Sirius222 / spll

SOP+Partial label learning

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

cifar 10

Cifar10_aggre 0.956(sop)

Cifar10_random 0.953 (sop)

cifar10_worse_label 93.4(sop+comatch)

worse_label

固定种子没有重复实验

SOP best acc 93%

50%数据监督训练:

监督训练(50%样本+mixup+增强): best acc:93.07% (213服务器 train_sup_mixup.py)

监督训练(50%样本):beast acc:90.42 (213服务器 clean train.py)

SOP+偏标记(do not work)

100 epoch的loss,predicition集成:

监督训练(50%样本+mixup+增强) + 偏标记训练(50%样本+候选集为4):best acc 91.15%

监督训练(50%样本+mixup+增强) + 偏标记训练(50%样本+候选集为3):best acc 91.95%

监督训练(50%样本+mixup+增强) + 偏标记训练(50%样本+候选集为2):best acc 93.03%

SOP+半监督

50%干净 50%无标记 comatch :best ema acc is 93.48% (ying-peng dir:comatch)

About

SOP+Partial label learning


Languages

Language:Python 100.0%