facebookresearch/deit

This repository contains practical implementations and trained models for several modern approaches to image recognition, the task of teaching computers to identify what's in a picture. Rather than requiring massive amounts of labeled training data like older methods did, these approaches are designed to work well even when you have limited data available. At a high level, the repository gives you working code and pre-trained models based on research papers published by Meta (formerly Facebook) researchers between 2021 and 2023. The core innovation across these projects is finding smarter ways to train image recognition systems so they need less data and compute time. Some approaches swap traditional convolutional networks (the older standard) for transformer-based architectures (originally developed for language), while others mix transformer concepts into convolutional designs. The README lists seven different research projects, each with its own folder and documentation, covering various experimental directions, like DeiT (the original data-efficient transformer approach), CaiT (making transformers deeper), ResMLP (using simpler feedforward networks), and others. The intended users are researchers, machine learning engineers, and practitioners who want to either build image classification systems or study how these newer architectures work. If you're training a model to classify product photos but only have a few thousand labeled examples (instead of millions), or if you want to understand how vision transformers compare to traditional convolutional networks, you'd use this repository. It provides the training scripts you need to fine-tune these models on your own data, as well as already-trained models you can download and use immediately. A practical aspect worth noting: because these are research implementations from a top institution, they're well-documented and actively maintained. The code comes with clear instructions in separate README files for each project, pre-trained model weights you can download, and the underlying academic papers so you can understand the theory behind each approach. This makes it useful both for practitioners who just want to use the models and for researchers who want to dive into the implementation details.