Pytorch技法之继承Subset类完成自定义数据拆分

from torchvision.datasets import FashionMNIST
from torchvision.transforms import Compose, ToTensor, Normalize
RAW_DATA_PATH = './rawdata'
transform = Compose(
        [ToTensor(),
         Normalize((0.1307,), (0.3081,))
         ]
    )
train_data = FashionMNIST(
        root=RAW_DATA_PATH,
        download=True,
        train=True,
        transform=transform
    )

classes = train_data.classes
num_features = train_data.data[0].shape[0]
train_labels = train_data.targets

print(classes)
print(num_features)
print(train_labels)

from torch.utils.data import random_split
k = 10000
train_data, valid_data = random_split(train_data, [k, len(train_data)-k])

<class 'torch.utils.data.dataset.Subset'>

valid_target = valid_data.target

from torch.utils.data import Subset
class CustomSubset(Subset):
    '''A custom subset class'''
    def __init__(self, dataset, indices):
        super().__init__(dataset, indices)
        self.targets = dataset.targets # 保留targets属性
        self.classes = dataset.classes # 保留classes属性

    def __getitem__(self, idx): #同时支持索引访问操作
        x, y = self.dataset[self.indices[idx]]      
        return x, y 

    def __len__(self): # 同时支持取长度操作
        return len(self.indices)

import numpy as np
from copy import deepcopy
origin_data = deepcopy(train_data)
train_data = CustomSubset(origin_data, np.arange(k))
valid_data = CustomSubset(origin_data, np.arange(k, len(origin_data))-k)

print(valid_data.classes)
print(valid_data.targets)

['T-shirt/top', 'Trouser', 'Pullover', 'Dress', 'Coat', 'Sandal', 'Shirt', 'Sneaker', 'Bag', 'Ankle boot']
tensor([9, 0, 0,  ..., 3, 0, 5])

class CustomSubset(Subset):
    '''A custom subset class with customizable data transformation'''
    def __init__(self, dataset, indices, subset_transform=None):
        super().__init__(dataset, indices)
        self.targets = dataset.targets
        self.classes = dataset.classes
        self.subset_transform = subset_transform

    def __getitem__(self, idx):
        x, y = self.dataset[self.indices[idx]]
        
        if self.subset_transform:
            x = self.subset_transform(x)
      
        return x, y   
    
    def __len__(self): 
        return len(self.indices)

from torchvision import transforms
valid_data.subset_transform = transforms.Compose(\
    [transforms.RandomRotation((180,180))])

print(valid_data[0])