自己写dataloader，pytorch数据集读取，resnet50在standford cars数据集上

 作者用resnet50跑了一下standford cars数据集。该数据集共有196类，16185张图片，其中训练集有8144张，测试集有8041张图片。

 训练集是这样的：  

       train/00001.jpg

       train/00002.jpg

       ……

 对应的标签放在另一个mat文件中。

读取mat文件的代码如下，将其写进txt文件，一行只有一个正整数，表示对应的一个标签

import scipy.io

data = scipy.io.loadmat('cars_train_annos.mat')
annotations = data['annotations']
f_train = open('./train.txt','w')

for i in range(annotations.shape[1]):
    num = int(annotations[0,i][4])
    num = str(num)
    print(i,num)
    f_train.write(num+'\n')

f_train.close()

有了图像和对应的标签之后，就可以开始写dataloader类了，代码如下：

import torch
import torch.nn as nn
import torch.optim as optim
from torch.optim import lr_scheduler
from torch.autograd import Variable

from torch.utils.data import Dataset
from torchvision import models,transforms
import os
import time

from PIL import Image


# use PIL Image to read iamge
def default_loader(path):
    try:
        img = Image.open(path)
        return img.convert('RGB')
    except:
        print("Cannot read image:{}".format(path))


class customData(Dataset):
    def __init__(self,img_path,txt_path,dataset='',data_transforms=None,loader=default_loader):
        with open(txt_path) as input_file:
            lines = input_file.readlines()
            self.img_label = [int(line.strip()) for line in lines]
        self.img_name = []
        for root,dirs,files in os.walk(img_path):
            for name in sorted(files):
                self.img_name.append(os.path.join(img_path,name))
        self.data_transforms = data_transforms
        self.dataset = dataset
        self.loader = loader
        # 最主要的是将图片路径存入self.img_name,将对应的标签写入self.img_label里面

    def __len__(self):
        return len(self.img_name)

    def __getitem__(self, item):
        img_name = self.img_name[item]
        label = self.img_label[item]
        img = self.loader(img_name)

        if self.data_transforms is not None:
            try:
                img = self.data_transforms[self.dataset](img)
            except:
                print("Cannot transform images: ()".format(img_name))
        return img,label
        #每次读出一个图片和标签

下面定义train的主函数，因为没有设置验证集，下面代码中关于val的都需要注释掉，

def train_model(model, criterion, optimizer, scheduler, num_epochs, use_gpu):
    since = time.time()
    best_model_wts = model.state_dict()
    best_acc = 0.0

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

        # Each epoch has a training and validation phase
        # for phase in ['train','val']:
        for phase in ['train']:
            count_batch = 0
            if phase == 'train':
                scheduler.step()
                model.train(True)  # Set model to training mode
            else:
                model.train(False)  # Set model to evaluate mode

            running_loss = 0.0
            running_corrects = 0.0

            # Iterate over data.
            for data in dataloders[phase]:
                count_batch += 1
                # get the inputs
                inputs, labels = data

                # wrap them in Variable
                if use_gpu:
                    inputs = Variable(inputs.cuda())
                    labels = Variable(labels.cuda())
                else:
                    inputs, labels = Variable(inputs), Variable(labels)

                # zero the parameter gradients
                optimizer.zero_grad()

                # forward
                outputs = model(inputs)
                # print(outputs.data)
                _, preds = torch.max(outputs.data, 1)
                loss = criterion(outputs, labels)

                # backward + optimize only if in training phase
                if phase == 'train':
                    loss.backward()
                    optimizer.step()
                # statistics
                running_loss += loss.item()
                running_corrects += torch.sum(preds == labels.data).to(torch.float32)

                # print result every 10 batch
                if count_batch%10 == 0:
                    batch_loss = running_loss / (batch_size*count_batch)
                    batch_acc = running_corrects / (batch_size*count_batch)
                    print('{} Epoch [{}] Batch [{}] Loss: {:.4f} Acc: {:.4f} Time: {:.4f}s'. \
                          format(phase, epoch, count_batch, batch_loss, batch_acc, time.time()-begin_time))
                    begin_time = time.time()

            epoch_loss = running_loss / dataset_sizes[phase]
            epoch_acc = running_corrects / dataset_sizes[phase]
            print('{} Loss: {:.4f} Acc: {:.4f}'.format(phase, epoch_loss, epoch_acc))

            # save model
            if phase == 'train':
                if not os.path.exists('output'):
                    os.makedirs('output')
                torch.save(model, 'output/resnet_epoch{}.pkl'.format(epoch))

            # deep copy the model
            # if phase == 'val' and epoch_acc > best_acc:
            if epoch_acc > best_acc:
                best_acc = epoch_acc
                best_model_wts = model.state_dict()

    time_elapsed = time.time() - since
    print('Training complete in {:.0f}m {:.0f}s'.format(
        time_elapsed // 60, time_elapsed % 60))
    # print('Best val Acc: {:4f}'.format(best_acc))
    print('Best acc: {:4f}'.format(best_acc))
    # load best model weights
    model.load_state_dict(best_model_wts)
    return model

最后是主函数，同样，val的部分也注释掉了

if __name__ == '__main__':

    data_transforms = {
        'train': transforms.Compose([
            transforms.RandomResizedCrop(224),
            transforms.RandomHorizontalFlip(),
            transforms.ToTensor(),
            transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
        ]),
       # 'val': transforms.Compose([
           # transforms.Resize(256),
           # transforms.CenterCrop(224),
           # transforms.ToTensor(),
           # transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
       # ]),
    }

    # use_gpu = torch.cuda.is_available()   # 用gpu跑的话，可以把注释去掉
    # print(use_gpu)

    batch_size = 32
    num_class = 197  # 用196会报错

    # image_datasets = {x: customData(img_path='/ImagePath',
    #                                 txt_path=('/TxtFile/' + x + '.txt'),
    #                                 data_transforms=data_transforms,
    #                                 dataset=x) for x in ['train', 'val']}

    image_datasets = {x: customData(img_path='./'+x,
                                    txt_path='./'+ x + '.txt',
                                    data_transforms=data_transforms,
                                    dataset=x) for x in ['train']}  # ['train', 'val']

    # wrap your data and label into Tensor
    dataloders = {x: torch.utils.data.DataLoader(image_datasets[x],
                                                 batch_size=batch_size,
                                                 shuffle=True) for x in ['train']} # ['train', 'val']

    dataset_sizes = {x: len(image_datasets[x]) for x in ['train']}  # ['train', 'val']

    # get model and replace the original fc layer with your fc layer
    model_ft = models.resnet50(pretrained=True)
    num_ftrs = model_ft.fc.in_features
    # print(num_ftrs)
    model_ft.fc = nn.Linear(num_ftrs, num_class)   # 将resnet最后全连接层的输出换成数据集的类别总数

    # if use gpu
    # if use_gpu:
    #     model_ft = model_ft.cuda()

    # define cost function
    criterion = nn.CrossEntropyLoss()

    # Observe that all parameters are being optimized
    optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.005, momentum=0.9)

    # Decay LR by a factor of 0.2 every 5 epochs
    exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=5, gamma=0.2)

    # multi-GPU
    # model_ft = torch.nn.DataParallel(model_ft, device_ids=[0,1])

    # train model
    model_ft = train_model(model=model_ft,
                           criterion=criterion,
                           optimizer=optimizer_ft,
                           scheduler=exp_lr_scheduler,
                           num_epochs=25,
                           use_gpu=False)




    # save best model
    torch.save(model_ft,"output/best_resnet.pkl")

如果需要用gpu来跑代码，其中主函数关于gpu的部分可以把注释给消掉。

另外，作者发现这个代码训练的准确率不够高。

实验发现，将
transform.RandomResizedCrop(224)
改成
transform.Resized((256,256),interpolation=3)
transform.RandomCrop(224)
准确率将大幅提升，能达到0.98左右