在 K 折交叉验证中仅增加训练集 [英] Augmenting only the training set in K-folds cross validation

问题描述

我正在尝试为不平衡数据集(第 0 类 = 4000 张图像，第 1 类 = 大约 250 张图像)创建二进制 CNN 分类器，我想对其执行 5 折交叉验证.目前，我正在将训练集加载到 ImageLoader 中，该 ImageLoader 应用我的转换/增强(?)并将其加载到 DataLoader 中.然而，这导致我的训练分割和验证分割都包含增强数据.

I am trying to create a binary CNN classifier for an unbalanced dataset (class 0 = 4000 images, class 1 = around 250 images), which I want to perform 5-fold cross validation on. Currently I am loading my training set into an ImageLoader that applies my transformations/augmentations(?) and loads it into a DataLoader. However, this results in both my training splits and validation splits containing the augmented data.

我最初应用离线转换(离线增强?)来平衡我的数据集，但是从这个线程 (https://stats.stackexchange.com/questions/175504/how-to-do-data-augmentation-and-train-validate-split)，似乎最好只增加训练集.我也更愿意在单独的增强训练数据上训练我的模型，然后在非增强数据上进行 5 折交叉验证

I originally applied transformations offline (offline augmentation?) to balance my dataset, but from this thread (https://stats.stackexchange.com/questions/175504/how-to-do-data-augmentation-and-train-validate-split), it seems it would be ideal to only augment the training set. I would also prefer to train my model on solely augmented training data and then validate it on non-augmented data in a 5-fold cross validation

我的数据按根/标签/图像组织，其中有 2 个标签文件夹(0 和 1)，图像分类到各自的标签中.

My data is organized as root/label/images, where there are 2 label folders (0 and 1) and images sorted into the respective labels.

total_set = datasets.ImageFolder(ROOT, transform = data_transforms['my_transforms'])

//Eventually I plan to run cross-validation as such:
splits = KFold(cv = 5, shuffle = True, random_state = 42)

for train_idx, valid_idx in splits.split(total_set):
    train_sampler = SubsetRandomSampler(train_idx)
    valid_sampler = SubsetRandomSampler(valid_idx)

    train_loader = torch.utils.data.DataLoader(total_set, batch_size=32, sampler=train_sampler)
    val_loader = torch.utils.data.DataLoader(total_set, batch_size=32, sampler=valid_sampler)

model.train()
//Model train/eval works but may be overpredict 

我确定我在这段代码中做了一些次优或错误的事情，但我似乎找不到任何关于专门增加交叉验证中的训练分割的文档！

I'm sure I'm doing something sub-optimally or wrong in this code, but I can't seem to find any documentation on specifically augmenting only the training splits in cross-validation!

任何帮助将不胜感激！

推荐答案

一种方法是实现一个包装 Dataset 类，该类将转换应用于 ImageFolder 数据集的输出.例如

One approach is to implement a wrapper Dataset class that applies transforms to the output of your ImageFolder dataset. For example

class WrapperDataset:
    def __init__(self, dataset, transform=None, target_transform=None):
        self.dataset = dataset
        self.transform = transform
        self.target_transform = target_transform

    def __getitem__(self, index):
        image, label = self.dataset[index]
        if self.transform is not None:
            image = self.transform(image)
        if self.target_transform is not None:
            label = self.target_transform(label)
        return image, label

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

然后你可以在你的代码中使用它，方法是用不同的转换包装更大的数据集.

Then you could use this in your code by wrapping the larger dataset with different transforms.

total_set = datasets.ImageFolder(ROOT)

# Eventually I plan to run cross-validation as such:
splits = KFold(cv = 5, shuffle = True, random_state = 42)

for train_idx, valid_idx in splits.split(total_set):
    train_sampler = SubsetRandomSampler(train_idx)
    valid_sampler = SubsetRandomSampler(valid_idx)

    train_loader = torch.utils.data.DataLoader(
        WrapperDataset(total_set, transform=data_transforms['train_transforms']),
        batch_size=32, sampler=train_sampler)
    valid_loader = torch.utils.data.DataLoader(
        WrapperDataset(total_set, transform=data_transforms['valid_transforms']),
        batch_size=32, sampler=valid_sampler)

    # train/validate now

我没有测试这段代码，因为我没有你的完整代码/模型，但概念应该很清楚.

I haven't tested this code since I don't have your full code/models but the concept should be clear.

这篇关于在 K 折交叉验证中仅增加训练集的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！