LightlyDataset with filenames bug on subfolders

LightlyDataset with filenames bug on subfolders

For

dataset = LightlyDataset(
    input_dir=tmp_dir,
    filenames=filenames
)

I get an error when filenames doesn’t contain at least one image from every subfolder, e.g.

data/
L subfolder_1/
    L  img0.png
    L  img1.png
L subfolder_2/
    L img0.png

Then filenames=[subfolder_1/img0.png'] will fail.

[masc-it]

Hi @masc-it

Probably torchvision just assumes the directory ./test/.ipynb_checkpoints should contain images when that’s obviously not true.

You can see all the directories in ./test with the command

ls -a ./test

To fix your problem simply remove the .ipynb_checkpoints directory.

yep, have alread found out, forgot to edit the comment! Thanks anyways and keep up with the good work 😄

[MalteEbner]

I have discussed this issue in length with some developers from torchvision: https://github.com/pytorch/vision/issues/4925

The key takeaway is: Do not use LightlyDataset(input_dir) on input_dirs which contain subdirectories without images. This is needed to ensures that the labels returned when iterating over the LightlyDataset are continuous, start with 0 and that there is at least one image for every label.