How to rearrange the sample order of a torch dataloader?

Question

I have a "torch.utils.data.DataLoader". I want to rearrange the order of the samples. Is it possible?

By rearrange, do you mean shuffle at every epoch? You can set `shuffle=True` to achieve this. — Jake Tae, Mar 28 '22 at 03:12
But I don't want to randomly shuffle them. I want to re order them in a particular way. — odbhut.shei.chhele, Mar 28 '22 at 04:34
@JakeTae But I don't want to randomly shuffle them. I want to re order them in a particular way. — odbhut.shei.chhele, Mar 28 '22 at 04:34
Then you have to reorder the `Dataset` which is used to initialize the `Dataloader`. Your custom `Dataset` class should have an `__getitem__` method, which determines the order of how the samples are retrieved. — Jake Tae, Mar 28 '22 at 05:34
@JakeTae Yes it is there. Is there any code example that I can see? Thank you very much. — odbhut.shei.chhele, Mar 28 '22 at 15:13

score 1 · Accepted Answer · answered Apr 01 '22 at 12:57

Yes, you can use torch.utils.data.Subset and specify the indices.

import numpy as np
import torch
from torch.utils.data import DataLoader, Subset, TensorDataset


data = np.arange(5) ** 2
dataset = TensorDataset(torch.tensor(data))

# Subset with entire Dataset in rearranged order
dataset_ordered = Subset(dataset, indices=[2, 1, 3, 4, 0])

for x in DataLoader(dataset_ordered):
    print(x)

# [tensor([4])]
# [tensor([1])]
# [tensor([9])]
# [tensor([16])]
# [tensor([0])]

How to rearrange the sample order of a torch dataloader?

1 Answers1

Linked