Introdunction

This is the official implementation of the paper "Query2Label: A Simple Transformer Way to Multi-Label Classification".

Abstract

This paper presents a simple and effective approach to solving the multi-label classification problem. The proposed approach leverages Transformer decoders to query the existence of a class label. The use of Transformer is rooted in the need of extracting local discriminative features adaptively for different labels, which is a strongly desired property due to the existence of multiple objects in one image. The built-in cross-attention module in the Transformer decoder offers an effective way to use label embeddings as queries to probe and pool class-related features from a feature map computed by a vision backbone for subsequent binary classifications. Compared with prior works, the new framework is simple, using standard Transformers and vision backbones, and effective, consistently outperforming all previous works on five multi-label classification data sets, including MS-COCO, PASCAL VOC, NUS-WIDE, and Visual Genome. Particularly, we establish 91.3% mAP on MS-COCO. We hope its compact structure, simple implementation, and superior performance serve as a strong baseline for multi-label classification tasks and future studies.

Results on MS-COCO:

Quick start

0. (optional) Star this repo.

1. Clone this repo:

git clone [email protected]:SlongLiu/query2labels.git
cd query2labels

2. Install cuda, PyTorch and torchvision.

Please make sure they are compatible. We test our models on two envs and other configs may also work:

cuda==11, torch==1.9.0, torchvision==0.10.0, python==3.7.3
or
cuda==10.2, torch==1.6.0, torchvision==0.7.0, python==3.7.3

3. Install other needed packages.

pip install -r requirments.txt

4. Data preparation.

Download MS-COCO 2014.

5. Train

# single process
python main_mlc.py \
--dataset_dir '/path/to/COCO14/' \
--backbone resnet101 --dataname coco14 --batch-size 64 --print-freq 100 \
--output "path/to/output" \
--world-size 1 --rank 0 --dist-url tcp://127.0.0.1:3717 \
--gamma_pos 0 --gamma_neg 2 --dtgfl \
--epochs 80 --lr 1e-4 --optim AdamW --pretrained \
--num_class 80 --img_size 448 --weight-decay 1e-2 \
--cutout --n_holes 1 --cut_fact 0.5 \
--hidden_dim 2048 --dim_feedforward 8192 \
--enc_layers 1 --dec_layers 2 --nheads 4 \
--early-stop --amp 

# single node multi processes
python -m torch.distributed.launch --nproc_per_node=2 \
main_mlc.py \
--dataset_dir '/path/to/COCO14/' \
--backbone resnet101 --dataname coco14 --batch-size 128 --print-freq 100 \
--output "path/to/output" \
--world-size 1 --rank 0 --dist-url tcp://127.0.0.1:3717 \
--gamma_pos 0 --gamma_neg 2 --dtgfl \
--epochs 80 --lr 1e-4 --optim AdamW --pretrained \
--num_class 80 --img_size 448 --weight-decay 1e-2 \
--cutout --n_holes 1 --cut_fact 0.5 \
--hidden_dim 2048 --dim_feedforward 8192 \
--enc_layers 1 --dec_layers 2 --nheads 4 \
--early-stop --amp 

Pretrianed Models

Download pretrained models.

Modelname	mAP	link(Tsinghua-cloud)
Q2L-R101-448	84.9	model, log
Q2L-R101-576	86.5	model, log
Q2L-TResL-448	87.3	model, log
Q2L-TResL_22k-448	89.2	model, log
Q2L-SwinL-384	90.5	model, log
Q2L-CvT_w24-384	91.3	model, log

Test pretrained models

python q2l_infer.py -a modelname --config /path/to/json/file --resume /path/to/pkl/file [other args]
e.g.
python q2l_infer.py -a 'Q2L-R101-448' --config "pretrained/Q2L-R101-448/config_new.json" -b 16 --resume 'pretrained/Q2L-R101-448/checkpoint.pkl'

BibTex

@misc{liu2021query2label,
      title={Query2Label: A Simple Transformer Way to Multi-Label Classification}, 
      author={Shilong Liu and Lei Zhang and Xiao Yang and Hang Su and Jun Zhu},
      year={2021},
      eprint={2107.10834},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Acknowledgement

We thank the authors of ASL, TResNet, detr, CvT, and Swin-Transformer for their great works and codes. Thanks to @mrT23 for sharing training tricks and providing a useful script for training.