Siyuan Li*,1,2, Di Wu*,1,2, Fang Wu1,3, Zelin Zang1,2, Stan Z. Li†,1
1Westlake University, 2Zhejiang University, 3Tsinghua University
Masked image modeling (MIM), an emerging self-supervised pre-training method, has shown impressive success across numerous downstream vision tasks with Vision transformers (ViT). Its underlying idea is simple: a portion of the input image is randomly masked out and then reconstructed via the pre-text task. However, why MIM works well is not well explained, and previous studies insist that MIM primarily works for the Transformer family but is incompatible with CNNs. In this paper, we first study interactions among patches to understand what knowledge is learned and how it is acquired via the MIM task. We observe that MIM essentially teaches the model to learn better middle-level interactions among patches and extract more generalized features. Based on this fact, we propose an Architecture-Agnostic Masked Image Modeling framework (A2MIM), which is compatible with not only Transformers but also CNNs in a unified way. Extensive experiments on popular benchmarks show that our A2MIM learns better representations and endows the backbone model with the stronger capability to transfer to various downstream tasks for both Transformers and CNNs.
Table of Contents
We have released implementations of A2MIM based on OpenMixup. In the future, we plan to add A2MIM implementations to MMPretrain. Pre-trained and fine-tuned models are released in GitHub / Baidu Cloud.
- Update camera-ready version of A2MIM [arXiv] [poster]
- ImageNet pre-training and fine-tuning with OpenMixup [config_pretrain] [config_finetune]
- ImageNet pre-training and fine-tuning with MMPretrain
- Downstream Transfer to Object Detection on COCO with MMDetection [config]
- Downstream Transfer to Semantic Segmentation on ADE20K MMSegmentation [config]
- Analysis tools and results [rep_bottleneck]
- Visualization of pre-training on Google Colab and Notebook Demo
Please refer to INSTALL.md for installation instructions.
We provide scripts for multiple GPUs pre-training and the specified CONFIG_FILE.
bash tools/dist_train.sh ${CONFIG_FILE} ${GPUS} [optional arguments]For example, you can run the script below to pre-train ResNet-50 with A2MIM on ImageNet with 8 GPUs:
PORT=29500 bash tools/dist_train.sh configs/openmixup/pretrain/a2mim/imagenet/r50_l3_sz224_init_8xb256_cos_ep300.py 8After pre-trianing, you can fine-tune and evaluate the models with the corresponding script:
python tools/model_converters/extract_backbone_weights.py work_dirs/openmixup/pretrain/a2mim/imagenet/r50_l3_sz224_init_8xb256_cos_ep300/latest.pth ${PATH_TO_CHECKPOINT}
PORT=29500 bash tools/dist_train_ft_8gpu.sh configs/openmixup/finetune/imagenet/r50_rsb_a3_ft_sz160_4xb512_cos_fp16_ep100.py ${PATH_TO_CHECKPOINT}- A2MIM model: In a2mim.py, the A2MIM method takes input samples, applies masking, encodes and caculates the MIM losses.
- A2MIM head: In mim_head.py, two MIM losses are computed, where regression_loss.py caculates the L1 loss and focal_loss.py caculates for the Fourier domain loss.
- Dataloader: In masked_image.py, loading the processed RGB images with the masked RGB mean.
We provide the summarization of pre-training (800 or 300 epochs) and fine-tuning (100 or 300 epochs) results of A2MIM and baselines on ImageNet-1K.
| Methods | # Params. | Supervision | SimMIM | A2MIM |
|---|---|---|---|---|
| Target | (M) | Label | RGB | RGB |
| ViT-S | 48.8 | 79.9 | 81.7 | 82.1 |
| ViT-B | 86.7 | 81.8 | 83.8 | 84.2 |
| ViT-L | 304.6 | 82.6 | 85.6 | 86.1 |
| ResNet-50 | 25.6 | 79.8 | 79.9 | 80.4 |
| ResNet-101 | 44.5 | 81.3 | 81.3 | 81.9 |
| ResNet-152 | 60.2 | 81.8 | 81.9 | 82.5 |
| ResNet-200 | 64.7 | 82.1 | 82.2 | 83.0 |
| ConvNeXt-S | 50.2 | 83.1 | 83.2 | 83.7 |
| ConvNeXt-B | 88.6 | 83.5 | 83.6 | 84.1 |
Config files, models, logs, and visualization of reconstructions are provided as follows. These files can also be downloaded from a2mim-in1k-weights, OpenMixup-a2mim-in1k-weights or Baidu Cloud: A2MIM (3q5i).
ViT-S/B/L on ImageNet-1K.
| Method | Backbone | PT Epoch | FT Top-1 | Pre-training | Fine-tuning | Results |
|---|---|---|---|---|---|---|
| SimMIM | ViT-Small | 800 | 81.7 | config | ckpt | vis | config | ckpt | log |
| A2MIM | ViT-Small | 800 | 82.1 | config | ckpt | vis | config | ckpt | log |
| SimMIM | ViT-Base | 800 | 83.8 | config | ckpt | vis | config | ckpt | log |
| A2MIM | ViT-Base | 800 | 84.3 | config | ckpt | vis | config | ckpt | log |
| SimMIM | ViT-Large | 800 | 85.6 | config | ckpt | config | log |
| A2MIM | ViT-Large | 800 | 86.1 | config | ckpt | vis | config | log |
ResNet-50/101/152/200 on ImageNet-1K.
| Method | Backbone | PT Epoch | FT (A2) Top-1 | Pre-training | Fine-tuning | Results |
|---|---|---|---|---|---|---|
| SimMIM | ResNet-50 | 300 | 79.9 | config | ckpt | vis | RSB A2 | - |
| A2MIM | ResNet-50 | 100 | 78.8 | config | ckpt | vis | RSB A3 | ckpt | log |
| A2MIM | ResNet-50 | 300 | 80.4 | config | ckpt | vis | RSB A2 | ckpt | log |
| SimMIM | ResNet-101 | 300 | 81.3 | config | ckpt | RSB A2 | ckpt (A3) | log (A3) |
| A2MIM | ResNet-101 | 300 | 81.9 | config | ckpt (300ep) | ckpt (800ep) | RSB A2 | ckpt (A2) | log (A2) |
| SimMIM | ResNet-152 | 300 | 81.9 | config | ckpt | RSB A2 | log (A3) |
| A2MIM | ResNet-152 | 300 | 82.5 | config | ckpt (300ep) | ckpt (800ep) | RSB A2 | ckpt (A2) | log (A2) |
| SimMIM | ResNet-200 | 300 | 82.2 | config | ckpt | vis | RSB A2 | ckpt | log |
| A2MIM | ResNet-200 | 300 | 83.0 | config | ckpt | vis | RSB A2 | ckpt | log |
ConvNeXt-S/B on ImageNet-1K.
| Method | Backbone | PT Epoch | FT (A2) Top-1 | Pre-training | Fine-tuning | Results |
|---|---|---|---|---|---|---|
| SimMIM | ConvNeXt-S | 300 | 83.2 | config | ckpt | vis | RSB A2 | - |
| A2MIM | ConvNeXt-S | 300 | 83.7 | config | ckpt | vis | RSB A2 | ckpt | log |
| SimMIM | ConvNeXt-B | 300 | 83.6 | config | ckpt | RSB A2 | ckpt | log |
| A2MIM | ConvNeXt-B | 300 | 84.1 | config | ckpt | RSB A2 | ckpt (A2) | ckpt (A3) | log (A2) | log (A3) |
Following RepBottleneck, we provided interpretation of how masked image modeling works with representation bottleneck based on ViTs and CNNs. As shown in Figure 1/5/A1/A2 in A2MIM and following figures, we visualize the multi-order interation strengths with representation_bottleneck. Following How ViT works, we also provided analysis from frequency perspectives in Figure A3/A4 in A2MIM based on fourier_analysis.
This project is released under the Apache 2.0 license.
Our implementation is mainly based on the following codebases. We gratefully thank the authors for their wonderful works.
- OpenMixup: Open-source toolbox for supervised and self-supervised visual representation learning.
- pytorch-image-models: PyTorch image models, scripts, pretrained weights.
- SimMIM: Official PyTorch implementation of SimMIM.
- MMPretrain: OpenMMLab Pre-training Toolbox and Benchmark.
- MMDetection: OpenMMLab Detection Toolbox and Benchmark.
- MMSegmentation: OpenMMLab Semantic Segmentation Toolbox and Benchmark.
If you find this repository helpful, please consider citing our paper:
@inproceedings{icml2023a2mim,
title={Architecture-Agnostic Masked Image Modeling -- From ViT back to CNN},
author={Li, Siyuan and Wu, Di and Wu, Fang and Zang, Zelin and Li, Stan. Z.},
booktitle={International Conference on Machine Learning},
year={2023},
}