Awesome

Multi-Grained Contrast for Data-Efficient Unsupervised Representation Learning

Official implementation and pre-trained models for paper MGC: "Multi-Grained Contrast for Data-Efficient Unsupervised Representation Learning".

[Paper] [BibTex] [Blog(CN)]

MGC is a novel multi-grained contrast method for unsupervised representation learning. MGC constructs delicate multi-grained correspondences between positive views and then conduct multi-grained contrast by the correspondences to learn more general unsupervised representations. Without pretrained on large-scale dataset, MGC significantly outperforms the existing state-of-the-art methods on extensive downstream tasks, including object detection, instance segmentation, scene parsing, semantic segmentation and keypoint detection.

Requirements

conda create -n mgc python=3.8
pip install -r requirements.txt

Datasets

MGC is pretrained on COCO 2017, VOC and ADE20K datasets, respectively. The root paths of data are respectively set to ./dataset/coco2017, ./dataset/vocand ./dataset/ade20k.

Pre-training

To start the MGC pre-training, run MGC pretraining code on COCO for 800 epochs with the following command.

python main_pretrain.py --arch='vit-small' --dataset='coco2017' --data-root='./dataset/coco2017' --nepoch=800

Pre-trained Model Weights

All models are trained on ViT-S/16 for 800 epochs. For detection, segmentation and keypoint detection downstream tasks, please check evaluation/detection, evaluation/segmentation and evaluation/pose.

Note:

• If you don't have an mircosoft office account, you can download the trained model weights by this link.

• If you have an mircosoft office account, you can download the trained model weights by the links in the following tables.

Evaluation

Our method are evaluate on the following downstream tasks, including object detection, instance segmentation, semantic segmentation and keypoint detection.

1. Object Detection & Instance Segmentation

step 1. Install mmdetection

git clone https://github.com/open-mmlab/mmdetection.git
cd mmdetection
git checkout v2.26.0

step 2. Fine-tune on the COCO dataset

tools/dist_train.sh configs/mgc/mask_rcnn_vit_small_12_p16_1x_coco.py [number of gpu] --seed 0 --work-dir /path/to/saving_dir    --options model.pretrained=/path/to/model_dir

step 3. Fine-tune on the Cityscapes dataset

tools/dist_train.sh configs/mgc/mask_rcnn_vit_small_12_p16_1x_cityscape.py [number of gpu] --seed 0 --work-dir /path/to/saving_dir    --options model.pretrained=/path/to/model_dir

Model weights finetuned on COCO 2017 and Cityscapes:

2. Semantic Segmentation

step 1. Install mmsegmentation

git clone https://github.com/open-mmlab/mmsegmentation.git
cd mmsegmentation
git checkout v0.30.0

step 2. Fine-tune on the ADE20K dataset

tools/dist_train.sh configs/mgc/vit_small_512_160k_ade20k.py [number of gpu] --seed 0 --work-dir /path/to/saving_dir    --options model.pretrained=/path/to/model_dir

step 3. Fine-tune on the VOC07+12 dataset

tools/dist_train.sh configs/mgc/mask_rcnn_vit_small_12_p16_1x_voc.py [number of gpu] --seed 0 --work-dir /path/to/saving_dir    --options model.pretrained=/path/to/model_dir

Model weights finetuned on ADE20K and VOC:

3. Keypoint Detection

step 1. Install mmpose

git clone https://github.com/open-mmlab/mmpose.git
cd mmpose
git checkout v1.2.0

step 2. Fine-tune on the COCO dataset

tools/dist_train.sh configs/mgc/td-hm_ViTPose-small_8xb64-210e_coco-256x192.py [number of gpu] --seed 0 --work-dir /path/to/saving_dir    --options model.pretrained=/path/to/model_dir

Model weights finetuned on COCO 2017:

Acknowledgement

This code is built on mmdetection, mmsegmentation and mmpose libray.

LICENSE

This project is under the CC-BY-NC 4.0 license. See LICENSE for details.

Citation

@article{shen2024multi,
  author  = {Shen, Chengchao and Chen, Jianzhong and Wang, Jianxin},
  title   = {Multi-Grained Contrast for Data-Efficient Unsupervised Representation Learning},
  journal = {arXiv preprint arXiv:2407.02014},
  year    = {2024},
}