lzxzy / hier-cvpr23

Official PyTorch Implementation of HIER: Metric Learning Beyond Class Labels via Hierarchical Regularization, CVPR 2023

HIER: Metric Learning Beyond Class Labels
via Hierarchical Regularization (CVPR 2023)

Sungyeon Kim, Boseung Jeong, Suha Kwak

[paper], [project hompage]

This repository contains the official PyTorch implementation of the paper HIER: Metric Learning Beyond Class Labels via Hierarchical Regularization, which is published in CVPR 2023. Our proposed approach, called HIER, is a new regularization method for metric learning. It aims to discover the latent semantic hierarchy of training data and deploy the hierarchy to provide richer and more fine-grained supervision than inter-class separability induced by common metric learning losses.

Overview

What is HIER

A conceptual illustration of the learning objective in HIER loss. Each hierarchical proxy is colored in black and different colors indicate different classes. The associations defined by the losses are expressed by edges, where the red solid line means the pulling and the blue dotted line is the pushing. Relevant pairs are pulled into their LCA, and the remaining sample is pulled into LCA of the entire triplet.

What HIER does

A tree-like hierarchical structure between data and hierarchical proxies is constructed in the embedding space learned by HIER. The middle figures are enlarged figures of part of the embedding space, which shows that samples in a sub-tree share common semantics, although not of the same class. This visualization proves that HIER discovers and deploys latent and meaningful hierarchies between data in the embedding space.

See this example for how to produce these visualizations.

Experimental Results

Installation

See INSTALL.md for installation details.

Datasets

1. Download four public benchmarks for deep metric learning

   • CUB-200-2011
   • Cars-196 (Img, Annotation)
   • Stanford Online Products (Link)
   • In-shop Clothes Retrieval (Link)

2. Extract the tgz or zip file into ./data/ (Exceptionally, for Cars-196, put the files in a ./data/cars196)

Training Embedding Network

Train embedding model with Proxy-Anchor loss and HIER on a single node with 4 gpus for 100 epochs run:

python -m torch.distributed.launch \
    --master_port=${MASTER_PORT:-$RANDOM} \
    --nproc_per_node=4 train.py \
    --data_path /path/to/dataset \
    --lr {1e-4 1e-5} \
    --num_hproxies 512 \
    --weight_decay {1e-2 1e-4} \
    --lambda1 1 --lambda2 {0.5 1} \
    --dataset {CUB Cars SOP Inshop} \
    --model {resnet50 deit_small_distilled_patch16_224 vit_small_patch16_224 dino_vits16} \
    --IPC {0 2} --loss {MS PA} \
    --emb {128 384 512} \ 
    --bn_freeze True 

For details, see scripts folder.

We also support training embedding model with Hyp:

python -m torch.distributed.launch \
    --master_port=${MASTER_PORT:-$RANDOM} \
    --nproc_per_node=4 train.py \
    --data_path /path/to/dataset \
    --lr {1e-4 1e-5} \
    --num_hproxies 512 \
    --weight_decay {1e-2 1e-4} \
    --lambda1 1 --lambda2 0 \
    --dataset {CUB Cars SOP Inshop} \
    --model {resnet50 deit_small_distilled_patch16_224 vit_small_patch16_224 dino_vits16} \
    --IPC 8 --loss SupCon \
    --emb {128 384 512} \ 
    --bn_freeze True 

Acknowledgements

Our source code is modified and adapted on these great repositories:

• Hyperbolic Vision Transformers: Combining Improvements in Metric Learning
• Proxy Anchor Loss for Deep Metric Learning
• PyTorch Metric learning

Citation

If you use this method or this code in your research, please cite as:

@inproceedings{kim2023hier,
  title={HIER: Metric Learning Beyond Class Labels via Hierarchical Regularization},
  author={Kim, Sungyeon and Jeong, Boseung and Kwak, Suha},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  year={2023}
}