This repo is the official Pytorch implementation of our paper:
Roll with the Punches: Expansion and Shrinkage of Soft Label Selection for Semi-supervised Fine-Grained Learning
Authors: Yue Duan, Zhen Zhao, Lei Qi, Luping Zhou, Lei Wang and Yinghuan Shi
-
🔗 Quick links: [PDF/Abs-arXiv | PDF/Abs-Published | Poster/Slides | Video | Code Download]
-
📰 Latest news:
-
📑 Related works:
- Interested in robust SSL in MNAR setting with mismatched distributions? 👉 Check out our ICCV'23 paper PRG [PDF-arXiv | Code] or our ECCV'22 paper RDA [PDF-arXiv | Code].
- Interested in the conventional SSL or application of complementary label in SSL? 👉 Check out our TNNLS paper MutexMatch [PDF-arXiv | Code].
- numpy==1.20.3
- Pillow==9.3.0
- scikit_learn==1.1.3
- torch==1.8.0
- torchvision==0.9.0
-
--net [resnet18/resnet50/resnet101/wrn/wrnvar/preresnet/cnn13]
: By default, ResNet-50 is used for experiments. We provide alternatives as follows: ResNet-18/101, Wide ResNet, Wide ResNet-Var (WRN-37-2), PreAct ResNet and CNN-13. -
--dataset [semi-fungi/semi-aves]
and--data_dir
: Your dataset name and path. We support two datasets: Semi-Fungi and Semi-Aves (See "A Realistic Evaluation of Semi-Supervised Learning for Fine-Grained Classification" for details). -
--unlabel [in/inout]
: The unlabeled data used for training. The unlabeled data of Semi-Fungi and Semi-Aves contains two subsets. One has in-distribution data only, the other one has both in-distribution and out-of-distribution data. -
--alpha
:$\alpha$ for confidence-aware k selection. -
--num_tracked_batch
:$N_{B}$ for class transition tracking (CTT). -
--resume
,--load_path
,--load_path_soc
and--pretrained
:- If you want to resume training without
soc.pkl
, i.e., saved data ofcentroids
(for CTT-based k-means clustering) andlabel_matrix
(for CTT), please set--resume --load_path @path to your checkpoint@
. - If you want to resume training with
soc.pkl
, please set--resume --load_path @path to your checkpoint@ --load_path_soc @path to your soc.pkl@
. - If you want to use the expert model or MoCo model, please set
--resume --load_path @path to expert model/MoCo model@ --pretrained
.
- If you want to resume training without
python train_soc.py --rank 0 --gpu [0/1/...] @@@other args@@@
- Using DataParallel
python train_soc.py --world-size 1 --rank 0 @@@other args@@@
- Using DistributedDataParallel with single node
python train_soc.py --world-size 1 --rank 0 --multiprocessing-distributed @@@other args@@@
To better reproduce our experimental results, it is recommended to use multi-GPUs with DataParallel for training.
python train_soc.py --world-size 1 --rank 0 --seed 1 --num_eval_iter 2000 --overwrite --save_name aves_in_sc --dataset semi_aves --num_classes 200 --unlabel in
python train_soc.py --world-size 1 --rank 0 --seed 1 --num_eval_iter 1000 --overwrite --save_name aves_in_sc_moco --dataset semi_aves --num_classes 200 --unlabel in --resume --load_path @path to MoCo pre-trained model@ --pretrained --num_train_iter 200000
python train_soc.py --world-size 1 --rank 0 --seed 1 --num_eval_iter 500 --overwrite --save_name aves_in_pre --dataset semi_aves --num_classes 200 --unlabel in --resume --load_path @path to pre-trained model@ --pretrained --lr 0.001 --num_train_iter 50000
The expert models and MoCo models can be obtained here (provided by https://github.com/cvl-umass/ssl-evaluation).
python train_soc.py --world-size 1 --rank 0 --seed 1 --num_eval_iter 2000 --overwrite --save_name aves_inout_sc --dataset semi_aves --num_classes 200 --unlabel inout
Each time you start training, the evaluation results of the current model will be displayed. If you want to evaluate a model, use its checkpoints to resume training, i.e., use --resume --load_path @path to your checkpoint@
.
Dateset | Unlabeled Data | Pre-training | Top1-Acc (%) | Top5-Acc (%) | Checkpoint |
---|---|---|---|---|---|
Semi-Aves | in-distribution | - | 32.3 | 55.5 | here |
MoCo | 39.5 | 62.5 | here | ||
ImageNet | 56.8 | 79.1 | here | ||
ImageNet + MoCo | 57.1 | 79.1 | here | ||
iNat | 71.0 | 88.4 | here | ||
iNat + MoCo | 70.2 | 88.3 | here | ||
out-of-distribution | - | 27.5 | 50.7 | here | |
MoCo | 40.4 | 65.9 | here | ||
Semi-Fungi | in-distribution | - | 38.50 | 61.35 | here |
MoCo | 46.9 | 71.4 | here | ||
ImageNet | 61.1 | 83.2 | here | ||
ImageNet + MoCo | 61.8 | 85.9 | here | ||
iNat | 62.3 | 85.0 | here | ||
iNat + MoCo | 62.2 | 84.4 | here | ||
out-of-distribution | - | 35.6 | 60.6 | here | |
MoCo | 50.0 | 74.8 | here |
Please cite our paper if you find SoC useful:
@inproceedings{duan2024roll,
title={Roll with the Punches: Expansion and Shrinkage of Soft Label Selection for Semi-supervised Fine-Grained Learning},
author={Duan, Yue and Zhao, Zhen and Qi, Lei and Zhou, Luping and Wang, Lei and Shi, Yinghuan},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
volume={38},
number={10},
pages={11829--11837},
year={2024}
}
or
@article{duan2023roll,
title={Roll With the Punches: Expansion and Shrinkage of Soft Label Selection for Semi-supervised Fine-Grained Learning},
author={Duan, Yue and Zhao, Zhen and Qi, Lei and Zhou, Luping and Wang, Lei and Shi, Yinghuan},
journal={arXiv preprint arXiv:2312.12237},
year={2023}
}