balanced-laion5b

This repository shall help finding a good distribution for huge datasets like LAION-5B for more efficient training.

Project outline

• Literature research
• Evaluate possible balancing strategies
• Find a good sample / test dataset, e.g. a subset of LAION-5B
• Create adjective + noun dictionary
• Run model/algorithm on LAION-400M
• Run model/algorithm on LAION-5B

Literature

• https://arxiv.org/abs/2106.09994
• https://onlinelibrary.wiley.com/doi/abs/10.1002/cpe.603
• https://www.frontiersin.org/articles/10.3389/fpubh.2020.00274/full
• ...

Methods

Idea 1

1. Cluster text/image embeddings (e.g., KMM/UMAP/PaCMAP/...) and reduce dense regions
2. Compare distribution of nouns/adjectives in whole dataset and subset

Idea 2

1. ...
2. ...