python - 使用 PyTorch 和 TorchVision 对自定义数据集进行训练-验证-测试拆分

Question

我有一些用于二进制分类任务的图像数据，图像被组织到 2 个文件夹中，分别是 data/model_data/class-A 和 data/model_data/class-B。

总共有N张图片。我想对训练/验证/测试进行 70/20/10 拆分。
我正在使用 PyTorch 和 Torchvision 来完成这项任务。这是我到目前为止的代码。

from torch.utils.data import Dataset, DataLoader
from torchvision import transforms, utils, datasets, models

data_transform = transforms.Compose([
    transforms.RandomResizedCrop(224),
    transforms.RandomHorizontalFlip(),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])])

model_dataset = datasets.ImageFolder(root, transform=data_transform) 
train_count = int(0.7 * total_count) 
valid_count = int(0.2 * total_count)
test_count = total_count - train_count - valid_count
train_dataset, valid_dataset, test_dataset = torch.utils.data.random_split(model_dataset, (train_count, valid_count, test_count))
train_dataset_loader = torch.utils.data.DataLoader(train_dataset, batch_size=BATCH_SIZE, shuffle=True, num_workers=NUM_WORKER)  
valid_dataset_loader = torch.utils.data.DataLoader(valid_dataset, batch_size=BATCH_SIZE, shuffle=True, num_workers=NUM_WORKER) 
test_dataset_loader  = torch.utils.data.DataLoader(test_dataset , batch_size=BATCH_SIZE, shuffle=False,num_workers=NUM_WORKER)
dataloaders = {'train': train_dataset_loader, 'val': valid_dataset_loader, 'test': test_dataset_loader}

我觉得这不是正确的做法，原因有两个。

我对所有分割应用相同的变换。 (这显然不是我想要做的!解决方案很可能就是 here 。)

通常人们首先将原始数据分成测试/训练，然后他们
将 train 分成 train/val，而我直接将
原始数据进入训练/验证/测试。 (这是正确的吗？)

所以，我的问题是，我所做的是否正确？ (可能不是)
如果它不正确，我该如何编写数据加载器以实现所需的拆分，以便我可以对每个训练/测试/验证应用单独的转换？

Answer 1

Usually people first separate the original data into test/train and then they separate train into train/val, whereas I am directly separating the original data into train/val/test. (Is this correct?)

是的，它完全正确，可读且完全没问题

I am applying the same transform to all the splits. (This is not what I want to do, obviously! The solution for this is most probably the answer here.)

是的，这个答案是有可能的，但它是毫无意义的冗长 tbh。您可以使用第三方工具 torchdata ，只需安装:

pip install torchdata

可以在 here 找到文档(同样免责声明:我是作者)。

它允许您轻松地将转换映射到任何 torch.utils.data.Dataset (在本例中为 train )。您的代码看起来像这样(只需更改两行，检查注释，并格式化您的代码以使其更容易遵循):

import torch
import torchvision

import torchdata as td

data_transform = torchvision.transforms.Compose(
    [
        torchvision.transforms.RandomResizedCrop(224),
        torchvision.transforms.RandomHorizontalFlip(),
        torchvision.transforms.ToTensor(),
        torchvision.transforms.Normalize(
            mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
        ),
    ]
)

# Single change, makes an instance of torchdata.Dataset
# Works just like PyTorch's torch.utils.data.Dataset, but has
# additional capabilities like .map, cache etc., see project's description
model_dataset = td.datasets.WrapDataset(torchvision.datasets.ImageFolder(root))
# Also you shouldn't use transforms here but below
train_count = int(0.7 * total_count)
valid_count = int(0.2 * total_count)
test_count = total_count - train_count - valid_count
train_dataset, valid_dataset, test_dataset = torch.utils.data.random_split(
    model_dataset, (train_count, valid_count, test_count)
)

# Apply transformations here only for train dataset

train_dataset = train_dataset.map(data_transform)

# Rest of the code goes the same

train_dataset_loader = torch.utils.data.DataLoader(
    train_dataset, batch_size=BATCH_SIZE, shuffle=True, num_workers=NUM_WORKER
)
valid_dataset_loader = torch.utils.data.DataLoader(
    valid_dataset, batch_size=BATCH_SIZE, shuffle=True, num_workers=NUM_WORKER
)
test_dataset_loader = torch.utils.data.DataLoader(
    test_dataset, batch_size=BATCH_SIZE, shuffle=False, num_workers=NUM_WORKER
)
dataloaders = {
    "train": train_dataset_loader,
    "val": valid_dataset_loader,
    "test": test_dataset_loader,
}

是的，我同意在拆分之前指定 transform 不太清楚，而且 IMO 这更具可读性。