Candidate Label Set Pruning: A Data-centric Perspective for Deep Partial-label Learning (ICLR2024)

Requirements

python=3.9.12

torch=1.12.1

protobuf=3.20.1

lavis

faiss-gpu=1.7.2

betaincder=0.1.1

scipy

Datasets

CIFAR-10

CIFAR-100

Tiny-ImageNet

PASCAL_VOC

Deep partial-label learning methods

CC

PRODEN

CAVL

LWS

PiCO

CRDPLL

ABLE

IDGP

SoLa

RECORDS

POP

Note that we use the same training schedule (e.g, same models, optimizers, and hyperparameters) of these methods before and after pruning.

Feature extractors

ResNet-S: trained by supervised learning.

ResNet-SSL: trained by the self-supervised learning method SimCLR

ResNet-I: using the model weight pre-trained on ImageNet-1K

These model weights can be found in google drive.

CLIP

ALBEF

BLIP-2

Run the pruning algorithm

Dataset path: {dataset_root}

Please carefully select parameters used in run.sh.

sh run.sh {gpu_device} {dataset} {partial_rate} {imb_rate} {tau} {k}

For examples ({model_name}='blip2' {model_type}='pretrain'):

Uniform:

sh run.sh 0 cifar10 0.4 0.0 0.6 150

sh run.sh 0 cifar10 0.6 0.0 0.6 150

sh run.sh 0 cifar100 0.01 0.0 0.6 150

sh run.sh 0 cifar100 0.05 0.0 0.6 150

sh run.sh 0 tiny-imagenet 0.01 0.0 0.4 150

sh run.sh 0 tiny-imagenet 0.05 0.0 0.4 150

Instance-dependent (LD):

sh run.sh 0 cifar10 0.0 0.0 0.6 50

sh run.sh 0 cifar100 0.5 0.0 0.6 150

Label-dependent (ID):

sh run.sh 0 cifar10 -1.0 0.0 0.2 5

sh run.sh 0 cifar100 -1.0 0.0 0.2 5

sh run.sh 0 tiny-imagenet -1.0 0.0 0.2 50

Long-tailed (LT):

sh run.sh 0 cifar10 0.3 0.01 0.2 50

sh run.sh 0 cifar10 0.3 0.02 0.2 50

sh run.sh 0 cifar10 0.5 0.01 0.2 50

sh run.sh 0 cifar10 0.5 0.02 0.2 50

sh run.sh 0 cifar100 0.01 0.01 0.2 50

sh run.sh 0 cifar100 0.01 0.02 0.2 50

sh run.sh 0 cifar100 0.05 0.01 0.2 50

sh run.sh 0 cifar100 0.05 0.02 0.2 50

PASCAL_VOC:

sh run.sh 0 voc 0.0 0.0 0.1 5

Original and pruned candidate labels

Original and pruned candidate labels used in the experiment can be found in google drive.

Upper bound in Figure 1

Using cal_er.py to reproduce the Figure 1.

Contact us

If you have any further questions, please feel free to send an e-mail to: shuohe123@gmail.com.