ResNet50的猫狗分类训练及预测

相比于之前写的ResNet18，下面的ResNet50写得更加工程化一点，这还适用与其他分类。

我的代码文件结构

1. 数据处理

　　首先已经对数据做好了分类

 　　文件夹结构是这样

　　开始划分数据集

　　split_data.py

import os
import random
import shutil


def move_file(target_path, save_train_path, save_val_pathm, scale=0.1):

    file_list = os.listdir(target_path)
    random.shuffle(file_list)

    number = int(len(file_list) * scale)
    train_list = file_list[number:]
    val_list = file_list[:number]

    for file in train_list:
        target_file_path = os.path.join(target_path, file)
        save_file_path = os.path.join(save_train_path, file)
        shutil.copyfile(target_file_path, save_file_path)
    for file in val_list:
        target_file_path = os.path.join(target_path, file)
        save_file_path = os.path.join(save_val_pathm, file)
        shutil.copyfile(target_file_path, save_file_path)


def split_classify_data(base_path, save_path, scale=0.1):
    folder_list = os.listdir(base_path)
    for folder in folder_list:
        target_path = os.path.join(base_path, folder)
        save_train_path = os.path.join(save_path, 'train', folder)
        save_val_path = os.path.join(save_path, 'val', folder)
        if not os.path.exists(save_train_path):
            os.makedirs(save_train_path)
        if not os.path.exists(save_val_path):
            os.makedirs(save_val_path)
        move_file(target_path, save_train_path, save_val_path, scale)
        print(folder, 'finish!')


if __name__ == '__main__':
    base_path = r'C:\Users\Administrator.DESKTOP-161KJQD\Desktop\save_dir'
    save_path = r'C:\Users\Administrator.DESKTOP-161KJQD\Desktop\dog_cat'
    # 验证集比例
    scale = 0.1
    split_classify_data(base_path, save_path, scale)

　　运行完以上代码的到的文件夹结构

 　　一个训练集数据，一个验证集数据

2.数据集的导入

　　我这个文件写了一个数据集的导入和一个学习率更新的函数。数据导入是通用的

　　tools.py

import os
import time

import cv2
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import torchvision
from torch.autograd.variable import Variable
from torch.utils.tensorboard import SummaryWriter
from torchvision import datasets, transforms
from torch.utils.data import Dataset, DataLoader
from torch.optim.lr_scheduler import ExponentialLR, LambdaLR
from torchvision.models import ResNet50_Weights
from tqdm import tqdm
from classify_cfg import *

mean = MEAN
std = STD


def get_dataset(base_dir='', input_size=160):
    dateset = dict()
    transform_train = transforms.Compose([
        # 分辨率重置为input_size
        transforms.Resize(input_size),
        transforms.RandomRotation(15),
        # 对加载的图像作归一化处理， 并裁剪为[input_sizexinput_sizex3]大小的图像(因为这图片像素不一致直接统一)
        transforms.CenterCrop(input_size),
        transforms.ToTensor(),
        transforms.Normalize(mean=mean, std=std)
    ])

    transform_val = transforms.Compose([
        transforms.Resize(input_size),
        transforms.RandomRotation(15),
        transforms.CenterCrop(input_size),
        transforms.ToTensor(),
        transforms.Normalize(mean=mean, std=std)
    ])
    base_dir_train = os.path.join(base_dir, 'train')
    train_dataset = datasets.ImageFolder(root=base_dir_train, transform=transform_train)
    # print("train_dataset=" + repr(train_dataset[1][0].size()))
    # print("train_dataset.class_to_idx=" + repr(train_dataset.class_to_idx))
    # print(train_dataset.classes)
    classes = train_dataset.classes
    # classes = train_dataset.class_to_idx
    classes_num = len(train_dataset.classes)

    base_dir_val = os.path.join(base_dir, 'val')
    val_dataset = datasets.ImageFolder(root=base_dir_val, transform=transform_val)

    dateset['train'] = train_dataset
    dateset['val'] = val_dataset

    return dateset, classes, classes_num


def update_lr(epoch, epochs):
    """
    假设开始的学习率lr是0.001，训练次数epochs是100
    当epoch<33时是lr * 1
    当33<=epoch<=66 时是lr * 0.5
    当66<epoch时是lr * 0.1
    """
    if epoch == 0 or epochs // 3 > epoch:
        return 1
    elif (epochs // 3 * 2 >= epoch) and (epochs // 3 <= epoch):
        return 0.5
    else:
        return 0.1

3.训练模型

　　数据集导入好了以后，选择模型，选择优化器等等，然后开始训练。

　　mytrain.py

import os
import time

import cv2
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
import torchvision
from torch.autograd.variable import Variable
from torch.utils.tensorboard import SummaryWriter
from torch.utils.data import Dataset, DataLoader
from torch.optim.lr_scheduler import ExponentialLR, LambdaLR
from torchvision.models import ResNet50_Weights
# from tqdm import tqdm
from classify_cfg import *
from tools import get_dataset, update_lr


def train(model, dateset, epochs, batch_size, device, optimizer, scheduler, criterion, save_path):
    train_loader = DataLoader(dateset.get('train'), batch_size=batch_size, shuffle=True)
    val_loader = DataLoader(dateset.get('val'), batch_size=batch_size, shuffle=True)

    # 保存为tensorboard文件
    write = SummaryWriter(save_path)
    # 训练过程写入txt
    f = open(os.path.join(save_path, 'log.txt'), 'w', encoding='utf-8')

    best_acc = 0
    for epoch in range(epochs):
        train_correct = 0.0
        model.train()
        sum_loss = 0.0
        accuracy = -1
        total_num = len(train_loader.dataset)
        # print(total_num, len(train_loader))
        # loop = tqdm(enumerate(train_loader), total=len(train_loader))
        batch_count = 0
        for batch_idx, (data, target) in enumerate(train_loader):
            start_time = time.time()
            data, target = Variable(data).to(device), Variable(target).to(device)
            output = model(data)
            loss = criterion(output, target)
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()

            print_loss = loss.data.item()
            sum_loss += print_loss
            train_predict = torch.max(output.data, 1)[1]
            if torch.cuda.is_available():
                train_correct += (train_predict.cuda() == target.cuda()).sum()
            else:
                train_correct += (train_predict == target).sum()
            accuracy = (train_correct / total_num) * 100
            # loop.set_description(f'Epoch [{epoch+1}/{epochs}]')
            # loop.set_postfix(loss=loss.item(), acc='{:.3f}'.format(accuracy))
            batch_count += len(data)
            end_time = time.time()
            s = f'Epoch:[{epoch+1}/{epochs}] Batch:[{batch_count}/{total_num}] train_acc: {"{:.2f}".format(accuracy)} ' \
                f'train_loss: {"{:.3f}".format(loss.item())} time: {int((end_time-start_time)*1000)} ms'
            # print(f'Epoch:[{epoch+1}/{epochs}]', f'Batch:[{batch_count}/{total_num}]',
            #       'train_acc:', '{:.2f}'.format(accuracy), 'train_loss:', '{:.3f}'.format(loss.item()),
            #       'time:', f'{int((end_time-start_time)*1000)} ms')
            print(s)
            f.write(s+'\n')

        write.add_scalar('train_acc', accuracy, epoch)
        write.add_scalar('train_loss', loss.item(), epoch)
        # print(optimizer.param_groups[0]['lr'])
        scheduler.step()
        if best_acc < accuracy:
            best_acc = accuracy
            torch.save(model, os.path.join(save_path, 'best.pt'))

        if epoch+1 == epochs:
            torch.save(model, os.path.join(save_path, 'last.pt'))

        # 预测验证集
        # if (epoch+1) % 5 == 0 or epoch+1 == epochs:
        model.eval()
        test_loss = 0.0
        correct = 0.0
        total_num = len(val_loader.dataset)
        # print(total_num, len(val_loader))
        with torch.no_grad():
            for data, target in val_loader:
                data, target = Variable(data).to(device), Variable(target).to(device)
                output = model(data)
                loss = criterion(output, target)
                _, pred = torch.max(output.data, 1)
                if torch.cuda.is_available():
                    correct += torch.sum(pred.cuda() == target.cuda())
                else:
                    correct += torch.sum(pred == target)
                print_loss = loss.data.item()
                test_loss += print_loss
            acc = correct / total_num * 100
            avg_loss = test_loss / len(val_loader)
        s = f"val acc: {'{:.2f}'.format(acc)} val loss: {'{:.3f}'.format(avg_loss)}"
        # print('val acc: ', '{:.2f}'.format(acc), 'val loss: ', '{:.3f}'.format(avg_loss))
        print(s)
        f.write(s+'\n')
        write.add_scalar('val_acc', acc, epoch)
        write.add_scalar('val_loss', avg_loss, epoch)
        # loop.set_postfix(val_loss='{:.3f}'.format(avg_loss), val_acc='{:.3f}'.format(acc))

    f.close()


if __name__ == '__main__':
    device = DEVICE
    epochs = EPOCHS
    batch_size = BATCH_SIZE
    input_size = INPUT_SIZE
    lr = LR
    # ---------------------------训练-------------------------------------
    # 图片的路径
    base_dir = r'C:\Users\Administrator.DESKTOP-161KJQD\Desktop\dog_cat'
    # 保存的路径
    save_path = r'C:\Users\Administrator.DESKTOP-161KJQD\Desktop\dog_cat_save'
    dateset, classes, classes_num = get_dataset(base_dir, input_size=input_size)
    # model = torchvision.models.resnet50(pretrained=True)
    model = torchvision.models.resnet50(weights=ResNet50_Weights.IMAGENET1K_V1)
    num_ftrs = model.fc.in_features
    model.fc = nn.Linear(num_ftrs, classes_num)
    model.to(DEVICE)
    # # 损失函数,交叉熵损失函数
    criteon = nn.CrossEntropyLoss()
    # 选择优化器
    optimizer = optim.SGD(model.parameters(), lr=lr)
    # 学习率更新
    # scheduler = ExponentialLR(optimizer, gamma=0.9)
    scheduler = LambdaLR(optimizer, lr_lambda=lambda epoch: update_lr(epoch, epochs))
    # 开始训练
    train(model, dateset, epochs, batch_size, device, optimizer, scheduler, criteon, save_path)
    # 将label保存起来
    with open(os.path.join(save_path, 'labels.txt'), 'w', encoding='utf-8') as f:
        f.write(f'{classes_num} {classes}')

　　训练结束以后，在保存路径下会得到下面的文件

　　最好的模型，最后一次的模型，标签的列表，训练的记录和tensorboard记录

　　在该路径下执行 tensorboard --logdir=.  

　　然后在浏览器打开给出的地址，即可看到数据训练过程的绘图

 4.对图片进行预测

　　考虑对于用户来说，用户是在网页或者手机上上传一张图片进行预测，所以这边是采用二进制数据。

　　mypredict.py

import cv2
import numpy as np
import torch

from classify_cfg import *



def img_process(img_betys, img_size, device):

    img_arry = np.asarray(bytearray(img_betys), dtype='uint8')
    # im0 = cv2.imread(img_betys)
    im0 = cv2.imdecode(img_arry, cv2.IMREAD_COLOR)
    image = cv2.resize(im0, (img_size, img_size))
    image = np.float32(image) / 255.0
    image[:, :, ] -= np.float32(mean)
    image[:, :, ] /= np.float32(std)
    image = image.transpose((2, 0, 1))
    im = torch.from_numpy(image).unsqueeze(0)
    im = im.to(device)
    return im


def predict(model_path, img, device):
    model = torch.load(model_path)
    model.to(device)
    model.eval()
    predicts = model(img)
    # print(predicts)
    _, preds = torch.max(predicts, 1)
    pred = torch.squeeze(preds)
    # print(pred)
    return pred


if __name__ == '__main__':
    mean = MEAN
    std = STD
    device = DEVICE
    classes = ['狗', '猫']
    # # 预测
    model_path = r'C:\Users\Administrator.DESKTOP-161KJQD\Desktop\dog_cat_save\best.pt'
    img_path = r'C:\Users\Administrator.DESKTOP-161KJQD\Desktop\save_dir\狗\000000.jpg'
    with open(img_path, 'rb') as f:
        img_betys = f.read()
    img =img_process(img_betys, 160, device)
    # print(img.shape)
    # print(img)
    pred = predict(model_path, img, device)
    print(classes[int(pred)])