Why do we need the training labels/annotations if this is a self-supervised training?

[chandlerbing65nm]

I'm confused on why we need to add the training labels/annotations for training.

[fangyixiao18]

I'm confused on why we need to add the training labels/annotations for training.

Can you provide the code scripts? Self-supervised learning doesn't need ground-truth labels, but some methods may use some pseudo labels.

[chandlerbing65nm]

I'm confused on why we need to add the training labels/annotations for training.

Can you provide the code scripts? Self-supervised learning doesn't need ground-truth labels, but some methods may use some pseudo labels.

@fangyixiao18 this one:

mmselfsup/configs/selfsup/_base_/datasets/imagenet_simclr.py

Line 39 in 684875d

ann_file='data/imagenet/meta/train.txt',

Why do I need the training labels for imagenet?

[fangyixiao18]

we use ann_file to load image according to the image path in the file, but we don't use labels during training.

[fangyixiao18]

https://github.com/open-mmlab/mmselfsup/blob/1.x/docs/en/user_guides/4_pretrain_custom_dataset.md
you can train the model on your dataset, following this document on 1.x branch

[chandlerbing65nm]

oh thanks @fangyixiao18. So, if I use my own dataset and format it similar to imagenet, I just need the names of the images listed in a txt file?

[fangyixiao18]

Yes, you could try it