Different Transformations for Different Classes

vision
1

I would like to perform different transformations on 5 different classes in a multi-class classification problem. What’s the cleanest way to do this using ImageFolder and DataLoader?

Note: The classes also have quite different sizes, ranging from 50-3,000 images
How can I set this up to not over fit on the classes with more data?

My data structure looks like this:

/data
    /train
        /class1
        /class2
        /class3
        /class4
        /class5
    /test
        /class1
        /class2
        /class3
        /class4
        /class5
2

The easiest way that comes to my mind at the moment is to create a new Dataset instance and explicitly use different transformations on different classes.
For the unbalanced dataset, you may want to use WeightedRandomSampler

3

Currently my setup is something like this:

class ConcatDataset(Dataset):
    '''Concatenates Datasets'''
    
    def __init__(self, *datasets):
        self.datasets = datasets

    def __getitem__(self, i):
        return tuple(d[i] for d in self.datasets)

    def __len__(self):
        return min(len(d) for d in self.datasets)

train_1_transforms = ...
train_2_transforms = ...
train_3_transforms = ...
train_4_transforms = ...
train_5_transforms = ...

train_1_data = ImageFolder(train_1_dir, transform=train_1_transforms)
train_2_data = ImageFolder(train_2_dir, transform=train_2_transforms)
train_3_data = ImageFolder(train_3_dir, transform=train_3_transforms)
train_4_data = ImageFolder(train_4_dir, transform=train_4_transforms)
train_5_data = ImageFolder(train_5_dir, transform=train_5_transforms)

train_concat_data = ConcatDataset(train_1_data, train_2_data, train_3_data, train_4_data, train_5_data)

train_loader = DataLoader(train_concat_data, batch_size=64, shuffle=True, pin_memory=True, num_workers=4)

I’m getting error:

RuntimeError: Found 0 files in subfolders of: ./data/train/class_1

I don’t have any idea how else to apply separate transformations to each class.

4

Dont use ImageFolder, just read the images using os.walk for example and store their paths and labels (based on which class they reside in you can give them a number) and then in __getitem__, after you have read the image, based on the label that you also have, you can apply the needed transformation and then return the pair.