README.md

Informative Sample-Aware Proxy for Deep Metric Learning

Official PyTorch implementation of ACM MM Asia 2022 paper Informative Sample-Aware Proxy for Deep Metric Learning.

Requirements

• Python3
• PyTorch (> 1.6)
• NumPy
• tqdm
• wandb
• plotly (needed if you want to visualize with t-SNE)

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 ./data/cars196)

Training Embedding Network

CUB-200-2011

• Train a embedding network of Inception-BN using Proxy-ISA

python train.py --gpu_id 0 \
                --loss ProxyISA \
                --model bn_inception \
                --embedding_size 512 \
                --batch_size 128 \
                --lr 1e-4 \
                --dataset cub \
                --warm 0 \
                --lr_decay_step 10 \
                --enableMemory True

• Train a embedding network of ResNet-50 using Proxy-ISA

python train.py --gpu_id 0 \
                --loss ProxyISA \
                --model resnet50 \
                --embedding_size 512 \
                --batch_size 128 \
                --lr 1e-4 \
                --dataset cub \
                --warm 0 \
                --lr_decay_step 5 \
                --enableMemory True

Cars-196

• Train a embedding network of Inception-BN using Proxy-ISA

python train.py --gpu_id 0 \
                --loss ProxyISA \
                --model bn_inception \
                --embedding_size 512 \
                --batch_size 128 \
                --lr 1e-4 \
                --dataset cars \
                --warm 0 \
                --lr_decay_step 20 \
                --enableMemory True \
                --k 0.4

• Train a embedding network of ResNet-50 using Proxy-ISA

python train.py --gpu_id 0 \
                --loss ProxyISA \
                --model resnet50 \
                --embedding_size 512 \
                --batch_size 128 \
                --lr 1e-4 \
                --dataset cars \
                --warm 0 \
                --lr_decay_step 10 \
                --enableMemory True \
                --k 0.4

Stanford Online Products

• Train a embedding network of Inception-BN using Proxy-ISA

python train.py --gpu_id 0 \
                --loss ProxyISA \
                --model bn_inception \
                --optimizer adamw \
                --embedding_size 512 \
                --batch_size 128 \
                --lr 6e-4 \
                --dataset SOP \
                --warm 1 \
                --bn_freeze False \
                --lr_decay_step 20 \
                --lr_decay_gamma 0.25 \
                --enableMemory True

In-Shop Clothes Retrieval

• Train a embedding network of Inception-BN using Proxy-ISA

python train.py --gpu_id 0 \
                --loss ProxyISA \
                --model bn_inception \
                --optimizer adamw \
                --embedding_size 512 \
                --batch_size 128 \
                --lr 6e-4 \
                --dataset Inshop \
                --warm 1 \
                --bn_freeze False \
                --lr_decay_step 20 \
                --lr_decay_gamma 0.25 \
                --enableMemory True \
                --k 0.1

Evaluating Image Retrieval

Follow the below steps to evaluate the trained model.

Trained best model will be saved in ./logs/folder_name.

# The parameters should be changed according to the model to be evaluated.
python evaluate.py --gpu_id 0 \
                   --batch_size 128 \
                   --model bn_inception \
                   --embedding_size 512 \
                   --dataset cub \
                   --resume /PATH/TO/YOUR/Model.pth

Embedding Space Visualization

t-SNE visualization of 512-dimensional embedding space for the Cars-196 dataset (during training).

Left: Proxy-Anchor loss (Kim et al. CVPR 2020); Right: Proxy-ISA (Ours)

 

Citation

@InProceedings{Li_2022_MMAsia,
  title = {Informative Sample-Aware Proxy for Deep Metric Learning},
  author = {Li, Aoyu and Sato, Ikuro and Ishikawa, Kohta and Kawakami, Rei and Yokota, Rio},
  booktitle = {ACM Multimedia Asia (MMAsia '22)},
  year = {2022},
  doi = {10.1145/3551626.3564942}
}