Embedding Transfer with Label Relaxation for Improved Metric Learning

Official PyTorch implementation of CVPR 2021 paper Embedding Transfer with Label Relaxation for Improved Metric Learning.

Embedding trnasfer with Relaxed Contrastive Loss improves performance, or reduces sizes and output dimensions of embedding model effectively.

This repository provides source code of experiments on three datasets (CUB-200-2011, Cars-196 and Stanford Online Products) including relaxed contrastive loss, relaxed MS loss, and 6 other knowledge distillation or embedding transfer methods such as:

• FitNet, Fitnets: hints for thin deep nets
• Attention, Paying More Attention to Attention: Improving the Performance of Convolutional Neural Networks via Attention Transfer
• CRD, Contrastive Representation Distillation
• DarkRank, Darkrank: Accelerating Deep Metric Learning via Cross Sample Similarities Transfer
• PKT, Learning Deep Representations with Probabilistic Knowledge Transfer
• RKD, Relational Knowledge Distillation

Overview

Relaxed Contrastive Loss

• Relaxed contrastive loss exploits pairwise similarities between samples in the source embedding space as relaxed labels, and transfers them through a contrastive loss used for learning target embedding models.

Experimental Restuls

• Our method achieves the state of the art when embedding dimension is 512, and is as competitive as recent metric learning models even with a substantially smaller embedding dimension. In all experiments, it is superior to other embedding transfer techniques.

Requirements

• Python3
• PyTorch (> 1.0)
• NumPy
• tqdm
• wandb
• Pytorch-Metric-Learning

Prepare Datasets

1. Download three public benchmarks for deep metric learning.

   • CUB-200-2011
   • Cars-196 (Img, Annotation)
   • Stanford Online Products

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

Prepare Pretrained Source models

Download the pretrained source models using ./scripts/download_pretrained_source_models.sh.

sh scripts/download_pretrained_source_models.sh

Training Target Embedding Network with Relaxed Contrastive Loss

Self-transfer Setting

• Transfer the knowledge of source model to target model with the same architecture and embedding dimension for performance improvement.
• Source Embedding Network (BN–Inception, 512 dim) 🠢 Target Embedding Network (BN–Inception, 512 dim)

CUB-200-2011

python code/train_target.py --gpu-id 0 --loss Relaxed_Contra --model bn_inception \
--embedding-size 512 --batch-size 90 --IPC 2 --dataset cub --epochs 90 \
--source-ckpt ./pretrained_source/bn_inception/cub_bn_inception_512dim_Proxy_Anchor_ckpt.pth \
--view 2 --sigma 1 --delta 1 --save 1

Cars-196

python code/train_target.py --gpu-id 0 --loss Relaxed_Contra --model bn_inception \ 
--embedding-size 512 --batch-size 90 --IPC 2 --dataset cars --epochs 90 \
--source-ckpt ./pretrained_source/bn_inception/cars_bn_inception_512dim_Proxy_Anchor_ckpt.pth \
--view 2 --sigma 1 --delta 1 --save 1

SOP

python code/train_target.py --gpu-id 0 --loss Relaxed_Contra --model bn_inception \
--embedding-size 512 --batch-size 90 --IPC 2 --dataset SOP --epochs 150 \
--source-ckpt ./pretrained_source/bn_inception/SOP_bn_inception_512dim_Proxy_Anchor_ckpt.pth \
--view 2 --sigma 1 --delta 1 --save 1

Dimensionality Reduction Setting

• Transfer to the same architecture with a lower embedding dimension for efficient image retrieval.
• Source Embedding Network (BN–Inception, 512 dim) 🠢 Target Embedding Network (BN–Inception, 64 dim)

CUB-200-2011

python code/train_target.py --gpu-id 0 --loss Relaxed_Contra --model bn_inception \
--embedding-size 64 --batch-size 90 --IPC 2 --dataset cub --epochs 90 \
--source-ckpt ./pretrained_source/bn_inception/cub_bn_inception_512dim_Proxy_Anchor_ckpt.pth \
--view 2 --sigma 1 --delta 1 --save 1

Cars-196

python code/train_target.py --gpu-id 0 --loss Relaxed_Contra --model bn_inception \
--embedding-size 64 --batch-size 90 --IPC 2 --dataset cars --epochs 90 \
--source-ckpt ./pretrained_source/bn_inception/cars_bn_inception_512dim_Proxy_Anchor_ckpt.pth \
--view 2 --sigma 1 --delta 1 --save 1

SOP

python code/train_target.py --gpu-id 0 --loss Relaxed_Contra --model bn_inception \
--embedding-size 64 --batch-size 90 --IPC 2 --dataset SOP --epochs 150 \
--source-ckpt ./pretrained_source/bn_inception/SOP_bn_inception_512dim_Proxy_Anchor_ckpt.pth \
--view 2 --sigma 1 --delta 1 --save 1

Model Compression Setting

• Transfer to a smaller network with a lower embedding dimension for usage in low-power and resource limited devices.
• Source Embedding Network (ResNet50, 512 dim) 🠢 Target Embedding Network (ResNet18, 128 dim)

CUB-200-2011

python code/train_target.py --gpu-id 0 --loss Relaxed_Contra --model resnet18 \
--embedding-size 128 --batch-size 90 --IPC 2 --dataset cub --epochs 90 \
--source-ckpt ./pretrained_source/resnet50/cub_resnet50_512dim_Proxy_Anchor_ckpt.pth \
--view 2 --sigma 1 --delta 1 --save 1

Cars-196

python code/train_target.py --gpu-id 0 --loss Relaxed_Contra --model resnet18 \
--embedding-size 128 --batch-size 90 --IPC 2 --dataset cars --epochs 90 \
--source-ckpt ./pretrained_source/resnet50/cars_resnet50_512dim_Proxy_Anchor_ckpt.pth \
--view 2 --sigma 1 --delta 1 --save 1

SOP

python code/train_target.py --gpu-id 0 --loss Relaxed_Contra --model resnet18 \
--embedding-size 128 --batch-size 90 --IPC 2 --dataset SOP --epochs 150 \
--source-ckpt ./pretrained_source/resnet50/SOP_resnet50_512dim_Proxy_Anchor_ckpt.pth \
--view 2 --sigma 1 --delta 1 --save 1

Train Source Embedding Network

This repository also provides code for training source embedding network with several losses as well as proxy-anchor loss. For details on how to train the source embedding network, please see the Proxy-Anchor Loss repository.

• For example, training source embedding network (BN–Inception, 512 dim) with Proxy-Anchor Loss on the CUB-200-2011 as

python code/train_source.py --gpu-id 0 --loss Proxy_Anchor --model bn_inception \
--embedding-size 512 --batch-size 180 --lr 1e-4 --dataset cub \
--warm 1 --bn-freeze 1 --lr-decay-step 10 

Evaluating Image Retrieval

Follow the below steps to evaluate the trained model.
Trained best model will be saved in the ./logs/folder_name.

# The parameters should be changed according to the model to be evaluated.
python code/evaluate.py --gpu-id 0 \
                   --batch-size 120 \
                   --model bn_inception \
                   --embedding-size 512 \
                   --dataset cub \
                   --ckpt /set/your/model/path/best_model.pth

Acknowledgements

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

• Proxy Anchor Loss for Deep Metric Learning
• No Fuss Distance Metric Learning using Proxies
• Contrastive Representation Distillation
• Relational Knowledge Distillation
• PyTorch Metric learning

Citation

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

@inproceedings{kim2021embedding,
  title={Embedding Transfer with Label Relaxation for Improved Metric Learning},
  author={Kim, Sungyeon and Kim, Dongwon and Cho, Minsu and Kwak, Suha},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  year={2021}
}