"""
Compares the best learning rate and regularization parameter for multiple-cases.
Calculate AUC-score and accuracy score.
"""
train_accuracy = np.zeros((len(eta_vals), len(lmbd_vals)))
test_accuracy = np.zeros((len(eta_vals), len(lmbd_vals)))
roc_score_test = np.zeros((len(eta_vals), len(lmbd_vals)))
roc_score_train = np.zeros((len(eta_vals), len(lmbd_vals)))
NN_numpy = np.zeros((len(eta_vals), len(lmbd_vals)), dtype=object)
#grid search
for i in range(len(eta_vals)):
for j in range(len(lmbd_vals)):
self.create_biases_and_weights()
#self.print_values()
self.train(iterations)
test_pred = self.predict(X_test)
train_pred = self.predict(X_train)
accuracy = accuracy_score(y_test, test_pred)
#train_accuracy[i][j] = metrics.accuracy_score(y_train,train_pred.round(), normalize=False)
#test_accuracy[i][j] = metrics.accuracy_score(y_test,test_pred)
train_accuracy[i][j] = accuracy_score(y_train, train_pred)
test_accuracy[i, j] = accuracy_score(y_test, test_pred)
roc_score_test[i, j] = metrics.roc_auc_score(y_test, test_pred)
roc_score_train[i, j] = metrics.roc_auc_score(
y_train, train_pred)
if accuracy > self.best_accuracy:
self.best_accuracy = accuracy
self.best_lmbd = lmbd_vals[j]
self.best_eta = eta_vals[i]
#print('Accuracy score on test data:', accuracy)
print('best accuracy:', self.best_accuracy)
print('lambda:', self.best_lmbd)
print('Learning rate:', self.best_eta)
#print('Train Area ratio:', np.mean(roc_score_train))
#print('Test Area ratio:', np.mean(roc_score_test))
sns.set()
sns.heatmap(train_accuracy, annot=True, cmap="viridis", fmt='.4g')
plt.title("Training Accuracy")
plt.ylabel('Learning rate: $\\eta$')
plt.xlabel('Regularization Term: $\\lambda$')
b, t = plt.ylim()
b += 0.5 # Add 0.5 to the bottom
t -= 0.5 # Subtract 0.5 from the top
plt.ylim(b, t)
#plt.savefig('traing_accuracy_cc_nn.png')
plt.show()
sns.heatmap(test_accuracy, annot=True, cmap="viridis", fmt='.4g')
plt.title("Test Accuracy")
plt.ylabel('Learning rate: $\\eta$')
plt.xlabel('Regularization Term: $\\lambda$')
b, t = plt.ylim()
b += 0.5 # Add 0.5 to the bottom
t -= 0.5 # Subtract 0.5 from the top
plt.ylim(b, t)
#plt.savefig('test_accuracy_cc_nn.png')
plt.show()
sns.heatmap(roc_score_train, annot=True, cmap="viridis", fmt='.4g')
plt.title("AUC Train")
plt.ylabel('Learning rate: $\\eta$')
plt.xlabel('Regularization Term: $\\lambda$')
b, t = plt.ylim()
b += 0.5 # Add 0.5 to the bottom
t -= 0.5 # Subtract 0.5 from the top
plt.ylim(b, t)
#plt.savefig('traing_auc_cc_nn.png')
plt.show()
sns.heatmap(roc_score_test, annot=True, cmap="viridis", fmt='.4g')
plt.title("AUC Test")
plt.ylabel('Learning rate: $\\eta$')
plt.xlabel('Regularization Term: $\\lambda$')
b, t = plt.ylim()
b += 0.5 # Add 0.5 to the bottom
t -= 0.5 # Subtract 0.5 from the top
plt.ylim(b, t)
#plt.savefig('test_auc_cc_nn.png')
plt.show()
Confusion_Matrix(y_test, test_pred)
#test_pred = np.argmax(test_pred, axis=1)
diff = np.concatenate((1 - test_pred, test_pred), axis=1)
plot_cumulative_gain(y_test, diff)
plot_roc(y_test, diff, plot_micro=False, plot_macro=False)
plt.show()
def LogisticRegression_self_test(X_train, X_test, y_train, y_test,
learning_rates, epochs, iteration):
"""
Logistic regression with stochastic gradient descent and gradient descent.
"""
# scoping number of training samples
n_inputs = X_train.shape[0]
n_features = X_train.shape[1]
eta_ = 1e-12
beta_opt = np.random.randn(X_train.shape[1], 2)
calc_beta_GD, norm = GradientDescent(X_train, beta_opt, y_train, iteration,
eta_)
prob_GD, predict_GD = Probability_GD(
X_test, calc_beta_GD) #defining values to be between 0 and 1
#yPred_GD = (predict_GD >= 0.5).astype(int) # converting to just 0 or 1
#Define Logistic regression
clf = LogisticRegression(solver='lbfgs', max_iter=1e5)
clf = clf.fit(X_train, np.ravel(y_train))
pred_sklearn = clf.predict(X_test)
prob_sklearn = clf.predict_proba(X_test)
#print(prob_sklearn)
#for eta in np.logspace(np.log10(1e-6), np.log10(1e0), 7):
accuracy = np.zeros(len(learning_rates))
auc_score = np.zeros(len(learning_rates))
for i, eta in enumerate(learning_rates):
beta_SGD = stochastic_gradient_descent(X_train, beta_opt, y_train, eta,
epochs, iteration)
prob_SGD, predict_SGD = Probability(
X_test, beta_SGD) #defining values to be between 0 and 1
accuracy[i] = metrics.accuracy_score(y_test, predict_SGD)
auc_score[i] = metrics.roc_auc_score(y_test, predict_SGD)
difference = y_test - predict_SGD
if i > 0 and auc_score[i] > auc_score[i - 1]:
best_pred_SGD = predict_SGD
best_prob_SGD = prob_SGD
print('Accuracy {}, learning rate= {}, iterations = {}'.format(
accuracy[i], eta, iteration))
print('Auc score: {}'.format(auc_score[i]))
"""
plt.plot(yPred, label='predict')
plt.plot(optimal_beta, label ='optimal beta')
plt.plot(y_test, label='test')
plt.show()
"""
sns.set()
sns.heatmap(pd.DataFrame(accuracy), annot=True, fmt='.4g')
plt.title('Grid-search for logistic regression')
plt.ylabel('Learning rate: $\\eta$')
plt.xlabel('Regularization Term: $\\lambda$')
#plt.xticks(ticks=np.arange(len(learning_rates)) + 0.5, labels=learning_rates)
#plt.yticks(ticks=np.arange(len(lambda_values)) + 0.5, labels=lambda_values)
b, t = plt.ylim() # discover the values for bottom and top
b += 0.5 # Add 0.5 to the bottom
t -= 0.5 # Subtract 0.5 from the top
plt.ylim(b, t) # update the ylim(bottom, top) values
#plt.savefig('accuracy_logreg.png')
plt.show()
sns.heatmap(pd.DataFrame(auc_score), annot=True, fmt='.4g')
plt.title('Grid-search for logistic regression')
plt.ylabel('Learning rate: $\\eta$')
plt.xlabel('Regularization Term: $\\lambda$')
#plt.xticks(ticks=np.arange(len(learning_rates)) + 0.5, labels=learning_rates)
#plt.yticks(ticks=np.arange(len(lambda_values)) + 0.5, labels=lambda_values)
b, t = plt.ylim() # discover the values for bottom and top
b += 0.5 # Add 0.5 to the bottom
t -= 0.5 # Subtract 0.5 from the top
plt.ylim(b, t) # update the ylim(bottom, top) values
#plt.savefig('auc_score_logreg.png')
plt.show()
#plot confusion matrix
Confusion_Matrix(y_test, predict_GD)
#Confusion_Matrix(y_test, best_pred_SGD)
#Confusion_Matrix(y_test, pred_sklearn)
#diff = np.concatenate((1- predict, predict), axis=1)
diff_sklearn = np.concatenate((1 - prob_sklearn, prob_sklearn), axis=1)
diff_GD = np.concatenate((1 - prob_GD, prob_GD), axis=1)
diff_SGD = np.concatenate((1 - best_prob_SGD, best_prob_SGD), axis=1)
#plot roc curves
plot_roc(y_test, prob_sklearn)
plot_roc(y_test, diff_SGD)
plot_roc(y_test, prob_GD)
plt.show()
#plot cumulative gain curves
plot_cumulative_gain(y_test, prob_sklearn)
ax = plot_cumulative_gain(y_test, diff_SGD)
plot_cumulative_gain(y_test, prob_GD)
#plt.show()
"""
#plot roc curves
plot_roc(y_test, diff_sklearn, plot_micro=False, plot_macro= False)
plot_roc(y_test, diff_GD, plot_micro=False, plot_macro= False)
plot_roc(y_test, diff_SGD, plot_micro=False, plot_macro= False)
plt.show()
#plot cumulative gain curves
plot_cumulative_gain(y_test, diff_sklearn)
plot_cumulative_gain(y_test, diff_GD)
plot_cumulative_gain(y_test, diff_SGD)
plt.show()
"""
model_curve = auc_score
area_baseline = 0.5
area_ratio = (model_curve - area_baseline) / (area_baseline)
print('Area Ratio:', area_ratio)
return accuracy, learning_rates
