DataLoader transformations not working

1

I have this code where I tested Normalize and LinearTranformation.
torchvision.transforms.Normalize and torchvision.transforms.LinearTransformation to be more precise.

import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import DataLoader, Dataset, TensorDataset
from torch.optim import *
import torchvision

trans = torch.tensor([[1/128., 1.], [1., 1.]])
dl = DataLoader(
  torchvision.datasets.MNIST('/data/mnist', train=True, download=True,
                             transform=torchvision.transforms.Compose([
                               torchvision.transforms.ToTensor(),
                               torchvision.transforms.Normalize((0.5,), (0.5,)),
                               torchvision.transforms.LinearTransformation(trans)  
                             ])), shuffle=False)



tensor = dl.dataset.data
tensor = tensor.to(dtype=torch.float32)
tr = tensor.reshape(tensor.size(0), -1) 
# tr = tr/128 # tr = tr/255
targets = dl.dataset.targets
targets = targets.to(dtype=torch.long)

x_train = tr[0:50000-1]
print(x_train[0])

The output print(x_train[0]) is always the same, even if I eliminate these two transformations, or if I eliminate any of these two. Do you have any idea why is this so?

By LinearTransformation I tried to do a transform that would be identical to scaling down the tensor values like tr = tr/128 for instance in the commented part of the code.

2

You are accessing the dataset directly. You need to access it via the DataLoader object, for the transforms to be applied.

For example, you could do:

next(iter(dl))

This should return a Tensor to which transforms have been applied.

More generally, to iterate through your dataset, do:

for i, sample in enumerate(dl):
    print(sample) # Or any other processing you'd like to perform

Aside

Your usage of LinearTransformation() appears to be inconsistent with the latest API. LinearTransformation() expects a transformation_matrix and a mean_vector. It seems like you’re providing only the transformation_matrix.

1 Like
3

I tried to find where I could match the example from PyTorch website and I found an instance having Compose:

image
image972Ă—493 40.6 KB

I am using the DataLoader already, where exactly the first parameter is a Dataset.
I provided that first parameter because torchvision.datasets.MNIST(... is a Dataset.

4

Yes, I tried to do that but if I recall, when I tried to set the mean, I got this error.

TypeError: init() takes 2 positional arguments but 3 were given
Seams that I cannot provide second parameter at all.

PyTorch ver 1.0.0

5

Does using the transform in here:

dl = DataLoader(
  torchvision.datasets.MNIST('/data/mnist', train=True, download=True,
                             transform=torchvision.transforms.Compose([
                               torchvision.transforms.ToTensor(),
                               torchvision.transforms.Normalize((0.5,), (0.5,)),
                               torchvision.transforms.LinearTransformation(trans)  
                             ])), shuffle=False)

really has any sense?
Try the example and you will see nothing will be transformed. Am I missing something obvious?

6

Your dataloader transforms the data. What’s incorrect is that you’re accessing the train set directly, and not the output of the dataloader.

More concretely, you must not access tr[0] directly. You should rather access the dataloader output. (accessing the zeroth index is as simple as calling next(iter(dl)))

1 Like