Python利用Pytorch实现绘制ROC与PR曲线图

Pytorch 多分类模型绘制 ROC, PR 曲线（代码 亲测 可用）

ROC曲线

示例代码

import torch
import torch.nn as nn
import os
import numpy as np
from torchvision.datasets import ImageFolder
from utils.transform import get_transform_for_test
from senet.se_resnet import FineTuneSEResnet50
from scipy import interp
import matplotlib.pyplot as plt
from itertools import cycle
from sklearn.metrics import roc_curve, auc, f1_score, precision_recall_curve, average_precision_score

os.environ['CUDA_VISIBLE_DEVICES'] = "0"

data_root = r'D:\TJU\GBDB\set113\set113_images\test1'    # 测试集路径
test_weights_path = r"C:\Users\admin\Desktop\fsdownload\epoch_0278_top1_70.565_'checkpoint.pth.tar'"    # 预训练模型参数
num_class = 113    # 类别数量
gpu = "cuda:0"  

# mean=[0.948078, 0.93855226, 0.9332005], var=[0.14589554, 0.17054074, 0.18254866]
def test(model, test_path):
   # 加载测试集和预训练模型参数
   test_dir = os.path.join(data_root, 'test_images')
   class_list = list(os.listdir(test_dir))
   class_list.sort()
   transform_test = get_transform_for_test(mean=[0.948078, 0.93855226, 0.9332005],
                                           var=[0.14589554, 0.17054074, 0.18254866])
   test_dataset = ImageFolder(test_dir, transform=transform_test)
   test_loader = torch.utils.data.DataLoader(
       test_dataset, batch_size=1, shuffle=False, drop_last=False, pin_memory=True, num_workers=1)
   checkpoint = torch.load(test_path)
   model.load_state_dict(checkpoint['state_dict'])
   model.eval()

score_list = []     # 存储预测得分
   label_list = []     # 存储真实标签
   for i, (inputs, labels) in enumerate(test_loader):
       inputs = inputs.cuda()
       labels = labels.cuda()

outputs = model(inputs)
       # prob_tmp = torch.nn.Softmax(dim=1)(outputs) # (batchsize, nclass)
       score_tmp = outputs  # (batchsize, nclass)

score_list.extend(score_tmp.detach().cpu().numpy())
       label_list.extend(labels.cpu().numpy())

score_array = np.array(score_list)
   # 将label转换成onehot形式
   label_tensor = torch.tensor(label_list)
   label_tensor = label_tensor.reshape((label_tensor.shape[0], 1))
   label_onehot = torch.zeros(label_tensor.shape[0], num_class)
   label_onehot.scatter_(dim=1, index=label_tensor, value=1)
   label_onehot = np.array(label_onehot)

print("score_array:", score_array.shape)  # (batchsize, classnum)
   print("label_onehot:", label_onehot.shape)  # torch.Size([batchsize, classnum])

# 调用sklearn库，计算每个类别对应的fpr和tpr
   fpr_dict = dict()
   tpr_dict = dict()
   roc_auc_dict = dict()
   for i in range(num_class):
       fpr_dict[i], tpr_dict[i], _ = roc_curve(label_onehot[:, i], score_array[:, i])
       roc_auc_dict[i] = auc(fpr_dict[i], tpr_dict[i])
   # micro
   fpr_dict["micro"], tpr_dict["micro"], _ = roc_curve(label_onehot.ravel(), score_array.ravel())
   roc_auc_dict["micro"] = auc(fpr_dict["micro"], tpr_dict["micro"])

# macro
   # First aggregate all false positive rates
   all_fpr = np.unique(np.concatenate([fpr_dict[i] for i in range(num_class)]))
   # Then interpolate all ROC curves at this points
   mean_tpr = np.zeros_like(all_fpr)
   for i in range(num_class):
       mean_tpr += interp(all_fpr, fpr_dict[i], tpr_dict[i])
   # Finally average it and compute AUC
   mean_tpr /= num_class
   fpr_dict["macro"] = all_fpr
   tpr_dict["macro"] = mean_tpr
   roc_auc_dict["macro"] = auc(fpr_dict["macro"], tpr_dict["macro"])

# 绘制所有类别平均的roc曲线
   plt.figure()
   lw = 2
   plt.plot(fpr_dict["micro"], tpr_dict["micro"],
            label='micro-average ROC curve (area = {0:0.2f})'
                  ''.format(roc_auc_dict["micro"]),
            color='deeppink', linestyle=':', linewidth=4)

plt.plot(fpr_dict["macro"], tpr_dict["macro"],
            label='macro-average ROC curve (area = {0:0.2f})'
                  ''.format(roc_auc_dict["macro"]),
            color='navy', linestyle=':', linewidth=4)

colors = cycle(['aqua', 'darkorange', 'cornflowerblue'])
   for i, color in zip(range(num_class), colors):
       plt.plot(fpr_dict[i], tpr_dict[i], color=color, lw=lw,
                label='ROC curve of class {0} (area = {1:0.2f})'
                      ''.format(i, roc_auc_dict[i]))
   plt.plot([0, 1], [0, 1], 'k--', lw=lw)
   plt.xlim([0.0, 1.0])
   plt.ylim([0.0, 1.05])
   plt.xlabel('False Positive Rate')
   plt.ylabel('True Positive Rate')
   plt.title('Some extension of Receiver operating characteristic to multi-class')
   plt.legend(loc="lower right")
   plt.savefig('set113_roc.jpg')
   plt.show()

if __name__ == '__main__':
   # 加载模型
   seresnet = FineTuneSEResnet50(num_class=num_class)
   device = torch.device(gpu)
   seresnet = seresnet.to(device)
   test(seresnet, test_weights_path)

运行结果：

PR曲线

示例代码

import torch
import torch.nn as nn
import os
import numpy as np
from torchvision.datasets import ImageFolder
from utils.transform import get_transform_for_test
from senet.se_resnet import FineTuneSEResnet50
import matplotlib.pyplot as plt
from sklearn.metrics import roc_curve, auc, f1_score, precision_recall_curve, average_precision_score

os.environ['CUDA_VISIBLE_DEVICES'] = "0"

data_root = r'D:\TJU\GBDB\set113\set113_images\test1'    # 测试集路径
test_weights_path = r"C:\Users\admin\Desktop\fsdownload\epoch_0278_top1_70.565_'checkpoint.pth.tar'"    # 预训练模型参数
num_class = 113    # 类别数量
gpu = "cuda:0"    

# mean=[0.948078, 0.93855226, 0.9332005], var=[0.14589554, 0.17054074, 0.18254866]
def test(model, test_path):
   # 加载测试集和预训练模型参数
   test_dir = os.path.join(data_root, 'test_images')
   class_list = list(os.listdir(test_dir))
   class_list.sort()
   transform_test = get_transform_for_test(mean=[0.948078, 0.93855226, 0.9332005],
                                           var=[0.14589554, 0.17054074, 0.18254866])
   test_dataset = ImageFolder(test_dir, transform=transform_test)
   test_loader = torch.utils.data.DataLoader(
       test_dataset, batch_size=1, shuffle=False, drop_last=False, pin_memory=True, num_workers=1)
   checkpoint = torch.load(test_path)
   model.load_state_dict(checkpoint['state_dict'])
   model.eval()

score_list = []     # 存储预测得分
   label_list = []     # 存储真实标签
   for i, (inputs, labels) in enumerate(test_loader):
       inputs = inputs.cuda()
       labels = labels.cuda()

outputs = model(inputs)
       # prob_tmp = torch.nn.Softmax(dim=1)(outputs) # (batchsize, nclass)
       score_tmp = outputs  # (batchsize, nclass)

score_list.extend(score_tmp.detach().cpu().numpy())
       label_list.extend(labels.cpu().numpy())

score_array = np.array(score_list)
   # 将label转换成onehot形式
   label_tensor = torch.tensor(label_list)
   label_tensor = label_tensor.reshape((label_tensor.shape[0], 1))
   label_onehot = torch.zeros(label_tensor.shape[0], num_class)
   label_onehot.scatter_(dim=1, index=label_tensor, value=1)
   label_onehot = np.array(label_onehot)
   print("score_array:", score_array.shape)  # (batchsize, classnum) softmax
   print("label_onehot:", label_onehot.shape)  # torch.Size([batchsize, classnum]) onehot

# 调用sklearn库，计算每个类别对应的precision和recall
   precision_dict = dict()
   recall_dict = dict()
   average_precision_dict = dict()
   for i in range(num_class):
       precision_dict[i], recall_dict[i], _ = precision_recall_curve(label_onehot[:, i], score_array[:, i])
       average_precision_dict[i] = average_precision_score(label_onehot[:, i], score_array[:, i])
       print(precision_dict[i].shape, recall_dict[i].shape, average_precision_dict[i])

# micro
   precision_dict["micro"], recall_dict["micro"], _ = precision_recall_curve(label_onehot.ravel(),
                                                                             score_array.ravel())
   average_precision_dict["micro"] = average_precision_score(label_onehot, score_array, average="micro")
   print('Average precision score, micro-averaged over all classes: {0:0.2f}'.format(average_precision_dict["micro"]))

# 绘制所有类别平均的pr曲线
   plt.figure()
   plt.step(recall_dict['micro'], precision_dict['micro'], where='post')

plt.xlabel('Recall')
   plt.ylabel('Precision')
   plt.ylim([0.0, 1.05])
   plt.xlim([0.0, 1.0])
   plt.title(
       'Average precision score, micro-averaged over all classes: AP={0:0.2f}'
       .format(average_precision_dict["micro"]))
   plt.savefig("set113_pr_curve.jpg")
   # plt.show()

if __name__ == '__main__':
   # 加载模型
   seresnet = FineTuneSEResnet50(num_class=num_class)
   device = torch.device(gpu)
   seresnet = seresnet.to(device)
   test(seresnet, test_weights_path)

运行结果：

来源：https://blog.csdn.net/Vertira/article/details/128482515