How to transform data in a flexible way?

vision
1

Currently, it seems only one way to transforms the data, in ImageFolder.
But it is so inconvenient for different purposes.
For example, I want to import the data first, and then transform it, so I want to perform data transforms on tensor.
Could it be realized in the next release? I wish a flexible method to transform data.

2

Iā€™m not sure I understand what you ask. Transforms can be use flexibly, like so:

    from torchvision.transforms import transforms
    from yourdataset import yourdatasetclass

    tf = transforms.ToTensor()
    dset = yourdatasetclass(root, transforms=tf)
    dloader = torch.utils.data.Dataloader(dset, batch_size=2,shuffle=True,num_workers=2)

You can use transforms.Compose to chain your transforms sequentially.

3

What is the yourdatasetclass?
Can I define it by myself?
It seems like importing dataset from root. Is it ImageFolder?
I want to import dataset first, and then transform it in the tensor.

4

you can define ImageFolder by yourself, it is very easy. This is the source code, you only need to define init, get_item and len, it is subclass of data.Dataset

1 Like
5

Thanks for your kind answer~
It is indeed an effective solution.
Besides, I conceive that a more flexible transform operation should be added to the new release for convenience.

6

I see!
I can define mydatasetclass reference to dset.CIFAR10(root, train=True, transform=None, target_transform=None, download=False).
Thank you very much!

7

Datasets are really simple.
Transforms are also very flexible.

See http://pytorch.org/tutorials/beginner/data_loading_tutorial.html for tutorial on both datasets and transforms.