[Documentation request] How to create a dataset for training

[StoneCypher]

You do say how to train a new model from a dataset. You don't say how to create a dataset.

Is it just a directory of images? Do they need to be sized somehow? Is there a subdirectory tree? Are there text files containing metadata?

[patrickvonplaten]

Hey @StoneCypher,

That's a very good question and quite important indeed! I'll do my best to answer it here, but @anton-l might have a better answer.

So in short the google colab shows you how to train a simple unconditional diffusion model on any image datasets. All you need are actually images and that's it!

The training script is written in a way that expects you to use the Datasets library for your vision data. More specifically, the script expects that one key of your dataset is called image as can be seen here:

diffusers/examples/train_unconditional.py

Line 81 in 1a84bd2

images = [augmentations(image.convert("RGB")) for image in examples["image"]]

(note how the image key is called). This means your dataset should have an image key.

Now there are two options:

• Either you can use a pre-existing dataset that has images, e.g.:
  • https://huggingface.co/datasets/cifar10 (here you could just rename img to image to make it work. There are many other image datasets on the Hub or
  • You can create your own image dataset as is explained here: https://huggingface.co/docs/datasets/image_process#image-datasets

[patrickvonplaten]

@anton-l we should also try to make the training colab / script a bit easier to adapt to other datasets (with clear comments on how to do so)

[patrickvonplaten]

Let's maybe take the Transformers notebook as an example here: https://github.com/huggingface/transformers/blob/main/examples/pytorch/text-classification/run_glue_no_trainer.py#L53

[patrickvonplaten]

@StoneCypher Let me know if you have more in-detail questions! Happy to help

[StoneCypher]

@patrickvonplaten - Thank you.

PR to obviate.