分类评价指标（二分类）——f1score sensitivity specificity roc曲线 auc

注：此文章为记录学习历程，无参考价值。最近在学习分类（二分类）的时候，需要用到一些常用的分类指标，各大中文网站的的代码不仅看不懂，甚至还千篇一律，都不知道到底是谁抄的谁的，自己在一些英文网站看人家写的代码一点一点理解，总算是能弄出结果来了，故写此文章以记录

由于我也是新手，故关于理论的讲解请参考其他文章：
敏感性和特异性
roc曲线和auc，f1score等
(1)roc 曲线以及面积auc和f1score的实现
这三个方法的实现主要是用sklearn官方库里面的函数
f1_score, roc_curve, auc
建议去sklearn 官方查看示例官方roc_curve示例

from sklearn.metrics import roc_curve, f1_score, auc

fp, tp, thresholes = roc_curve(all_labele, all_prescore)
roc_auc = auc(fp, tp)
f1score = f1_score(all_labele, all_pre)

此处，all_lable代表数据数据的标签，all_pre代表预测出来的标签，也就是经过torch.argmax(outputs.data, 1)操作提取出来的标签，all_prescore是结果softmax回归后关于每一类预测的得分，也就是操作torch.max(outputs.data, 1)[0]
然后画出roc曲线：

    plt.plot(fp, tp, 'r', label='ROC curve (area = %0.2f)' % roc_auc)

    plt.legend(loc='lower right')
    plt.plot([0, 1], [0, 1], 'g')
    print(f'loss:{val_loss / len(val_loader):.4f}, acc:{correct / total}, auc:{roc_auc:.3f}, f1score:{f1score:.3f}')
    # plt.savefig('./result/1.png')
    plt.show()

下面给出完整的代码：

import numpy as np
import matplotlib.pyplot as plt
from sklearn.model_selection import cross_val_score
from sklearn.metrics import roc_curve, f1_score, auc, confusion_matrix
import torch
from torchvision import transforms
from torchvision.datasets import ImageFolder
from torch.utils.data import DataLoader
from torchvision import models
import torch.nn as nn
def prediction():
    device = torch.device("cuda:0")
    transform = transforms.Compose([
            transforms.Resize((224, 224)),
            transforms.ToTensor(),
            transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])
    val_dataset = ImageFolder('data', transform=transform)
    val_loader = DataLoader(val_dataset,
                            batch_size=40,
                            shuffle=False,
                            num_workers=0)
    net = models.resnet18(pretrained=False)
    in_channel = net.fc.in_features
    net.fc = nn.Linear(in_channel, 2)
    net.load_state_dict(torch.load('yourweight.pth', map_location=device))
    net.to(device)
    loss_function = nn.CrossEntropyLoss()
    net.eval()
    val_loss, correct, total =0, 0, 0
    all_labele, all_prescore, all_pre = np.array([]), np.array([]), np.array([])
    cfmetric = 0
    for step, batch in enumerate(val_loader):
        imgs, lables = batch
        imgs, lables = imgs.to(device), lables.to(device)
        with torch.no_grad():
            outputs = net(imgs)
        loss = loss_function(outputs, lables)
        val_loss += loss.item()
        predicted = torch.argmax(outputs, 1)
        total += lables.size(0)
        predicted_score = torch.max(outputs.data, 1)[0]
        correct += (predicted == lables).sum().item()
        all_pre = np.append(all_pre, np.array(predicted.to('cpu')))
        all_labele = np.append(all_labele, np.array(lables.to('cpu')))
        all_prescore = np.append(all_prescore, np.array(predicted_score.to('cpu')))

    fp, tp, thresholes = roc_curve(all_labele, all_prescore)
    roc_auc = auc(fp, tp)
    f1score = f1_score(all_labele, all_pre)
    plt.plot(fp, tp, 'r', label='ROC curve (area = %0.2f)' % roc_auc)

    plt.legend(loc='lower right')
    plt.plot([0, 1], [0, 1], 'g')
    print(f'loss:{val_loss / len(val_loader):.4f}, acc:{correct / total}, auc:{roc_auc:.3f}, f1score:{f1score:.3f}')
    # plt.savefig('./result/1.png')
    plt.show()

  

if __name__=='__main__':
    prediction()

(2)敏感性，特异性求解：根据公式
sensitivuty = tp/(tp+fn)
specificity = tn/(tn+fp)
那我们利用混淆矩阵，将tp, fp, tn, fn求出来：参考官方示例
all_lable, all_pre同上

from sklearn.metrics import  confusion_matrix

cfmetric = confusion_matrix(all_labele, all_pre, labels=[1, 0])
tn, fp, fn, tp = cfmetric.ravel()
sensitivity = tp/(tp+fn)
specificity = tn/(tn+fp)
print(f'sensitivity:{sensitivity:.3f}, specificity:{specificity:.3f}')

给出完整代码：

import numpy as np
import matplotlib.pyplot as plt
from sklearn.model_selection import cross_val_score
from sklearn.metrics import roc_curve, f1_score, auc, confusion_matrix
import torch
from torchvision import transforms
from torchvision.datasets import ImageFolder
from torch.utils.data import DataLoader
from torchvision import models
import torch.nn as nn
def prediction():
    device = torch.device("cuda:0")
    transform = transforms.Compose([
           # transforms.Grayscale(1),
            transforms.Resize((224, 224)),
            transforms.ToTensor(),
            transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])
    val_dataset = ImageFolder('./dataset/MMQ/Axial/val', transform=transform)
    val_loader = DataLoader(val_dataset,
                            batch_size=40,
                            shuffle=False,
                            num_workers=0)
    net = models.resnet18(pretrained=False)
    # net.conv1 = nn.Conv2d(1, 64, kernel_size=7, stride=2, padding=3, bias=False)
    in_channel = net.fc.in_features
    net.fc = nn.Linear(in_channel, 2)
    net.load_state_dict(torch.load('./model_weight/HCC_MMQ_Axial.pth', map_location=device))
    net.to(device)
    loss_function = nn.CrossEntropyLoss()
    net.eval()
    val_loss, correct, total =0, 0, 0
    all_labele, all_prescore, all_pre = np.array([]), np.array([]), np.array([])
    cfmetric = 0
    for step, batch in enumerate(val_loader):
        imgs, lables = batch
        imgs, lables = imgs.to(device), lables.to(device)
        with torch.no_grad():
            outputs = net(imgs)
        loss = loss_function(outputs, lables)
        val_loss += loss.item()
        predicted = torch.argmax(outputs, 1)
        total += lables.size(0)
        predicted_score = torch.max(outputs.data, 1)[0]
        correct += (predicted == lables).sum().item()
        all_pre = np.append(all_pre, np.array(predicted.to('cpu')))
        all_labele = np.append(all_labele, np.array(lables.to('cpu')))
        all_prescore = np.append(all_prescore, np.array(predicted_score.to('cpu')))

    cfmetric = confusion_matrix(all_labele, all_pre, labels=[1, 0])
    tn, fp, fn, tp = cfmetric.ravel()
    sensitivity = tp/(tp+fn)
    specificity = tn/(tn+fp)
    print(f'sensitivity:{sensitivity:.3f}, specificity:{specificity:.3f}')
    

if __name__=='__main__':
    prediction()