Pytorch DataLoader shuffle验证方式

shuffle = False时，不打乱数据顺序

shuffle = True，随机打乱

import numpy as np
import h5py
import torch
from torch.utils.data import DataLoader, Dataset  
h5f = h5py.File('train.h5', 'w');
data1 = np.array([[1,2,3],
               [2,5,6],
              [3,5,6],
              [4,5,6]])
data2 = np.array([[1,1,1],
                   [1,2,6],
                  [1,3,6],
                  [1,4,6]])
h5f.create_dataset(str('data'), data=data1)
h5f.create_dataset(str('label'), data=data2)
class Dataset(Dataset):
    def __init__(self):
        h5f = h5py.File('train.h5', 'r')
        self.data = h5f['data']
        self.label = h5f['label']
    def __getitem__(self, index):
        data = torch.from_numpy(self.data[index])
        label = torch.from_numpy(self.label[index])
        return data, label
 
    def __len__(self):
        assert self.data.shape[0] == self.label.shape[0], "wrong data length"
        return self.data.shape[0] 
 
dataset_train = Dataset()
loader_train = DataLoader(dataset=dataset_train,
                           batch_size=2,
                           shuffle = True)
 
for i, data in enumerate(loader_train):
    train_data, label = data
    print(train_data)
 

pytorch DataLoader使用细节

背景：

我一开始是对数据扩增这一块有疑问, 只看到了数据变换(torchvisiom.transforms),但是没看到数据扩增, 后来搞明白了, 数据扩增在pytorch指的是torchvisiom.transforms + torch.utils.data.DataLoader+多个epoch共同作用下完成的,

数据变换共有以下内容

composed = transforms.Compose([transforms.Resize((448, 448)), #  resize
                               transforms.RandomCrop(300), # random crop
                               transforms.ToTensor(),
                               transforms.Normalize(mean=[0.5, 0.5, 0.5],  # normalize
                                                    std=[0.5, 0.5, 0.5])])

简单的数据读取类, 进返回PIL格式的image:

class MyDataset(data.Dataset):    
    def __init__(self, labels_file, root_dir, transform=None):
        with open(labels_file) as csvfile:
            self.labels_file = list(csv.reader(csvfile))
        self.root_dir = root_dir
        self.transform = transform
        
    def __len__(self):
        return len(self.labels_file)
    
    def __getitem__(self, idx):
        im_name = os.path.join(root_dir, self.labels_file[idx][0])
        im = Image.open(im_name)
        
        if self.transform:
            im = self.transform(im)
            
        return im

下面是主程序

labels_file = "F:/test_temp/labels.csv"
root_dir = "F:/test_temp"
dataset_transform = MyDataset(labels_file, root_dir, transform=composed)
dataloader = data.DataLoader(dataset_transform, batch_size=1, shuffle=False)
"""原始数据集共3张图片, 以batch_size=1, epoch为2 展示所有图片(共6张)  """
for eopch in range(2):
    plt.figure(figsize=(6, 6)) 
    for ind, i in enumerate(dataloader):
        a = i[0, :, :, :].numpy().transpose((1, 2, 0))
        plt.subplot(1, 3, ind+1)
        plt.imshow(a)

从上述图片总可以看到, 在每个eopch阶段实际上是对原始图片重新使用了transform, , 这就造就了数据的扩增

以上为个人经验，希望能给大家一个参考，也希望大家多多支持猪先飞。