字错误率（cer）
使用于中文等不以空格为分隔的语言
import evaluatemetric = evaluate.load("cer")
print(metric.compute(predictions=['你吃了吗', '今天我要去打篮球'], references=["我吃了么a" , '明天我要去打篮球']))词错误率（wer）
import evaluatemetric = evaluate.load("wer")
print(metric.compute(predictions=['a b c', 'd e f'], references=['a b c', '1 2 3']))获取最优f1
def find_best_f1_and_threshold(scores, labels, ):""":param scores: 正样本概率:param labels: 真实标签:return: """assert len(scores) == len(labels)scores = np.asarray(scores)labels = np.asarray(labels)rows = list(zip(scores, labels))rows = sorted(rows, key=lambda x: x[0], reverse=True)best_f1 = best_precision = best_recall = 0threshold = 0nextract = 0ncorrect = 0total_num_duplicates = sum(labels)for i in range(len(rows) - 1):score, label = rows[i]nextract += 1if label == 1:ncorrect += 1if ncorrect > 0:precision = ncorrect / nextractrecall = ncorrect / total_num_duplicatesf1 = 2 * precision * recall / (precision + recall)if f1 > best_f1:best_f1 = f1best_precision = precisionbest_recall = recallthreshold = (rows[i][0] + rows[i + 1][0]) / 2return best_f1, best_precision, best_recall, threshold寻找最优accuracy
from sklearn.metrics import accuracy_scoredef find_best_accuracy_and_threshold(scores, labels, ):""":param scores: 正样本概率:param labels: 真实标签:return:"""best_accuracy = 0best_th = 0for th in scores:pre_label = [1 if score >= th else 0 for score in scores]accuracy = accuracy_score(labels, pre_label)if accuracy > best_accuracy:best_accuracy = accuracybest_th = threturn best_accuracy, best_th
)
)
