Unsupervised Images Dataset?

[devforfu]

How do you think, would it be reasonable to have an "unsupervised" mode for DatasetFolder class? Now it expects each image to sit in its own folder but sometimes I would like to just load a bunch of images from the unlabelled test set, like:

dataset = DatasetFolder(path, loader=open_image, unsupervised=True)
image = dataset[i]  # or maybe (image, None), or (image, -1)

[fmassa]

Hi,

I believe we can always create a folder with a single folder in it (for example via symlinks if we don't want to copy the data), which makes it possible to use DatasetFolder straight away.

Would that be a solution? It would return (image, 0), which you can discard

[devforfu]

@fmassa Yes, agree, that sounds like a valid solution.

I guess the only drawback is that you need to remember about this additional step in data preprocessing pipeline each time when you start working on a new dataset :)

[fmassa]

Or a very basic implementation:

import os
import glob

class BunchOfImagesDataset(object):
    def __init__(self, folder, transforms=None):
        self.folder = folder
        self.files = glob.glob(os.path.join(folder, '*.jpg'))
    
    def __getitem__(self, idx):
        ...
    def __len__(self):
        return len(self.files)

I think that the proposed solutions are simple enough, so that it might not be necessary to add one more level of complexity.