Unsupervised pre-training on custom data

[rram12]

❓ Unsupervised pre-training on custom data

From the documentation, Colabs, and the issues I read in this repo, I don't understand why we need to provide the dataset in a structured way with the labels (the same as the following).

data/ ├── train/ │ ├── class0/		└── *.png │ └── class1/	└── *.png

└── val/ ├── class0/ | └── .png └── class1/ └── .png

In my example, I have an unlabeled dataset of 1 million images and I can't really label them. My question is can we use it directly with all the images under one folder data (or just data/train and data/test)?

[QuentinDuval]

Hi @rram12,

First of all, thank you for using VISSL!

As you noticed, this folder structure is indeed a bit awkward but will work in your case nevertheless. You can create the following structure by considering that "no label" is equivalent to "one dummy label":

data/
   train/
      0/
         *.png
   val/
      0/
         *.png

You can actually omit the val folder if you don't have any, there should be no problem with that.

Please tell me if this unblocks you,

Thank you, Quentin

[rram12]

@QuentinDuval Thank you for your response. That's the way I was doing the training with "no label". I was just wondering why people seem to structure their dataset in the "supervised learning" way (e.g. #323) even though they are doing the "unsupervised learning". It is all clear now!