import bisect from sklearn import metrics from sklearn.metrics import precision_recall_curve# 創建一個單獨的Metircs.py文件 def calc_auc(y_true, y_pred): # y_pred：[0,1]之間任何數return metrics.roc_auc_score(y_true, y_pred)def calc_ks(y_true, y_pred): # y_pred：[0,1]之間任何數fpr, tpr, thresholds = metrics.roc_curve(y_true, y_pred)return max(tpr - fpr)def calc_acc(y_true, y_pred): # y_pred：{0,1} 必須事先通過閾值轉變為0,1return metrics.accuracy_score(y_true, y_pred)def calc_f1(y_true, y_pred): # y_pred：{0,1} 必須事先通過閾值轉變為0,1return metrics.recall_score(y_true, y_pred), metrics.precision_score(y_true, y_pred), metrics.f1_score(y_true, y_pred)# 召回率大于0.9時的準確率 def precision_at_r9(y_true, y_pred):p, r, tresholds = precision_recall_curve(y_true, y_pred)ind = bisect.bisect(r, 0.9)if ind >= len(p)-1:return r[-2], p[-2], tresholds[-1]else:return r[ind], p[ind], tresholds[ind]# 準確率大于0.9時的召回率 def recall_at_p9(y_true, y_pred):p, r, tresholds = precision_recall_curve(y_true, y_pred)ind = bisect.bisect(p, 0.9)if ind >= len(p)-1:return r[-2], p[-2], tresholds[-1]else:return r[ind], p[ind], tresholds[ind]

NDCG@K

def get_dcg(y_pred, y_true, k):# 注意y_pred與y_true必須是一一對應的，而且y_pred越大越接近label=1(用相關性的說法就是，與label=1越相關)df = pd.DataFrame({"y_pred": y_pred, "y_true": y_true})df = df.sort_values(by="y_pred", ascending=False) # 對y_pred進行降序排列，越排在前面的，越接近label=1df = df.iloc[0:k, :] # 取前K個dcg = (2 ** df["y_true"] - 1) / np.log2(np.arange(1, df["y_true"].count() + 1) + 1) # 位置從1開始計數dcg = np.sum(dcg)return dcgdef calc_ndcg(y_true, y_pred, k):dcg = get_dcg(y_pred, y_true, k)idcg = get_dcg(y_true, y_true, k)ndcg = dcg / idcgreturn ndcg

Recall@K,Precision@K

def calc_f1(y_true, y_pred, k): df = pd.DataFrame({"y_pred": y_pred, "y_true": y_true})df = df.sort_values(by="y_pred", ascending=False) # 對y_pred進行降序排列，越排在前面的，越接近label=1df = df.iloc[0:k, :] # 取前K個y_true = np.array(df["y_true"])y_pred = np.round(np.array(df["y_pred"])) # y_pred：{0,1} 必須事先通過閾值轉變為0,1return recall_score(y_true, y_pred), precision_score(y_true, y_pred), f1_score(y_true, y_pred)

總結

以上是生活随笔為你收集整理的ML/DL常用评估方法的全部內容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網站內容還不錯，歡迎將生活随笔推薦給好友。