ColorMAE: Exploring data-independent masking strategies in Masked AutoEncoders

Image and Video Understanding Lab, AI Initiative, KAUST

Carlos Hinojosa, Shuming Liu, Bernard Ghanem

Paper · GitHub · Supplementary Material · Project · BibTeX

Can we enhance MAE performance beyond random masking without relying on input data or incurring additional computational costs?

We introduce ColorMAE, a simple yet effective data-independent method which generates different binary mask patterns by filtering random noise. Drawing inspiration from color noise in image processing, we explore four types of filters to yield mask patterns with different spatial and semantic priors. ColorMAE requires no additional learnable parameters or computational overhead in the network, yet it significantly enhances the learned representations.

Installation

To get started with ColorMAE, follow these steps to set up the required environment and dependencies. This guide will walk you through creating a Conda environment, installing necessary packages, and setting up the project for use.

1. Clone our repo to your local machine

git clone https://github.com/carlosh93/ColorMAE.git
cd ColorMAE

2. Create conda environment with python 3.10.12

conda create --prefix ./venv python=3.10.12 -y
conda activate ./venv

3. Install Pytorch 2.0.1 and mmpretrain 1.0.2:

pip install torch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 --index-url https://download.pytorch.org/whl/cu118

pip install -U openmim && mim install mmpretrain==1.0.2 mmengine==0.8.4 mmcv==2.0.1

pip install yapf==0.40.1

Note: You can install mmpretrain as a Python package (using the above commands) or from source (see here).

Getting Started

Setup Environment

At first, add the current folder to PYTHONPATH, so that Python can find your code. Run command in the current directory to add it.

Note: Please run it every time after you opened a new shell.

export PYTHONPATH=`pwd`:$PYTHONPATH

Data Preparation

Prepare the ImageNet-2012 dataset according to the instruction. We provide a script and step by step guide here.

Download Color Noise Patterns

The following table provides the color noise patterns used in the paper

Color Noise	Description	Link	Md5
Green Noise	Mid-frequency component of noise.	Download	a76e71
Blue Noise	High-frequency component of noise.	Download	ca6445
Purple Noise	Noise with only high and low-frequency content.	Download	590c8f
Red Noise	Low-frequency component of noise.	Download	1dbcaa

You can download these pre-generated color noise patterns and place them in the corresponding folder inside noise_colors directory of the project.

Models and results

In the following tables we provide the pretrained and finetuned models with their corresponding results presented in the paper.

Note: This Hugging Face repository is the canonical checkpoint archive. The OSF links remain available as legacy mirrors. Safetensors are the recommended, security-first downloads.

ColorMAE-Green is the recommended model family. Blue, Purple, and Red are provided as the color-mask ablations reported in Table 1 of the paper. Random-mask checkpoints are intentionally not duplicated here because they are baseline experiments rather than ColorMAE variants.

Hugging Face release index

Pretraining	Pretrained	ImageNet-1K	ADE20K	COCO 768	COCO 1024
100 epochs	checkpoint	81.82 top-1	42.24 mIoU	45.9 box AP / 40.9 mask AP	—
300 epochs	checkpoint	83.01 top-1	45.90 mIoU	48.7 box AP / 43.3 mask AP	50.4 box AP / 44.9 mask AP
800 epochs	checkpoint	83.61 top-1	49.18 mIoU	49.5 box AP / 43.7 mask AP	—
1600 epochs	checkpoint	83.77 top-1	49.26 mIoU	50.1 box AP / 44.3 mask AP	51.5 box AP / 45.7 mask AP

Safe and compatible formats

• Recommended: .safetensors contains only tensors and cannot execute pickle imports.
• Native MMEngine resume: the sibling .pth at the same path preserves the exact model, optimizer, parameter-scheduler, epoch/iteration, and safe MessageHub runtime state. It passes PickleScan and resumes normally with MMEngine 0.8.4. Historical scalar-log buffers are intentionally reset because their HistoryBuffer objects caused the original unsafe-pickle warning; this does not change model or optimization state.

To load safetensors directly, install safetensors, build the model from the corresponding config, and load the returned state dictionary:

from safetensors.torch import load_file

state_dict = load_file("checkpoint.safetensors", device="cpu")
model.load_state_dict(state_dict)

The corresponding Blue, Purple, and Red checkpoints for all four schedules are stored beside Green under ViT_Base/{100,300,800,1600}/{color}/. Each task directory has a canonical config.py; every raw resolved config and scalar-log shard is preserved under runs/<timestamp>/.

See the complete model zoo on GitHub, the checkpoint manifest, the security and resume conversion manifest, and the artifact manifest. The manifests record metrics, model/optimizer/scheduler fingerprints, source evidence, byte sizes, and SHA-256 hashes.

Pretrained models

Model	Params (M)	Flops (G)	Config	Download
colormae_vit-base-p16_8xb512-amp-coslr-300e_in1k.py	111.91	16.87	config	model | log
colormae_vit-base-p16_8xb512-amp-coslr-800e_in1k.py	111.91	16.87	config	model | log
colormae_vit-base-p16_8xb512-amp-coslr-1600e_in1k.py	111.91	16.87	config	model | log

Image Classification on ImageNet-1k

Model	Pretrain	Params (M)	Flops (G)	Top-1 (%)	Config	Download
vit-base-p16_colormae-green-300e-pre_8xb128-coslr-100e_in1k	ColorMAE-G 300-Epochs	86.57	17.58	83.01	config	model | log
vit-base-p16_colormae-green-800e-pre_8xb128-coslr-100e_in1k	ColorMAE-G 800-Epochs	86.57	17.58	83.61	config	model | log
vit-base-p16_colormae-green-1600e-pre_8xb128-coslr-100e_in1k	ColorMAE-G 1600-Epochs	86.57	17.58	83.77	config	model | log

Semantic Segmentation on ADE20K

Model	Pretrain	Params (M)	Flops (G)	mIoU (%)	Config	Download
name	ColorMAE-G 300-Epochs	xx.xx	xx.xx	45.80	config	N/A
name	ColorMAE-G 800-Epochs	xx.xx	xx.xx	49.18	config	N/A

Object Detection and Instance Segmentation on COCO

Model	Pretrain	Params (M)	Flops (G)	$AP^{bbox}$ (%)	Config	Download
name	ColorMAE-G 300-Epochs	xx.xx	xx.xx	48.70	config	N/A
name	ColorMAE-G 800-Epochs	xx.xx	xx.xx	49.50	config	N/A

Using the Models

Predict image

Download the vit-base-p16_colormae-green-300e-pre_8xb128-coslr-100e_in1k.pth pretrained classification model and place it inside the pretrained folder, then run:

from mmpretrain import ImageClassificationInferencer

image = 'https://github.com/open-mmlab/mmpretrain/raw/main/demo/demo.JPEG'
config = 'benchmarks/image_classification/configs/vit-base-p16_8xb128-coslr-100e_in1k.py'
checkpoint = 'pretrained/vit-base-p16_colormae-green-300e-pre_8xb128-coslr-100e_in1k.pth'
inferencer = ImageClassificationInferencer(model=config, pretrained=checkpoint, device='cuda')
result = inferencer(image)[0]
print(result['pred_class'])
print(result['pred_score'])

Use the pretrained model

Also, you can use the pretrained ColorMAE model to extract features.

import torch
from mmpretrain import get_model

config = "configs/colormae_vit-base-p16_8xb512-amp-coslr-300e_in1k.py"
checkpoint = "pretrained/colormae-green-epoch_300.pth"
model = get_model(model=config, pretrained=checkpoint)
inputs = torch.rand(1, 3, 224, 224)
out = model(inputs)
print(type(out))
# To extract features.
feats = model.extract_feat(inputs)
print(type(feats))

Pretraining Instructions

We use mmpretrain for pretraining the models similar to MAE. Please refer here for the instructions: PRETRAIN.md.

Finetuning Instructions

We evaluate transfer learning performance using our pre-trained ColorMAE models on different datasets and downstream tasks including: Image Classification, Semantic Segmentation, and Object Detection. Please refer to the FINETUNE.md file in the corresponding folder.

Acknowledgments

• Our code is based on the MAE implementation of the mmpretrain project: https://github.com/open-mmlab/mmpretrain/tree/main/configs/mae. We thank all contributors from MMPreTrain, MMSegmentation, and MMDetection.
• This work was supported by the KAUST Center of Excellence on GenAI under award number 5940.

How to cite

If you use our code or models in your research, please cite our work as follows:

@inproceedings{hinojosa2024colormae,
  title={ColorMAE: Exploring data-independent masking strategies in Masked AutoEncoders},
  author={Hinojosa, Carlos and Liu, Shuming and Ghanem, Bernard},
  booktitle={European Conference on Computer Vision},
  url={https://www.ecva.net/papers/eccv_2024/papers_ECCV/html/3072_ECCV_2024_paper.php}
  year={2024}
}

Troubleshooting

CuDNN Warning

If you encounter the following warning at the beginning of pretraining:

UserWarning: Applied workaround for CuDNN issue, install nvrtc.so (Triggered internally at /opt/conda/conda-bld/pytorch_1682343995026/work/aten/src/ATen/native/cudnn/Conv_v8.cpp:80.)
  return F.conv2d(input, weight, bias, self.stride,

Solution: This warning indicates a missing or incorrectly linked nvrtc.so library in your environment. To resolve this issue, create a symbolic link to the appropriate libnvrtc.so file. Follow these steps:

1. Navigate to the library directory of your virtual environment:

cd venv/lib/  # Adjust the path if your environment is located elsewhere

2. Create a symbolic link to libnvrtc.so.11.8.89:

ln -sfn libnvrtc.so.11.8.89 libnvrtc.so