Skip to content

SoongE/shot

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

9 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Shot

Pytorch implement of [ICML-2020] Do We Really Need to Access the Source Data? Source Hypothesis Transfer for Unsupervised Domain Adaptation

Experiment Details

범주 카테고리 내용
데이터 Train Augmentation Resize, RandomCrop, RandomHorizontalFlip, Normalize
Test Augmentation Resize, CenterCrop, Normalize
모델 구조 ResNet50 (Pre-trained on MS1V3)
학습도구 Optimizer SGD, lr=0.01, weight_decay=1e-3, momentum=0.9
Criterion(source train) CrossEntropyLoss With LabelSmooth(0.1)
Criterion(target train) CrossEntropyLoss, Entropy Loss, IM Loss
LR Scheduler 1 + gamma * iter_num / max_iter) ** power 만큼 epoch마다 감소
학습 Epoch(source/target) 50 / 15
Batch size 128
평가 Softmax CDA, PDA
Softmax & threshold 라벨에 해당하는 이미지는 softmax, 라벨에 해당하지 않는 이미지는 1,0

Result

Office-Home

PDA Ar->Cl Ar->Pr Ar->Re
source only (repository) 46.4 71.8 80.6
source only (paper) 45.2 70.4 81.0
SHOT (repository) 61.0 82.8
SHOT (paper) 64.8 85.2 92.7

RMFD

Dataset Description

# Image # class
Non Masked 87228 442
Masked 1945 442

Result

CDA Unmask -> Mask
source only(unmask) 9.05
SHOT 0.5 (학습 안됌)
arcfaceBackbone transfer learning unmask 22.7
arcfaceBackbone + SHOT 0.3 (학습 안됌)

Pseudo labeling이 제대로 되지 않음. 이전보다 정확도가 더 떨어짐. class의 수(443)가 너무 많아서 제대로 학습되지 않는 것으로 보임

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published