#ImageNet-1K

This repository features FastViT, a rapid hybrid vision transformer utilizing structural reparameterization to boost image classification accuracy. Models have been trained on ImageNet-1K and benchmarked for latency on an iPhone 12 Pro via the ModelBench app. The repository includes setup guides for configuring environments, training, and evaluating models, along with scripts for implementation. It provides a varied collection of pre-trained models tailored for classification tasks, including the option for knowledge distillation. Comprehensive dataset preparation and model export instructions are available, making this a versatile tool for tasks ranging from training to fine-tuning in machine learning.

RepViT-SAM addresses computational challenges in mobile vision tasks by replacing conventional image encoders with advanced RepViT models, enhancing segmentation speed and efficiency on devices such as iPhones. With this approach, RepViT-SAM achieves impressive zero-shot transfer performance and up to ten times faster inference. Leveraging state-of-the-art ViT and CNN integrations, the RepViT family sets a new standard in lightweight model performance, boasting over 80% top-1 accuracy on ImageNet while maintaining low latency.

This objective overview details several MetaFormer architectural baselines implemented in PyTorch that perform with high accuracy on ImageNet-1K. Models such as IdentityFormer, RandFormer, ConvFormer, and CAFormer utilize unique token mixers like identity mapping and global random mixing. CAFormer achieves 85.5% accuracy at 224x224 resolution in a standard training setup. Learn how these architectures assist in optimizing vision tasks, with integration in the timm library for improved machine vision capabilities.