README.md

IRGen

A PyTorch implementation of IRGen based on the paper IRGen: Generative Modeling for Image Retrieval.

Requirements

• Anaconda
• PyTorch

conda install pytorch torchvision cudatoolkit=10.0 -c pytorch
pip install -r requirements.txt

Datasets

CARS196, CUB200-2011 and In-shop Clothes are used in this repo.

You should download these datasets by yourself, and extract them into data/car, data/cub, data/isc directory. For each dataset, run the following instruction to generate the ground truth file.

python gnd_generater.py 

Usage

Train Tokenizer

python train_tokenizer.py 
optional arguments:
--data_path                   datasets path [default value is 'data']
--data_name                   dataset name [default value is 'isc'](choices=['car', 'cub', 'isc', 'imagenet', 'places'])
--feats                       initialize features for quantize [default value is '']
--output_dir                  saving output direction [default value is 'results']
--lr                          train learning rate [default value is 5e-4]
--batch_size                  train batch size [default value is 128]
--num_epochs                  train epoch number [default value is 200]
--rq_weight                   loss weight for rq reconsturcted features

Get image tokens

python rq.py --features 'isc_features.npy' --file_name 'isc_rq.pkl' --data_dir 'data/isc'
optional arguments:
--data_name                   dataset name [default value is 'isc'](choices=['car', 'cub', 'isc', 'imagenet', 'places'])

Train IRGen

python -m torch.distributed.launch --nproc_per_node=8 train_ar.py --file_name 'in-shop_clothes_retrieval_trainval.pkl' --codes 'isc_rq.pkl'
optional arguments:
--data_dir                    datasets path [default value is 'data/isc/Img']
--data_name                   dataset name [default value is 'isc'](choices=['car', 'cub', 'isc', 'imagenet', 'places'])
--output_dir                  saving output direction [default value is 'results']
--lr                          train learning rate [default value is 8e-5]
--batch_size                  train batch size [default value is 64]
--num_epochs                  train epoch number [default value is 200]
--smoothing                   smoothing value for label smoothing [default value is 0.1]

Test IRGen

python test_ar.py --file_name 'in-shop_clothes_retrieval_trainval.pkl' --codes 'isc_rq.pkl' --model_dir 'results/isc_rq_e200.pkl' 
optional arguments:
--data_dir                    datasets path [default value is 'data/isc/Img']
--data_name                   dataset name [default value is 'isc'](choices=['car', 'cub', 'isc', 'imagenet', 'places'])
--beam_size                   size for beam search [default value is 30]
--ks                          query number for test@k[default value is [1,10,20,30]]

Benchmarks

The models are trained on 8 NVIDIA Tesla V100 (32G) GPU.

In-shop

P refers to precision, R refers to recall.

P@1	P@10	P@20	P@30	R@1	R@10	R@20	R@30
92.4%	87.4%	87.0%	86.9%	92.4%	96.8%	97.6%	97.9%

CUB200

P@1	P@2	P@4	P@8	R@1	R@2	R@4	R@8
82.7%	82.7%	83.0%	82.8%	82.7%	86.4%	89.2%	91.4%

CARS196

P@1	P@2	P@4	P@8	R@1	R@2	R@4	R@8
90.1%	89.9%	90.2%	90.5%	90.1%	92.1%	93.2%	93.7%

Results

The Precision-Recall curve.

In-shop Clothes

CUB200

Cars196