使用PyTorch的DataLoaders进行k折交叉验证

Question

我已将我的训练数据集分成了80%的训练数据和20%的验证数据，并创建了下面所示的DataLoader。然而，我不想限制模型的训练。因此，我考虑将我的数据分成K（可能为5）个折叠，并执行交叉验证。但我不知道如何在将它们拆分后将数据集合并到我的dataloader中。

train_size = int(0.8 * len(full_dataset))
validation_size = len(full_dataset) - train_size
train_dataset, validation_dataset = random_split(full_dataset, [train_size, validation_size])

full_loader = DataLoader(full_dataset, batch_size=4,sampler = sampler_(full_dataset), pin_memory=True) 
train_loader = DataLoader(train_dataset, batch_size=4, sampler = sampler_(train_dataset))
val_loader = DataLoader(validation_dataset, batch_size=1, sampler = sampler_(validation_dataset))

- Suraj

---

请查看 torch.utils.data.ConcatDataset。 - Berriel

Answer 1

我刚刚写了一个可以与数据加载器和数据集一起使用的交叉验证函数。这是我的代码，希望对你有所帮助。

# define a cross validation function
def crossvalid(model=None,criterion=None,optimizer=None,dataset=None,k_fold=5):
    
    train_score = pd.Series()
    val_score = pd.Series()
    
    total_size = len(dataset)
    fraction = 1/k_fold
    seg = int(total_size * fraction)
    # tr:train,val:valid; r:right,l:left;  eg: trrr: right index of right side train subset 
    # index: [trll,trlr],[vall,valr],[trrl,trrr]
    for i in range(k_fold):
        trll = 0
        trlr = i * seg
        vall = trlr
        valr = i * seg + seg
        trrl = valr
        trrr = total_size
        # msg
#         print("train indices: [%d,%d),[%d,%d), test indices: [%d,%d)" 
#               % (trll,trlr,trrl,trrr,vall,valr))
        
        train_left_indices = list(range(trll,trlr))
        train_right_indices = list(range(trrl,trrr))
        
        train_indices = train_left_indices + train_right_indices
        val_indices = list(range(vall,valr))
        
        train_set = torch.utils.data.dataset.Subset(dataset,train_indices)
        val_set = torch.utils.data.dataset.Subset(dataset,val_indices)
        
#         print(len(train_set),len(val_set))
#         print()
        
        train_loader = torch.utils.data.DataLoader(train_set, batch_size=50,
                                          shuffle=True, num_workers=4)
        val_loader = torch.utils.data.DataLoader(val_set, batch_size=50,
                                          shuffle=True, num_workers=4)
        train_acc = train(res_model,criterion,optimizer,train_loader,epoch=1)
        train_score.at[i] = train_acc
        val_acc = valid(res_model,criterion,optimizer,val_loader)
        val_score.at[i] = val_acc
    
    return train_score,val_score
        

train_score,val_score = crossvalid(res_model,criterion,optimizer,dataset=tiny_dataset)

为了让您对我们的做法有一个正确的直观感受，请看下面的输出结果：

train indices: [0,0),[3600,18000), test indices: [0,3600)
14400 3600

train indices: [0,3600),[7200,18000), test indices: [3600,7200)
14400 3600

train indices: [0,7200),[10800,18000), test indices: [7200,10800)
14400 3600

train indices: [0,10800),[14400,18000), test indices: [10800,14400)
14400 3600

train indices: [0,14400),[18000,18000), test indices: [14400,18000)
14400 3600

- Skipper

---

非常好的例子，谢谢您提供这个。我认为将数据集拆分和训练分开会更好。例如： metrics = k_fold(full_dataset, train_fn, **other_options)，其中 k_fold 函数将负责数据集的拆分，并将 train_loader 和 val_loader 传递给 train_fn，并将其输出收集到指标中。train_fn 将负责实际的训练，并返回每个 K 的指标。 - 18augst

什么是准则和优化器？优化器是否类似于AdamW？ - GradStudent

Answer 2

请看使用pytorch和sklearn进行MNIST数据集的交叉验证。问题提问者实现了k折交叉验证。特别注意他自己的答案（于2019年11月23日10:34回答）。他不依赖于random_split()，而是使用sklearn.model_selection.KFold构建数据集，并从那里构造一个Dataloader。

- kradant

Answer 3

你可以使用sklearn中的KFOLD和dataloader来实现这一点。

import torch
from torch._six import int_classes as _int_classes
from torch import Tensor

from typing import Iterator, Optional, Sequence, List, TypeVar, Generic, Sized

T_co = TypeVar('T_co', covariant=True)

class Sampler(Generic[T_co]):
    r"""Base class for all Samplers.

    Every Sampler subclass has to provide an :meth:`__iter__` method, providing a
    way to iterate over indices of dataset elements, and a :meth:`__len__` method
    that returns the length of the returned iterators.

    .. note:: The :meth:`__len__` method isn't strictly required by
              :class:`~torch.utils.data.DataLoader`, but is expected in any
              calculation involving the length of a :class:`~torch.utils.data.DataLoader`.
    """

    def __init__(self, data_source: Optional[Sized]) -> None:
        pass

    def __iter__(self) -> Iterator[T_co]:
        raise NotImplementedError
        
class SubsetRandomSampler(Sampler[int]):
    r"""Samples elements randomly from a given list of indices, without replacement.

    Args:
        indices (sequence): a sequence of indices
        generator (Generator): Generator used in sampling.
    """
    indices: Sequence[int]

    def __init__(self, indices: Sequence[int], generator=None) -> None:
        self.indices = indices
        self.generator = generator

    def __iter__(self):
        return (self.indices[i] for i in torch.randperm(len(self.indices), generator=self.generator))

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


train_dataset = CustomDataset(data_dir=train_path, mode='train') )
val_dataset = CustomDataset(data_dir=train_path, mode='val') )

    fold = KFold(5, shuffle=True, random_state=random_seed)
    for fold,(tr_idx, val_idx) in enumerate(fold.split(dataset)):
        # initialize the model
        model = smp.FPN(encoder_name='efficientnet-b4', classes=12 , encoder_weights=None, activation='softmax2d')
    
 
     
        loss = BCEDiceLoss()
        optimizer = torch.optim.AdamW([
            {'params': model.decoder.parameters(), 'lr': 1e-07/2}, 
            {'params': model.encoder.parameters(), 'lr': 5e-07},  
        ])
        scheduler = ReduceLROnPlateau(optimizer, factor=0.15, patience=2)
    
  
    
        print('#'*35); print('############ FOLD ',fold+1,' #############'); print('#'*35);
        train_loader = torch.utils.data.DataLoader(dataset=train_dataset, 
                                               batch_size=batch_size,
                                               num_workers=1,
                                               sampler = SubsetRandomSampler(tr_idx)
                                            )
        val_loader = torch.utils.data.DataLoader(dataset=val_dataset, 
                                               batch_size=batch_size,
                                               num_workers=1,
                                               sampler = SubsetRandomSampler(val_idx)
                                            )

因此，在编写DataLoader部分时，请使用subsetRandomSampler。这样，dataloader中的采样器将始终随机采样由kfold函数生成的训练/验证索引。

- sfsf9797