This repository provides a PyTorch implementation and model weights for HCSC (Hierarchical Contrastive Selective Coding), whose details are in this paper (accepted by CVPR 2022).

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.

• [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).
• [2022/03/03] More checkpoints released! We update the 400 epochs single-crop model (with 71.0 linear classification accuracy, 64.1 KNN accuracy) and add two new pre-trained ResNet-50 models: 400 epochs (multi-crop, batch size: 256) and 800 epochs (single-crop, batch size: 256).
• [2022/03/22] 800 epochs multi-crop model (batch size: 256) released! This pre-trained model achieves 74.2 linear classification accuracy and 70.6 KNN accuracy.

• 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.

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

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.

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]

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)]

Note: for single-gpu evaluation, turn off the multiprocessing-distributed. For example, run:

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

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]

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

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]

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]

Note: for single-gpu evaluation, start the script without torch.distributed.launch. Remember to modify the --batch_size because it is the batch-size-per-gpu. Example:

python3 eval_semisup.py --labels_perc 1 \
--batch_size 256 --pretrained [your pretrained weights] \
[your ImageNet Folder]

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

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

Note: for single-gpu evaluation, see instruction above.

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

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

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

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.

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

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}
}