python - 决策树分类器我不断收到 NaN 错误
问题描述
我有一个小的决策树代码，我相信我将所有内容都转换为 int，并且我已经使用 isnan、max 等检查了我的训练/测试数据。
我真的不知道为什么它会给出这个错误。
所以我试图从决策树传递 Mnist 数据集，然后我将使用一个类进行攻击。
这是代码：
from AttackUtils import Attack
from AttackUtils import calc_output_weighted_weights, targeted_gradient, non_targeted_gradient, non_targeted_sign_gradient
(X_train_woae, y_train_woae), (X_test_woae, y_test_woae) = mnist.load_data()
X_train_woae = X_train_woae.reshape((len(X_train_woae), np.prod(X_train_woae.shape[1:])))
X_test_woae = X_test_woae.reshape((len(X_test_woae), np.prod(X_test_woae.shape[1:])))
from sklearn import tree
#model_woae = LogisticRegression(multi_class='multinomial', solver='lbfgs', fit_intercept=False)
model_woae = tree.DecisionTreeClassifier(class_weight='balanced')
model_woae.fit(X_train_woae, y_train_woae)
#model_woae.coef_ = model_woae.feature_importances_
coef_int = np.round(model_woae.tree_.compute_feature_importances(normalize=False) * X_train_woae.size).astype(int)
attack_woae = Attack(model_woae)
attack_woae.prepare(X_train_woae, y_train_woae, X_test_woae, y_test_woae)
weights_woae = attack_woae.weights
num_classes_woae = len(np.unique(y_train_woae))
attack_woae.create_one_hot_targets(y_test_woae)
attack_woae.attack_to_max_epsilon(non_targeted_gradient, 50)
non_targeted_scores_woae = attack_woae.scores
所以攻击类进行扰动和非目标梯度攻击。这是攻击类：
import numpy as np
from sklearn.metrics import accuracy_score
def calc_output_weighted_weights(output, w):
for c in range(len(output)):
if c == 0:
weighted_weights = output[c] * w[c]
else:
weighted_weights += output[c] * w[c]
return weighted_weights
def targeted_gradient(foolingtarget, output, w):
ww = calc_output_weighted_weights(output, w)
for k in range(len(output)):
if k == 0:
gradient = foolingtarget[k] * (w[k]-ww)
else:
gradient += foolingtarget[k] * (w[k]-ww)
return gradient
def non_targeted_gradient(target, output, w):
ww = calc_output_weighted_weights(output, w)
for k in range(len(target)):
if k == 0:
gradient = (1-target[k]) * (w[k]-ww)
else:
gradient += (1-target[k]) * (w[k]-ww)
return gradient
def non_targeted_sign_gradient(target, output, w):
gradient = non_targeted_gradient(target, output, w)
return np.sign(gradient)
class Attack:
def __init__(self, model):
self.fooling_targets = None
self.model = model
def prepare(self, X_train, y_train, X_test, y_test):
self.images = X_test
self.true_targets = y_test
self.num_samples = X_test.shape[0]
self.train(X_train, y_train)
print("Model training finished.")
self.test(X_test, y_test)
print("Model testing finished. Initial accuracy score: " + str(self.initial_score))
def set_fooling_targets(self, fooling_targets):
self.fooling_targets = fooling_targets
def train(self, X_train, y_train):
self.model.fit(X_train, y_train)
self.weights = self.model.coef_
self.num_classes = self.weights.shape[0]
def test(self, X_test, y_test):
self.preds = self.model.predict(X_test)
self.preds_proba = self.model.predict_proba(X_test)
self.initial_score = accuracy_score(y_test, self.preds)
def create_one_hot_targets(self, targets):
self.one_hot_targets = np.zeros(self.preds_proba.shape)
for n in range(targets.shape[0]):
self.one_hot_targets[n, targets[n]] = 1
def attack(self, attackmethod, epsilon):
perturbed_images, highest_epsilon = self.perturb_images(epsilon, attackmethod)
perturbed_preds = self.model.predict(perturbed_images)
score = accuracy_score(self.true_targets, perturbed_preds)
return perturbed_images, perturbed_preds, score, highest_epsilon
def perturb_images(self, epsilon, gradient_method):
perturbed = np.zeros(self.images.shape)
max_perturbations = []
for n in range(self.images.shape[0]):
perturbation = self.get_perturbation(epsilon, gradient_method, self.one_hot_targets[n], self.preds_proba[n])
perturbed[n] = self.images[n] + perturbation
max_perturbations.append(np.max(perturbation))
highest_epsilon = np.max(np.array(max_perturbations))
return perturbed, highest_epsilon
def get_perturbation(self, epsilon, gradient_method, target, pred_proba):
gradient = gradient_method(target, pred_proba, self.weights)
inf_norm = np.max(gradient)
perturbation = epsilon / inf_norm * gradient
return perturbation
def attack_to_max_epsilon(self, attackmethod, max_epsilon):
self.max_epsilon = max_epsilon
self.scores = []
self.epsilons = []
self.perturbed_images_per_epsilon = []
self.perturbed_outputs_per_epsilon = []
for epsilon in range(0, self.max_epsilon):
perturbed_images, perturbed_preds, score, highest_epsilon = self.attack(attackmethod, epsilon)
self.epsilons.append(highest_epsilon)
self.scores.append(score)
self.perturbed_images_per_epsilon.append(perturbed_images)
self.perturbed_outputs_per_epsilon.append(perturbed_preds)
这是它给出的回溯：
值错误
4 num_classes_woae = len(np.unique(y_train_woae)) 5 attack_woae.create_one_hot_targets(y_test_woae) ----> 6 attack_woae.attack_to_max_epsilon(non_targeted_gradient, 50) 7 non_targeted_scores_woae = attack_woae.scores 中的回溯（最后一次调用）
~\MULTIATTACK\AttackUtils.py in attack_to_max_epsilon(self, attackmethod, max_epsilon) 106 self.perturbed_outputs_per_epsilon = [] 107 for epsilon in range(0, self.max_epsilon): --> 108 perturbed_images, perturbed_preds, score, highest_epsilon = self.攻击（攻击方法，epsilon）109 self.epsilons.append（highest_epsilon）110 self.scores.append（score）
~\MULTIATTACK\AttackUtils.py in attack(self, attackmethod, epsilon) 79 def attack(self, attackmethod, epsilon): 80 perturbed_images,highest_epsilon = self.perturb_images(epsilon, attackmethod) ---> 81 perturbed_preds = self.model .predict(perturbed_images) 82 score = accuracy_score(self.true_targets, perturbed_preds) 83 返回 perturbed_images, perturbed_preds, score,highest_epsilon
...\appdata\local\programs\python\python35\lib\site-packages\sklearn\tree\tree.py in predict(self, X, check_input) 413 """ 414 check_is_fitted(self, 'tree_') - -> 415 X = self._validate_X_predict(X, check_input) 416 proba = self.tree_.predict(X) 417 n_samples = X.shape[0]
...\appdata\local\programs\python\python35\lib\site-packages\sklearn\tree\tree.py in _validate_X_predict(self, X, check_input) 374 """每当尝试预测时验证 X，应用， predict_proba""" 375 if check_input: --> 376 X = check_array(X, dtype=DTYPE, accept_sparse="csr") 377 if issparse(X) and (X.indices.dtype != np.intc or 378 X. indptr.dtype != np.intc):
...\appdata\local\programs\python\python35\lib\site-packages\sklearn\utils\validation.py in check_array(array, accept_sparse, accept_large_sparse, dtype, order, copy, force_all_finite, ensure_2d, allow_nd, ensure_min_samples, ensure_min_features, warn_on_dtype, estimator) 566 if force_all_finite: 567 _assert_all_finite(array, --> 568 allow_nan=force_all_finite == 'allow-nan') 569 570 shape_repr = _shape_repr(array.shape)
...\appdata\local\programs\python\python35\lib\site-packages\sklearn\utils\validation.py 在 _assert_all_finite(X, allow_nan) 54 不是 allow_nan 而不是 np.isfinite(X).all()) : 55 type_err = 'infinity' if allow_nan else 'NaN, infinity' ---> 56 raise ValueError(msg_err.format(type_err, X.dtype)) 57 58
ValueError：输入包含 NaN、无穷大或对于 dtype('float32') 来说太大的值。
编辑：
我已将系数编号添加为 0，现在它在该行下方给出了相同的错误，在attack.attack_to_max_epsilon(non_targeted_gradient, epsilon_number)
解决方案
在训练之前尝试将 one-hot enconde 应用于您的标签。
from sklearn.preprocessing import LabelEncoder
mylabels= ["label1", "label2", "label2"..."n.label"]
le = LabelEncoder()
labels = le.fit_transform(mylabels)
然后尝试拆分您的数据：
from sklearn.model_selection import train_test_split
(x_train, x_test, y_train, y_test) = train_test_split(data,
labels,
test_size=0.25)
现在你的标签可能会用数字编码，这对训练机器学习算法很有好处。