Production-ready data processing made easy and shareable
Explore the Fondant docs »

Dataset Card for fondant-ai/datacomp-small-clip

This is a dataset containing image urls and their CLIP embeddings, based on the datacomp_small dataset, and processed with fondant.

Dataset Details

Dataset Description

Large (image) datasets are often unwieldy to use due to their sheer size. Assume for instance that we would like to extract all the cat images from such a dataset. We would have to look at every image to classify if it's a cat image or not. And if we want to extract all the dog images next, we again need to look at every image.

Instead, we can look at every image once, and calculate a (CLIP) embedding representing its content. Combining these embeddings into an index, we can efficiently search through the dataset with a query, finding specific images, without having to look at each one.

This is what LAION did for their LAION-5b dataset, which made it possible to use, like we did in our ControlNet example. Unfortunately, the LAION-5b dataset and index have been taken offline (temporarily) and there aren't any alternatives. This is why we built an index for the Datacomp-12M dataset. While it is a lot smaller than LAION-5b, it should already enable a lot of use cases again, and can hopefully be the start towards building indices for more and larger datasets.

• License: cc-by-4.0

Dataset Sources

• Original data: datacomp_small
• Repository: fondant-clip-index

Uses

We provide an example use case which uses the FAISS index of this dataset to create a dataset of interior design images, used for the fine-tuning of a ControlNet model:

Dataset Structure

The data repository is structured as follows:

• data/: The dataset containing ids, urls, and CLIP embeddings
• faiss: The faiss index
• id_mapping/: The mapping of the faiss ids to the original urls

Dataset Creation

We leveraged Fondant to generate the CLIP index and published the pipeline as a git repository. The pipeline consists of 4 steps:

• A load_from_hf_hub operation that loads the datacomp_small dataset from huggingface into the Fondant workspace and format.
• A download_images operation which downloads the actual images from the urls in the dataset.
• A embed_images operation which embeds the downloaded images using a CLIP model.
• A write_to_file operation which writes the original urls and generated embeddings to the chosen destination.

After running the pipeline, we used autofaiss to build the CLIP index.

Execution details

Download images

We downloaded the images with 32 cores in parallel, each opening up to 25 concurrent connections, and achieved a success rate of 72%, resulting in 9.251.172 images.

The downloading was executed on a VM on GCP using the Fondant Docker runner. We originally planned to run this on Vertex AI, but moved to a VM when noticing lower network bandwidth on Vertex.

The success rate can probably be further improved by setting up a faster DNS resolver.

Embed images

We leveraged the laion/CLIP-ViT-B-32-laion2B-s34B-b79K CLIP model. We chose this model because of a couple of reasons. It is popular, which makes it easy to use with existing embeddings, it is small, which makes it cheap to run, and it is an open model trained on open data.

We appreciate any feedback on our choice of model, so we can take this into account if we generate indices for larger datasets in the future.

The embedding was executed on 4 T4 GPUs on Google Cloud using our Vertex AI runner, with a batch size of 32. The execution took 8:15 hours.

Terms and Conditions

Under no circumstances can Fondant be held liable by a third party for (i) the accuracy or correctness of the content, (ii) an alleged infringement of intellectual property rights or (iii) any other alleged claim, action, injunction or suit resulting from the publication or use of the dataset.

Dataset Card Contact

• Email: [email protected]
• Discord: https://discord.gg/HnTdWhydGp