81 lines
3.6 KiB
Python
81 lines
3.6 KiB
Python
|
import datasets
|
|||
|
|
|
|||
|
|
# 加载cifar100-enriched数据集的测试集，如同阅读一本心法
|
|||
|
|
ds = datasets.load_dataset("cifar100-enriched", split="test")
|
|||
|
|
|
|||
|
|
# 将数据集转换为Pandas DataFrame，就像将零散的招式编纂成连招
|
|||
|
|
df = ds.to_pandas()
|
|||
|
|
|
|||
|
|
import pandas as pd
|
|||
|
|
import numpy as np
|
|||
|
|
from cleanlab import Datalab
|
|||
|
|
|
|||
|
|
# 使用Datalab寻找数据中的问题，如同武林高手寻找对手的破绽
|
|||
|
|
lab = Datalab(data=df, label_name="fine_label")
|
|||
|
|
|
|||
|
|
# 将DataFrame中的"embedding"列转换为numpy数组，如同将散落的招式秘籍编纂成连贯的招式
|
|||
|
|
features = np.array([x.tolist() for x in df["embedding"]])
|
|||
|
|
|
|||
|
|
# 将DataFrame中的"probabilities"列也转换为numpy数组，如同收集各路信息，为出招做准备
|
|||
|
|
pred_probs = np.array([x.tolist() for x in df["probabilities"]])
|
|||
|
|
|
|||
|
|
# 调用find_issues方法寻找数据问题，如同武林侠客挑战江湖隐秘，揭示真相
|
|||
|
|
lab.find_issues(features=features, pred_probs=pred_probs)
|
|||
|
|
|
|||
|
|
# 将原始数据框与发现的问题合并，如同将新发现的武林秘籍融入已有的知识体系中
|
|||
|
|
df = pd.concat([df, lab.get_issues()], axis=1)
|
|||
|
|
|
|||
|
|
from renumics.spotlight.analysis import DataIssue
|
|||
|
|
|
|||
|
|
# 寻找标签问题的行，如同武林中的侠客追踪敌人的踪迹
|
|||
|
|
label_issue_rows = df[df["is_label_issue"]].sort_values("label_score").index.tolist()
|
|||
|
|
# 记录标签问题，如同将侦查到的情报记录在案
|
|||
|
|
label_issue = DataIssue(
|
|||
|
|
severity="medium",
|
|||
|
|
title="label-issue",
|
|||
|
|
rows=label_issue_rows,
|
|||
|
|
description="Label issue found by Cleanlab - Review and correct if necessary",
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
# 寻找异常值问题的行，如同寻找隐藏在暗处的敌人
|
|||
|
|
outlier_issue_row = (
|
|||
|
|
df[df["outlier_score"] < 0.6].sort_values("outlier_score").index.tolist()
|
|||
|
|
)
|
|||
|
|
# 记录异常值问题，如同标记出潜藏的危险
|
|||
|
|
outlier_issue = DataIssue(
|
|||
|
|
severity="medium",
|
|||
|
|
title="outlier-issue",
|
|||
|
|
rows=outlier_issue_row,
|
|||
|
|
description="Outlier score < 0.6 - Review and remove or collect more data",
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
# 寻找几乎重复问题的行，如同寻找武林中几乎看不出差别的影子剑客
|
|||
|
|
near_duplicate_issue_row = (
|
|||
|
|
df[df["is_near_duplicate_issue"]].sort_values("near_duplicate_score").index.tolist()
|
|||
|
|
)
|
|||
|
|
# 记录几乎重复的问题，如同揭露伪装的敌人
|
|||
|
|
near_duplicate_issue = DataIssue(
|
|||
|
|
severity="medium",
|
|||
|
|
title="near-duplicate-issue",
|
|||
|
|
rows=near_duplicate_issue_row,
|
|||
|
|
description="Near duplicate issue found by Cleanlab - Review and remove if necessary",
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
from renumics import spotlight
|
|||
|
|
|
|||
|
|
# 定义数据类型，如同为武林中的各种武器赋予名号
|
|||
|
|
dtypes = {
|
|||
|
|
"image": spotlight.Image, # 图像数据，如同武林中的画卷，记录着形象
|
|||
|
|
"full_image": spotlight.Image, # 完整图像数据，如同详尽的地图，展现全貌
|
|||
|
|
"embedding": spotlight.Embedding, # 嵌入数据，如同内力深厚的武功秘籍
|
|||
|
|
"embedding_reduced": spotlight.Embedding, # 简化的嵌入数据，精简而不失精髓
|
|||
|
|
"probabilities": spotlight.Embedding, # 概率数据，如同预测敌人下一步的先天能力
|
|||
|
|
}
|
|||
|
|
|
|||
|
|
# 展示数据及其问题，如同武林大会上展示武功，亮出绝技
|
|||
|
|
spotlight.show(
|
|||
|
|
df.rename(columns={"fine_label_str": "label", "fine_label_prediction_str": "pred"}), # 重命名列，如同为武功起名，更易理解
|
|||
|
|
dtype=dtypes, # 指定数据类型，如同明确每位武林人士的门派和武功
|
|||
|
|
layout="https://spotlight.renumics.com/resources/layout_data_issues.json", # 使用布局文件，如同选择武林大会的比武场地
|
|||
|
|
issues=[label_issue, outlier_issue, near_duplicate_issue], # 展示发现的问题，如同揭示对手的破绽
|
|||
|
|
)
|