Batch of images, change dimension

#1

Greetings!

I am trying to preprocess RGB images to create batches of the correct shape (as required by my network).
Example: CIFAR10

transfos = transforms.Compose([torchvision.transforms.Grayscale(),transforms.ToTensor(),transforms.Normalize((0.4749), (0.2382))])
train = datasets.CIFAR10(".", train = True, download = True, transform = transfos)
test = datasets.CIFAR10(".", train = False, download = True, transform = transfos)

batches_train = torch.utils.data.DataLoader(train, batch_size=bs, shuffle=True)
batches_test = torch.utils.data.DataLoader(test, batch_size=bs, shuffle=True)

The problem now is that each batch has dimension (batchsize, 1, H, W).
For all batches, I would like to have the dimension (batchsize, H*W). Is there a simple way to use transforms to do this?

Thank you!

PiF

#2

Does something like:

>>> import torch
>>> data = torch.randn(32, 1, 32, 32)
>>> data.shape
torch.Size([32, 1, 32, 32])
>>> data = data.reshape(32, -1)
>>> data.shape
torch.Size([32, 1024])

work?

#3

That’s how I do it right now, but it strikes me as very inefficient as I have to do this each time my Dataloader hands me a new batch. I was hoping there is a preprocessing solution so that my Dataloader hands me the batches in the correct shape (no more reshaping during training).

#4

You can also do the reshape in your dataloader if you think that is a significant source of inefficiency. However, this is unlikely because a reshape doesn’t actually change the data layout or require any copying of data. It just changes the strides for indexing.

#5

Alright, then it might not be computationally inefficient, but it would help readability if I could tell the dataloader from the get go what to do. How do I do it in the Dataloader? The Dataloader is not a simple tensor.

#6

You can add another transform function to your list of your transforms to do the reshape e.g.,

def reshape_transform():
    def fn(data):
        batch_size = data.size(0)
        return data.reshape(batch_size, -1)
    return fn

transfos = transforms.Compose([torchvision.transforms.Grayscale(),transforms.ToTensor(),transforms.Normalize((0.4749), (0.2382)), reshape_transform()])
1 Like
#7

This should do it!
I will try it asap. Thank you!