2.3 [pytorch] data preprocessing torchvision datasets. ImageFolder

When we were training , We have to read in the data , Different storage methods for data , We also have different reading methods , Thus, it is convenient for us to match the training data with their labels one by one .

Mode one ： It is used to read data from a folder containing all categories of data . By rewriting torch.utils.data.Dataset, Build data reading mode （ Do your own processing and match the data and labels one by one ）, Finally, through the iterator torch.utils.data.DataLoader Call to , according to batch_size Read data in batches . Here are two examples ：

• 【kaggle Data sets - dog breed give an example 】 Data preprocessing 、 rewrite Dataset、DataLoader Reading data
• torch.utils.data.Dataset and torch.utils.data.DataLoader Based on using

Mode two ： This is what we want to introduce in this article torchcvision.datasets.ImageFolder, It is used to read data from the classified folder , Examples are as follows

Data storage structure ：

from torchvision import datasets, transforms

my_trans = transforms.Compose([transforms.RandomResizedCrop(224),
                               transforms.ToTensor()])

train_data = datasets.ImageFolder('./animals/train', transform=my_trans)
print(train_data.class_to_idx)    #  View category name ( file name )  Corresponding tag value 
print(train_data[0][0].size())    #  The size of the first picture , It's us  transforms.RandomResizedCrop  Set crop size 

for i in range(len(train_data)):
    print(train_data[i][1])   #  View the tag values mapped to all pictures in the training dataset 

Output ：