sklearn踩坑confusion_matrix和KFold

sklearn.metrics.confusion_matrix(y_true, y_pred, labels=None, sample_weight=None)

        虽然sklearn.metrics.confusion_matrix很方便可以直接得到混淆矩阵，但是由于我实践后发现问题：sklearn的混淆矩阵实现会在只有一类的时候自动降维成1*1，于是我自己手撸一个：

def calculate_metric(gt, pred): 
    pred[pred>0.5]=1
    pred[pred<1]=0
    TP, FP, TN, FN = 0, 0, 0, 0

    for i in range(len(gt)):
        if gt[i] == 1 and pred[i] == 1:
           TP += 1
        if gt[i] == 0 and pred[i] == 1:
           FP += 1
        if gt[i] == 0 and pred[i] == 0:
           TN += 1
        if gt[i] == 1 and pred[i] == 0:
           FN += 1

    # confusion = confusion_matrix(gt,pred)
    # print(confusion.shape)
    # TP = confusion[1,1]
    # TN = confusion[0,0]
    # FP = confusion[0,1]
    # FN = confusion[1,0]

    return TP, FP, TN, FN

        但后来发现问题怪不到confusion_matrix，其实是因为数据样本较少，从而导致五折交叉验证时每一折不能保证每一折都包含两类样本，万幸：

 所以去了解了下KFold与StratifiedKFold，两者用法一样，只需要修改函数名就好：

from sklearn.model_selection import train_test_split, KFold, StratifiedKFold


kf = KFold(n_splits=5,random_state=2023,shuffle=True)

kf = StratifiedKFold(n_splits=5,random_state=2023,shuffle=True)

但在做split时两者有部分不同

#KFold不需要传入标签
for train_index, validate_index in kf.split(dataset):  
    pass
#StratifiedKFold需要传入标签
for train_index, validate_index in kf.split(dataset,dataset['label']):  
    pass

 换成StratifiedKFold分层采样就不会有一折只有一类的情况，从而也避免了confusion_matrix的降维问题，至此问题解决。