How to make multiple crops for images

1

Is there any implementation of multiple crops in pytorch? I’ve look up in pytorch document, but there are only

class torchvision.transforms.CenterCrop(size)

class torchvision.transforms.RandomCrop(size, padding=0)

class torchvision.transforms.RandomSizedCrop(size, interpolation=2)

I think all these transforms will return one crop of an image. So how to get multiple sub-crops such as 10 sub-crops(4 coners + 1 center with(out) flips).

Thanks!

2

you need to write your own transform.

3

OK, thanks for your reply

4

Has anyone implemented multi-crop testing transforms efficiently?

5

I have coded a custom transform class to generate N random crops with padding size and number of random crops:


and here’s a test script:

1 Like
6

You can place the data loading step inside a loop with number of loops equal number of crops required per image/

data_transforms = transforms.Compose([
        transforms.RandomCrop(200),
        transforms.RandomHorizontalFlip(),
        transforms.ToTensor()
    ])
image_dataset_hazy = torchvision.datasets.ImageFolder('/home/Desktop/Data3',transform= data_transforms)

Place the above code (and after it) inside loop.