Setting transforms attribute of dataset not working after dataset has been initialized

vision
1

Hello, I want to set the transform attribute of a Dataset object once I have already instantiated the object - I try to do that using the setattr method of python, but it does not work - below is a minimum working example.

import torch
import numpy as np
from torch.utils.data import Dataset
import torchvision.transforms as transforms
from torch.utils.data import TensorDataset

trans = transforms.Compose([transforms.Resize(size=224), transforms.CenterCrop(size=(224, 224)), transforms.ToTensor(), transforms.Normalize(mean=(0.48145466, 0.4578275, 0.40821073), std=(0.26862954, 0.26130258, 0.27577711))])
images = torch.randn(50000, 3, 32, 32)
labels = torch.randint(9, (50000,1))
dset = TensorDataset(images, labels)
setattr(dset, 'transform', trans)
for idx, (x, y) in enumerate(dset):
    print(x.shape)
    break

The output of this snippet gives me -
torch.Size([3, 32, 32])

but rather, we should be getting
torch.Size([3, 224, 224])

why is setting attribute externally not working?

Please let me know if I am missing anything and thanks for your time!

2

TensorDataset doesn’t have and use a self.transform attribute and will just directly index the passed tensors as seen here.
So technically you are creating this attribute, which is just never used.

3

Thank you! But, isnt TensorDataset inherited from Dataset, and hence it must have the transform attribute?

4

It it inherited from Dataset but this base class also doesn’t define the transform attribute as seen in its definition.

5

Oh, thank you, that clears it!

6

With that being said, the easiest approach to apply a transformation on each tensor would be by creating a custom Dataset.

1 Like