def remove_outliers(array):
if not isinstance(array, np.ndarray):
raise Exception('input type should be numpy ndarray, instead of {}'.format(type(array)))
Q1 = np.quantile(array,0.25)
Q3 = np.quantile(array,0.75)
IQR = Q3 - Q1
array = array[array<=Q3+1.5*IQR]
array = array[ array>= Q1-1.5*IQR]
return array
def _calc_wts(self, x_i):
distances = np.array(
[np.linalg.norm(self.X_[i, :] - x_i) for i in range(self.X_.shape[0])]
)
weights = np.exp(-(distances ** 2) / self.sigma)
if self.span:
weights = weights * (distances <= np.quantile(distances, q=self.span))
return weights
def get_anchor_points(self, actions):
""" Builds anchor action point sets for the direct estimator
Args:
actions (np.array): actions drawn from the logging policy to build quantiles on
"""
if self.mode == 'quantile':
self.quantiles = np.quantile(actions, np.linspace(0, 1, self.K+1))
self.action_set = np.pad(self.quantiles, 1, 'constant', constant_values=(self.eps, np.inf))
elif self.mode == 'grid':
self.action_set = np.arange(self.eps, np.max(actions) + self.stride, self.stride)
self.K = int(np.max(actions)/self.stride)
self.initialized = True
return self.action_set
def T_val(x, gamma):
mu = x[0]
sig = np.exp(x[1])
for m in range(K):
X = mu + (sig) * Z[:, m]
P = (stats.norm.pdf(X, 2, 0.8) + stats.norm.pdf(X, 0, 0.8) +
stats.norm.pdf(X, -2, 0.8) + stats.norm.pdf(X, -4, 0.8)) / 4
logQ = stats.t.logpdf(X, 10, mu, sig)
logF = np.log(
P[(P > 0) & (logQ > -np.inf)]) - logQ[(P > 0) & (logQ > -np.inf)]
T_value = np.quantile(
-logF[(~np.isnan(logF))], gamma, interpolation="lower") - np.log(2)
return (T_value)
def _setup(self):
""" Setup the experiments and creates the data
"""
# Actions
features, y = self.get_X_y_by_name()
potentials = self._get_potentials(y)
actions = self.rng.lognormal(mean=self.start_mu,
sigma=self.start_sigma,
size=potentials.shape[0])
rewards = self.get_rewards_from_actions(potentials, actions)
if self.discrete:
from scipy.stats import lognorm
rv = lognorm(s=self.start_sigma, scale=np.exp(self.start_mu))
quantiles = np.quantile(actions,
np.linspace(0, 1, self.discrete + 1))
action_anchors = np.pad(quantiles,
1,
'constant',
constant_values=(1e-7, np.inf))
bins = action_anchors[:-1]
inds = np.digitize(actions, bins, right=True)
inds_1 = inds - 1
inds_1[inds_1 == -1] = 0
pi_logging = rv.cdf(bins[inds]) - rv.cdf(bins[inds_1])
else:
pi_logging = Dataset.logging_policy(actions, self.start_mu,
self.start_sigma)
# Test train split
self.actions_train, self.actions_test, self.features_train, self.features_test, self.reward_train, \
self.reward_test, self.pi_0_train, self.pi_0_test, self.potentials_train, self.potentials_test, \
self.l_train, self.l_test = train_test_split(actions, features, rewards, pi_logging, potentials, y,
train_size=self.train_size, random_state=42)
self.actions_train, self.actions_valid, self.features_train, self.features_valid, self.reward_train, \
self.reward_valid, self.pi_0_train, self.pi_0_valid, self.potentials_train, self.potentials_valid, \
self.l_train, self.l_valid = train_test_split(self.actions_train, self.features_train, self.reward_train,
self.pi_0_train, self.potentials_train, self.l_train,
train_size=self.val_size, random_state=42)
min_max_scaler = MinMaxScaler(feature_range=(0, 1))
self.features_train = min_max_scaler.fit_transform(self.features_train)
self.features_valid = min_max_scaler.transform(self.features_valid)
self.features_test = min_max_scaler.transform(self.features_test)
self.baseline_reward_valid = np.mean(self.reward_valid)
self.baseline_reward_test = np.mean(self.reward_test)
def results_plot(best_arg):
y_test = best_arg[-1][-1]
# CI plot
pool = ProcessPool(num_cores)
result = pool.map_async(grad_func, best_arg)
result.wait()
pool.close()
pool.join()
grad_total = np.array(result.get())
grad_mu_total = np.mean(grad_total, 0)
pred_prob = grad_mu_total[-2]
std_sample = [j[-2].ravel() for j in grad_total]
prob_ub = np.quantile(std_sample, 0.975, axis=0)
prob_lb = np.quantile(std_sample, 0.025, axis=0)
plt.plot(range(len(pred_prob)), pred_prob)
plt.title('Probability Confidence Interval')
plt.fill_between(range(len(pred_prob)),
prob_ub,
prob_lb,
color='b',
alpha=.1)
plt.show()
"""
auc_list=[]
f_p_r_list=[]
t_p_r_list=[]
for sample_index in range(S):
false_positive_rate, true_positive_rate, thresholds = roc_curve(y_test,grad_total[sample_index][-2].ravel())
roc_auc=auc(false_positive_rate, true_positive_rate)
auc_list.append(roc_auc)
f_p_r_list.append(false_positive_rate)
t_p_r_list.append(true_positive_rate)
up_idx = int(0.975 * (len(auc_list) - 1))
np.argpartition(f_p_r_list, up_idx)
np.quantile(auc_list,0.025)
plt.title('Receiver Operating Characteristic')
1.96*np.std(auc_list)
plt.plot(false_positive_rate, true_positive_rate, 'b',
label='AUC = %0.2f'% np.mean(auc_list))
plt.plot(false_positive_rate, true_positive_rate, 'b',
label='AUC = %0.2f'% np.quantile(auc_list,0.975))
plt.plot(false_positive_rate, true_positive_rate, 'b',
label='AUC = %0.2f'% np.quantile(auc_list,0.025))
plt.legend(loc='lower right')
plt.plot([0,1],[0,1],'r--')
plt.xlim([-0.1,1.2])
plt.ylim([-0.1,1.2])
plt.ylabel('True Positive Rate')
plt.xlabel('False Positive Rate')
plt.show()
"""
# AUC plot
false_positive_rate, true_positive_rate, thresholds = roc_curve(
y_test, pred_prob)
roc_auc = auc(false_positive_rate, true_positive_rate)
plt.title('Receiver Operating Characteristic')
plt.plot(false_positive_rate,
true_positive_rate,
'b',
label='AUC = %0.2f' % roc_auc)
plt.legend(loc='lower right')
plt.plot([0, 1], [0, 1], 'r--')
plt.xlim([-0.1, 1.2])
plt.ylim([-0.1, 1.2])
plt.ylabel('True Positive Rate')
plt.xlabel('False Positive Rate')
plt.show()
accuracy = accuracy_score(pred_prob > 0.5, y_test)
print("accuracy is " + str(accuracy) + '')
return pred_prob, roc_auc
def set_list_quantiles(self, features):
self.list_quantiles = list(np.pad(np.array([np.quantile(features, (i + 1) / self.number_quantiles, axis=0) \
for i in range(self.number_quantiles - 1)]), 1, 'constant', \
constant_values=(-np.inf, np.inf)).T)[1:-1]
self.initialized = True
