基于WIN10的64位系统演示

一、写在前面

（1）Convolutional Vision Transformers

Convolutional Vision Transformer（ConViT）是一种结合了卷积神经网络（Convolutional Neural Networks，简称CNN）和Transformer结构的深度学习模型，由Facebook AI在2021年提出。该模型旨在利用CNN在局部特征提取方面的优势，同时保留Transformer在长距离依赖性处理上的优点，以此提升图像识别任务的性能。

在传统的Vision Transformer（ViT）模型中，输入图像会被划分为一系列的小块或者"patches"，然后用Transformer处理这些独立的patches。这种方法的一个问题是，它并未充分利用图像的局部信息，这些信息在CNN模型中得到了很好的处理。

ConViT模型提出了一种解决方案，即在每个Transformer的自注意力（Self-Attention）机制中引入卷积的归纳偏置（Convolutional Inductive Bias）。这种设计使得模型在处理每个patch时，能更加考虑到它与周围patch的关系，从而更好地捕捉图像的局部信息。同时，由于Transformer的全局自注意力机制，ConViT也能处理图像的长距离依赖性。

（2）Convolutional Vision Transformers的码源

本文继续使用Facebook的高级深度学习框架PyTorchImageModels (timm)，网址为：

https://github.com/huggingface/pytorch-image-models/blob/main/timm/models/convit.py

二、Convolutional Vision Transformers迁移学习代码实战

我们继续胸片的数据集：肺结核病人和健康人的胸片的识别。其中，肺结核病人700张，健康人900张，分别存入单独的文件夹中。

（a）导入包

import copy
import torch
import torchvision
import torchvision.transforms as transforms
from torchvision import models
from torch.utils.data import DataLoader
from torch import optim, nn
from torch.optim import lr_scheduler
import os
import matplotlib.pyplot as plt
import warnings
import numpy as np

warnings.filterwarnings("ignore")
plt.rcParams['font.sans-serif'] = ['SimHei']
plt.rcParams['axes.unicode_minus'] = False

# 设置GPU
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

（b）导入数据集

import torch
from torchvision import datasets, transforms
import os

# 数据集路径
data_dir = "./MTB"

# 图像的大小
img_height = 100
img_width = 100

# 数据预处理
data_transforms = {
    'train': transforms.Compose([
        transforms.RandomResizedCrop(img_height),
        transforms.RandomHorizontalFlip(),
        transforms.RandomVerticalFlip(),
        transforms.RandomRotation(0.2),
        transforms.ToTensor(),
        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
    ]),
    'val': transforms.Compose([
        transforms.Resize((img_height, img_width)),
        transforms.ToTensor(),
        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
    ]),
}

# 加载数据集
full_dataset = datasets.ImageFolder(data_dir)

# 获取数据集的大小
full_size = len(full_dataset)
train_size = int(0.7 * full_size)  # 假设训练集占80%
val_size = full_size - train_size  # 验证集的大小

# 随机分割数据集
torch.manual_seed(0)  # 设置随机种子以确保结果可重复
train_dataset, val_dataset = torch.utils.data.random_split(full_dataset, [train_size, val_size])

# 将数据增强应用到训练集
train_dataset.dataset.transform = data_transforms['train']

# 创建数据加载器
batch_size = 32
train_dataloader = torch.utils.data.DataLoader(train_dataset, batch_size=batch_size, shuffle=True, num_workers=4)
val_dataloader = torch.utils.data.DataLoader(val_dataset, batch_size=batch_size, shuffle=True, num_workers=4)

dataloaders = {'train': train_dataloader, 'val': val_dataloader}
dataset_sizes = {'train': len(train_dataset), 'val': len(val_dataset)}
class_names = full_dataset.classes

（c）导入Convolutional Vision Transformers

# 导入所需的库
import torch.nn as nn
import timm

# 定义Convolutional Vision Transformer模型
model = timm.create_model('convit_base', pretrained=True)  # 你可以选择适合你需求的ConViT版本，这里以'convit_base'为例
num_ftrs = model.head.in_features

# 根据分类任务修改最后一层
model.head = nn.Linear(num_ftrs, len(class_names))

# 将模型移至指定设备
model = model.to(device)

# 打印模型摘要
print(model)

（d）编译模型

# 定义损失函数
criterion = nn.CrossEntropyLoss()

# 定义优化器
optimizer = optim.Adam(model.parameters())

# 定义学习率调度器
exp_lr_scheduler = lr_scheduler.StepLR(optimizer, step_size=7, gamma=0.1)

# 开始训练模型
num_epochs = 10
best_model_wts = copy.deepcopy(model.state_dict())
best_acc = 0.0

# 初始化记录器
train_loss_history = []
train_acc_history = []
val_loss_history = []
val_acc_history = []

for epoch in range(num_epochs):
    print('Epoch {}/{}'.format(epoch, num_epochs - 1))
    print('-' * 10)

    # 每个epoch都有一个训练和验证阶段
    for phase in ['train', 'val']:
        if phase == 'train':
            model.train()  # Set model to training mode
        else:
            model.eval()   # Set model to evaluate mode

        running_loss = 0.0
        running_corrects = 0

        # 遍历数据
        for inputs, labels in dataloaders[phase]:
            inputs = inputs.to(device)
            labels = labels.to(device)

            # 零参数梯度
            optimizer.zero_grad()

            # 前向
            with torch.set_grad_enabled(phase == 'train'):
                outputs = model(inputs)
                _, preds = torch.max(outputs, 1)
                loss = criterion(outputs, labels)

                # 只在训练模式下进行反向和优化
                if phase == 'train':
                    loss.backward()
                    optimizer.step()

            # 统计
            running_loss += loss.item() * inputs.size(0)
            running_corrects += torch.sum(preds == labels.data)

        epoch_loss = running_loss / dataset_sizes[phase]
        epoch_acc = (running_corrects.double() / dataset_sizes[phase]).item()

        # 记录每个epoch的loss和accuracy
        if phase == 'train':
            train_loss_history.append(epoch_loss)
            train_acc_history.append(epoch_acc)
        else:
            val_loss_history.append(epoch_loss)
            val_acc_history.append(epoch_acc)

        print('{} Loss: {:.4f} Acc: {:.4f}'.format(phase, epoch_loss, epoch_acc))

        # 深拷贝模型
        if phase == 'val' and epoch_acc > best_acc:
            best_acc = epoch_acc
            best_model_wts = copy.deepcopy(model.state_dict())

    print()

print('Best val Acc: {:4f}'.format(best_acc))

（e）Accuracy和Loss可视化

epoch = range(1, len(train_loss_history)+1)

fig, ax = plt.subplots(1, 2, figsize=(10,4))
ax[0].plot(epoch, train_loss_history, label='Train loss')
ax[0].plot(epoch, val_loss_history, label='Validation loss')
ax[0].set_xlabel('Epochs')
ax[0].set_ylabel('Loss')
ax[0].legend()

ax[1].plot(epoch, train_acc_history, label='Train acc')
ax[1].plot(epoch, val_acc_history, label='Validation acc')
ax[1].set_xlabel('Epochs')
ax[1].set_ylabel('Accuracy')
ax[1].legend()

#plt.savefig("loss-acc.pdf", dpi=300,format="pdf")

观察模型训练情况：

蓝色为训练集，橙色为验证集。

（f）混淆矩阵可视化以及模型参数

from sklearn.metrics import classification_report, confusion_matrix
import math
import pandas as pd
import numpy as np
import seaborn as sns
from matplotlib.pyplot import imshow

# 定义一个绘制混淆矩阵图的函数
def plot_cm(labels, predictions):
    
    # 生成混淆矩阵
    conf_numpy = confusion_matrix(labels, predictions)
    # 将矩阵转化为 DataFrame
    conf_df = pd.DataFrame(conf_numpy, index=class_names ,columns=class_names)  
    
    plt.figure(figsize=(8,7))
    
    sns.heatmap(conf_df, annot=True, fmt="d", cmap="BuPu")
    
    plt.title('Confusion matrix',fontsize=15)
    plt.ylabel('Actual value',fontsize=14)
    plt.xlabel('Predictive value',fontsize=14)
    
def evaluate_model(model, dataloader, device):
    model.eval()   # 设置模型为评估模式
    true_labels = []
    pred_labels = []
    # 遍历数据
    for inputs, labels in dataloader:
        inputs = inputs.to(device)
        labels = labels.to(device)

        # 前向
        with torch.no_grad():
            outputs = model(inputs)
            _, preds = torch.max(outputs, 1)

        true_labels.extend(labels.cpu().numpy())
        pred_labels.extend(preds.cpu().numpy())
        
    return true_labels, pred_labels

# 获取预测和真实标签
true_labels, pred_labels = evaluate_model(model, dataloaders['val'], device)

# 计算混淆矩阵
cm_val = confusion_matrix(true_labels, pred_labels)
a_val = cm_val[0,0]
b_val = cm_val[0,1]
c_val = cm_val[1,0]
d_val = cm_val[1,1]

# 计算各种性能指标
acc_val = (a_val+d_val)/(a_val+b_val+c_val+d_val)  # 准确率
error_rate_val = 1 - acc_val  # 错误率
sen_val = d_val/(d_val+c_val)  # 灵敏度
sep_val = a_val/(a_val+b_val)  # 特异度
precision_val = d_val/(b_val+d_val)  # 精确度
F1_val = (2*precision_val*sen_val)/(precision_val+sen_val)  # F1值
MCC_val = (d_val*a_val-b_val*c_val) / (np.sqrt((d_val+b_val)*(d_val+c_val)*(a_val+b_val)*(a_val+c_val)))  # 马修斯相关系数

# 打印出性能指标
print("验证集的灵敏度为：", sen_val, 
      "验证集的特异度为：", sep_val,
      "验证集的准确率为：", acc_val, 
      "验证集的错误率为：", error_rate_val,
      "验证集的精确度为：", precision_val, 
      "验证集的F1为：", F1_val,
      "验证集的MCC为：", MCC_val)

# 绘制混淆矩阵
plot_cm(true_labels, pred_labels)

    
# 获取预测和真实标签
train_true_labels, train_pred_labels = evaluate_model(model, dataloaders['train'], device)
# 计算混淆矩阵
cm_train = confusion_matrix(train_true_labels, train_pred_labels)  
a_train = cm_train[0,0]
b_train = cm_train[0,1]
c_train = cm_train[1,0]
d_train = cm_train[1,1]
acc_train = (a_train+d_train)/(a_train+b_train+c_train+d_train)
error_rate_train = 1 - acc_train
sen_train = d_train/(d_train+c_train)
sep_train = a_train/(a_train+b_train)
precision_train = d_train/(b_train+d_train)
F1_train = (2*precision_train*sen_train)/(precision_train+sen_train)
MCC_train = (d_train*a_train-b_train*c_train) / (math.sqrt((d_train+b_train)*(d_train+c_train)*(a_train+b_train)*(a_train+c_train))) 
print("训练集的灵敏度为：",sen_train, 
      "训练集的特异度为：",sep_train,
      "训练集的准确率为：",acc_train, 
      "训练集的错误率为：",error_rate_train,
      "训练集的精确度为：",precision_train, 
      "训练集的F1为：",F1_train,
      "训练集的MCC为：",MCC_train)

# 绘制混淆矩阵
plot_cm(train_true_labels, train_pred_labels)

效果不错：

（g）AUC曲线绘制

from sklearn import metrics
import numpy as np
import matplotlib.pyplot as plt
from matplotlib.pyplot import imshow
from sklearn.metrics import classification_report, confusion_matrix
import seaborn as sns
import pandas as pd
import math

def plot_roc(name, labels, predictions, **kwargs):
    fp, tp, _ = metrics.roc_curve(labels, predictions)

    plt.plot(fp, tp, label=name, linewidth=2, **kwargs)
    plt.plot([0, 1], [0, 1], color='orange', linestyle='--')
    plt.xlabel('False positives rate')
    plt.ylabel('True positives rate')
    ax = plt.gca()
    ax.set_aspect('equal')


# 确保模型处于评估模式
model.eval()

train_ds = dataloaders['train']
val_ds = dataloaders['val']

val_pre_auc   = []
val_label_auc = []

for images, labels in val_ds:
    for image, label in zip(images, labels):      
        img_array = image.unsqueeze(0).to(device)  # 在第0维增加一个维度并将图像转移到适当的设备上
        prediction_auc = model(img_array)  # 使用模型进行预测
        val_pre_auc.append(prediction_auc.detach().cpu().numpy()[:,1])
        val_label_auc.append(label.item())  # 使用Tensor.item()获取Tensor的值
auc_score_val = metrics.roc_auc_score(val_label_auc, val_pre_auc)


train_pre_auc   = []
train_label_auc = []

for images, labels in train_ds:
    for image, label in zip(images, labels):
        img_array_train = image.unsqueeze(0).to(device) 
        prediction_auc = model(img_array_train)
        train_pre_auc.append(prediction_auc.detach().cpu().numpy()[:,1])  # 输出概率而不是标签！
        train_label_auc.append(label.item())
auc_score_train = metrics.roc_auc_score(train_label_auc, train_pre_auc)

plot_roc('validation AUC: {0:.4f}'.format(auc_score_val), val_label_auc , val_pre_auc , color="red", linestyle='--')
plot_roc('training AUC: {0:.4f}'.format(auc_score_train), train_label_auc, train_pre_auc, color="blue", linestyle='--')
plt.legend(loc='lower right')
#plt.savefig("roc.pdf", dpi=300,format="pdf")

print("训练集的AUC值为：",auc_score_train, "验证集的AUC值为：",auc_score_val)

ROC曲线如下：