Data Augmentation Tools for image classification

1

Hello, everyone.
I’d like to know tools or packages except default torch augmentation tools, such as albumentations, …

2

you could augment the data using the transforms pipeline available in Torchvision. You create a transorm(a set of operations an image would go through in the pipeline ) as follows -

transforms = transforms.Compose([
                              transforms.Resize((350, 350)),
                              transforms.RandomHorizontalFlip(),
                              transforms.RandomRotation(degrees=20),
                              transforms.RandomAffine(degrees=20),
                              transforms.ColorJitter(hue=0.5, saturation=.01),
                              # Convert the image to tensor,
                              transforms.ToTensor()
                                ])

to get a Dataloader containing both modified and the original images (hence augmenting the dataset) do the following -

def getConcatedDataset(root_dir):

    image_transformations = transforms.Compose([
        transforms.Resize((600, 600)),
        transforms.RandomHorizontalFlip(),
        transforms.RandomRotation(degrees=20),
        transforms.RandomAffine(degrees=20),
        transforms.ColorJitter(hue=0.5, saturation=.01),
        # Convert the image to tensor,
        transforms.ToTensor(),
        #transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
    ])

    minimum_transformations = transforms.Compose([
        transforms.ToTensor(),
        #transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
    ])

    dataset1 = ImageFolder(root_dir, transform=minimum_transformations)
    dataset2 = ImageFolder(root_dir, image_transformations)
    return ConcatDataset([dataset1, dataset2])

This way, you can create multiple transformation pipelines and append them into one big Dataloader.

3

thanks for your help.
Do you know external tools like Imgaug, albumentations?

4

Nope, Do not know any of them…