Image transforms

vision
1

Is there a need to resize images using (transforms.RandomResizedCrop(224)) in validation and testing datasets when using transfer learning with VGG16? or only using normalization and nothing else?

2

Hi,

If your input images are bigger than [224, 224] just normalization would be enough, otherwise you need to resize all image to the aforementioned size.

Here is the official source code of evaluation step:

normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
                                     std=[0.229, 0.224, 0.225])

transforms.Compose([
            transforms.Scale(256),  # optional
            transforms.CenterCrop(224),  # optional
            transforms.ToTensor(),
            normalize,
        ])

Source code of training using ImageNet dataset

Bests

1 Like
3

Many thanks!
Could you tell me how can I check the images sizes in a certain dataset?

4

If I understand correctly, you can find info about image sizes in the webpage of each dataset. If you are using your own custom dataset, you can check it in __getitem__ method by using CustomDataset. Here is post for such a situation: Custom Dataset with some preprocessing

Also, dataloading tutorial may help.

But a simple solution would be adding transforms.Resize(224) in the previous post right before ToTensor().

1 Like
5

Thank you, that’s what I’m looking for :smiley: