HCSC: Hierarchical Contrastive Selective Coding

This repository provides a PyTorch implementation and model weights for HCSC (Hierarchical Contrastive Selective Coding), whose details are in this paper.

HCSC is an effective and efficient method to pre-train image encoders in a self-supervised fashion. In general, this method seeks to learn image representations with hierarchical semantic structures. It utilizes hierarchical K-means to derive hierarchical prototypes, and these prototypes represent the hierarchical semantics underlying the data. On such basis, we perform Instance-wise and Prototypical Contrastive Selective Coding to inject the information within hierarchical prototypes into image representations. HCSC has achieved SOTA performance on the self-supervised pre-training of CNNs (e.g., ResNet-50), and we will further study its potential on pre-training Vision Transformers.

Roadmap

• [2022/02/01] The initial release! We release all source code for pre-training and downstream evaluation. We release three pre-trained ResNet-50 models: 200 epochs (single-crop), 200 epochs (multi-crop) and 400 epochs (single-crop, batch size: 256).

TODO

• Finish the pre-training of 400 epochs ResNet-50 models (multi-crop) and release.
• Finish the pre-training of 800 epochs ResNet-50 models (single- & multi-crop) and release.
• Support Vision Transformer backbones.
• Pre-train Vision Transformers with HCSC and release model weights under various configurations.

Model Zoo

We will continually release our pre-trained HCSC model weights and corresponding training configs. The current finished ones are as follows:

Installation

Use following command to install dependencies (python3.7 with pip installed):

pip3 install -r requirement.txt

If having trouble installing PyTorch, follow the original guidance (https://pytorch.org/). Notably, the code is tested with cudatoolkit version 10.2.

Pre-training on ImageNet

Download ImageNet dataset under [ImageNet Folder]. Go to the path "[ImageNet Folder]/val" and use this script to build sub-folders.

To train single-crop HCSC on 8 Tesla-V100-32GB GPUs for 200 epochs, run:

python3 -m torch.distributed.launch --master_port [your port] --nproc_per_node=8 \
pretrain.py [your ImageNet Folder]

To train multi-crop HCSC on 8 Tesla-V100-32GB GPUs for 200 epochs, run:

python3 -m torch.distributed.launch --master_port [your port] --nproc_per_node=8 \
pretrain.py --multicrop [your ImageNet Folder]

Downstream Evaluation

Evaluation: Linear Classification on ImageNet

With a pre-trained model, to train a supervised linear classifier with all available GPUs, run:

python3 eval_lincls_imagenet.py --data [your ImageNet Folder] \
--dist-url tcp://localhost:10001 --world-size 1 --rank 0 \
--pretrained [your pre-trained model (example:out.pth)]

Evaluation: KNN Evaluation on ImageNet

To reproduce the KNN evaluation results with a pre-trained model using a single GPU, run:

python3 -m torch.distributed.launch --master_port [your port] --nproc_per_node=1 eval_knn.py \
--checkpoint_key state_dict \
--pretrained [your pre-trained model] \
--data [your ImageNet Folder]

Evaluation: Semi-supervised Learning on ImageNet

To fine-tune a pre-trained model with 1% or 10% ImageNet labels with 8 Tesla-V100-32GB GPUs, run:

1% of labels:

python3 -m torch.distributed.launch --nproc_per_node 8 --master_port [your port] eval_semisup.py \
--labels_perc 1 \
--pretrained [your pretrained weights] \
[your ImageNet Folder]

10% of labels:

python3 -m torch.distributed.launch --nproc_per_node 8 --master_port [your port] eval_semisup.py \
--labels_perc 10 \
--pretrained [your pretrained weights] \
[your ImageNet Folder]

Evaluation: Transfer Learning - Classification on VOC / Places205

VOC

1. Download the VOC dataset.

2. Finetune and evaluate on PASCAL VOC (with a single GPU):

cd voc_cls/ 
python3 main.py --data [your voc data folder] \
--pretrained [your pretrained weights]

Places205

1. Download the Places205 dataset (resized 256x256 version)

2. Linear Classification on Places205 (with all available GPUs):

python3 eval_lincls_places.py --data [your places205 data folder] \
--data-url tcp://localhost:10001 \
--pretrained [your pretrained weights]

Evaluation: Transfer Learning - Object Detection on VOC / COCO

1. Download VOC and COCO Dataset (under ./detection/datasets).

2. Install detectron2.

3. Convert a pre-trained model to the format of detectron2:

cd detection
python3 convert-pretrain-to-detectron2.py [your pretrained weight] out.pkl

4. Train on PASCAL VOC/COCO:

Finetune and evaluate on VOC (with 8 Tesla-V100-32GB GPUs):

cd detection
python3 train_net.py --config-file ./configs/pascal_voc_R_50_C4_24k_hcsc.yaml \
--num-gpus 8 MODEL.WEIGHTS out.pkl

Finetune and evaluate on COCO (with 8 Tesla-V100-32GB GPUs):

cd detection
python3 train_net.py --config-file ./configs/coco_R_50_C4_2x_hcsc.yaml \
--num-gpus 8 MODEL.WEIGHTS out.pkl

Evaluation: Clustering Evaluation on ImageNet

To reproduce the clustering evaluation results with a pre-trained model using all available GPUs, run:

python3 eval_clustering.py --dist-url tcp://localhost:10001 \
--multiprocessing-distributed --world-size 1 --rank 0 \
--num-cluster [target num cluster] \
--pretrained [your pretrained model weights] \
[your ImageNet Folder]

In the experiments of our paper, we set --num-cluster as 25000 and 1000.

License

This repository is released under the MIT license as in the LICENSE file.

Citation

If you find this repository useful, please kindly consider citing the following paper:

@article{guo2022hcsc,
  title={HCSC: Hierarchical Contrastive Selective Coding},
  author={Guo, Yuanfan and Xu, Minghao and Li, Jiawen and Ni, Bingbing and Zhu, Xuanyu and Sun, Zhenbang and Xu, Yi},
  journal={arXiv preprint arXiv:2202.00455},
  year={2022}
}