data_modeling/evaluation/playground-series-s4e3_eval.py

import os.path

import numpy as np
import pandas as pd
import argparse
from sklearn.metrics import accuracy_score
from sklearn.metrics import roc_auc_score


# 计算多类对数损失
def multiclass_logloss(actuals, predictions):
    epsilon = 1e-15  # 避免对数运算中的数值问题
    predictions = np.clip(predictions, epsilon, 1 - epsilon)  # 限制预测概率的范围，防止对数为无穷
    predictions /= predictions.sum(axis=1)[:, np.newaxis]  # 归一化确保总和为1
    log_pred = np.log(predictions)
    loss = -np.sum(actuals * log_pred) / len(actuals)
    return loss


parser = argparse.ArgumentParser()

parser.add_argument('--path', type=str, required=True)
parser.add_argument('--name', type=str, required=True)
parser.add_argument('--answer_file', type=str, required=True)
parser.add_argument('--predict_file', type=str, required=True)

parser.add_argument('--value', type=str, default="NObeyesdad")

args = parser.parse_args()

actual = pd.read_csv(args.answer_file)
submission = pd.read_csv( args.predict_file)

# 定义要计算的类别
categories = ['Pastry', 'Z_Scratch', 'K_Scatch', 'Stains', 'Dirtiness', 'Bumps', 'Other_Faults']

# 提取数据并计算每个类别的 ROC AUC 分数
auc_scores = {}
for category in categories:
    y_true = actual[category].values
    y_pred = submission[category].values
    auc_scores[category] = roc_auc_score(y_true, y_pred)

# 计算平均 AUC 分数
performance = sum(auc_scores.values()) / len(auc_scores)

with open(os.path.join(args.path, args.name, "result.txt"), "w") as f:
    f.write(str(performance))