我正在使用sklearn的MLPClassifier((函数制作一个分类器，并有以下代码：
def train_NN(self, trained_data_list, test_size=0.25):
X = np.array([[7, 4], [3,7], [8,5], [2,3]])
y = np.array([0, 1, 1, 0])
X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=0)
#oddly, printing 'X_train' yeilds the following: [[2 3] [3 7] [7 4]] and seems to have dropped my point [8 5]
#printing 'y_train' yeilds the following: [0 1 0]
self.clf = MLPClassifier(solver='lbfgs', alpha=1e-10, hidden_layer_sizes=(2, 2), activation = 'tanh', random_state=5, max_iter=10000, learning_rate_init = 0.1)
self.clf.fit(X_train, y_train)
print(self.clf.predict(np.atleast_2d([7,4]))) #returns 0
print(self.clf.predict(np.atleast_2d([3,7]))) #returns 0
print(self.clf.predict(np.atleast_2d([8,5]))) #returns 0
print(self.clf.predict(np.atleast_2d([2,3]))) #returns 0
我在这里的理解是，"X"数组应该与"y"数组相关。例如，给clf.predict((值[3,7]应该返回"1"，因为这就是它们在定义NP数组时的关联方式。[8,5]应返回"1"，[2，3]应返回"0"等。
我计划在更大的规模上做这件事，这里的数据只是为了演示我的问题。我是NN领域的新手，所以我确信我只是错过了一些小东西。
发现了这个问题，以防有人偶然发现。我使用了错误的激活功能。事实证明，在这种分类场景中，"物流"比"tanh"效果好得多：
def train_NN(self, trained_data_list, test_size=0.25):
X = np.array([[7, 4], [3,7], [8,5], [2,3]])
y = np.array([0, 1, 1, 0])
X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=0)
self.clf = MLPClassifier(solver='lbfgs', alpha=1e-10, hidden_layer_sizes=(2, 2), activation = 'logistic', random_state=5, max_iter=10000, learning_rate_init = 0.1)
self.clf.fit(X_train, y_train)
print(self.clf.predict(np.atleast_2d([7,4]))) #returns 0
print(self.clf.predict(np.atleast_2d([3,7]))) #returns 1
print(self.clf.predict(np.atleast_2d([8,5]))) #returns 1
print(self.clf.predict(np.atleast_2d([2,3]))) #returns 0