ConvMAE

Why we included this project

ConvMAE is the official implementation of a NeurIPS 2022 masked autoencoder that combines convolutional stages with a transformer encoder. Pretraining reconstructs masked image patches on a hybrid, multi-scale architecture, and the hierarchical features it learns transfer well to downstream tasks. The repo includes pretraining and finetuning scripts for ImageNet-1K, plus configs for COCO object detection and instance segmentation with Mask R-CNN and ADE20K semantic segmentation with UperNet, so you can reproduce the reported numbers or reuse the released checkpoints in your own work. It is a research codebase, not a turnkey service: expect to handle configs, training logs, and GPU time yourself. If you are comparing self-supervised vision pretraining methods, having the full train-and-evaluate flow in one repo is a real convenience.