怎么使用Weights & Biases记录异常检测实验？

wandb

pip install wandb

import wandb
import numpy as np
import pandas as pd
from sklearn.ensemble import IsolationForest
from sklearn.metrics import average_precision_score, roc_auc_score, precision_recall_curve, roc_curve
import matplotlib.pyplot as plt
import seaborn as sns

# 假设你已经有了数据 X, y (y是真实标签，0为正常，1为异常)
# X: 特征数据，y: 标签
# 模拟一些数据
np.random.seed(42)
normal_data = np.random.randn(1000, 5) * 2
anomaly_data = np.random.randn(20, 5) * 5 + 10 # 异常点分布不同
X = np.vstack((normal_data, anomaly_data))
y = np.array([0]*1000 + [1]*20)

# 打乱数据
indices = np.arange(len(X))
np.random.shuffle(indices)
X = X[indices]
y = y[indices]

# 1. 初始化W&B
wandb.init(project="anomaly-detection-experiments",
           name="isolation_forest_run_01",
           config={
               "model_type": "IsolationForest",
               "n_estimators": 100,
               "max_features": 1.0,
               "contamination": 0.02, # 预估的异常比例
               "random_state": 42
           })

# 获取配置
config = wandb.config

# 2. 定义模型并训练
model = IsolationForest(n_estimators=config.n_estimators,
                        max_features=config.max_features,
                        contamination=config.contamination,
                        random_state=config.random_state)
model.fit(X)

# 获取异常分数
# IsolationForest的decision_function值越小，越可能是异常
# 为了方便理解，我们通常将其反转，让分数越大越异常
anomaly_scores = -model.decision_function(X)

# 3. 记录核心指标
# 异常检测中，PR-AUC通常比ROC-AUC更能反映模型在极度不平衡数据上的表现
pr_auc = average_precision_score(y, anomaly_scores)
roc_auc = roc_auc_score(y, anomaly_scores)

wandb.log({
    "pr_auc": pr_auc,
    "roc_auc": roc_auc,
    "max_anomaly_score": np.max(anomaly_scores),
    "min_anomaly_score": np.min(anomaly_scores),
    "mean_anomaly_score": np.mean(anomaly_scores)
})

# 4. 记录关键可视化
# PR曲线
precision, recall, _ = precision_recall_curve(y, anomaly_scores)
plt.figure(figsize=(8, 6))
plt.plot(recall, precision, label=f'PR Curve (AP = {pr_auc:.2f})')
plt.xlabel('Recall')
plt.ylabel('Precision')
plt.title('Precision-Recall Curve')
plt.legend()
wandb.log({"pr_curve": wandb.Image(plt)})
plt.close()

# ROC曲线
fpr, tpr, _ = roc_curve(y, anomaly_scores)
plt.figure(figsize=(8, 6))
plt.plot(fpr, tpr, label=f'ROC Curve (AUC = {roc_auc:.2f})')
plt.plot([0, 1], [0, 1], 'k--', label='Random Classifier')
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.title('Receiver Operating Characteristic (ROC) Curve')
plt.legend()
wandb.log({"roc_curve": wandb.Image(plt)})
plt.close()

# 异常分数分布
plt.figure(figsize=(8, 6))
sns.histplot(anomaly_scores[y == 0], color='blue', label='Normal', kde=True)
sns.histplot(anomaly_scores[y == 1], color='red', label='Anomaly', kde=True)
plt.title('Distribution of Anomaly Scores')
plt.xlabel('Anomaly Score')
plt.ylabel('Count')
plt.legend()
wandb.log({"anomaly_score_distribution": wandb.Image(plt)})
plt.close()

# 5. 记录模型和数据（可选但强烈推荐）
# 使用wandb.Artifacts来版本化你的模型和数据集
# 例如，保存训练好的模型
wandb.log_model(path="model.pkl", name="isolation_forest_model", metadata={"pr_auc": pr_auc})

# 结束W&B运行
wandb.finish()

wandb.init()

name

isolation_forest_v1_0_contam001

vae_anomaly_detection_latent32_epoch100

tags

model:isolation_forest

dataset:sensor_data

hyperparam_sweep