Chapter5 - Homework - loss_function&Deep_Metric_Learning - shenlanxueyuan
(https://drive.google.com/file/d/1wJC2aPA4AC0rI-tAL2BFs2M8vfcpX-w6/view?usp=sharing)
unzip casia-maxpy-clean.zip
cd casia-maxpy-clean
zip -F CASIA-maxpy-clean.zip --out CASIA-maxpy-clean_fix.zip
unzip CASIA-maxpy-clean_fix.zip
(https://pan.baidu.com/s/1Rue4FBmGvdGMPkyy2ZqcdQ)
SEResNet18 | Best acc on LFW | SEResNet34 | Best acc on LFW |
---|---|---|---|
Softmax | 0.851 | Softmax | 0.8578333333333333 |
NormFace | 0.8428333333333334 | NormFace | 0.8470000000000001 |
SpereFace | 0.8474999999999999 | SpereFace | 0.8651666666666665 |
CosFace | 0.8488333333333333 | CosFace | 0.8504999999999999 |
ArcFace | 0.8456666666666667 | ArcFace | 0.755 |
OHEM & NormFace | 0.8456666666666667 | OHEM & NormFace | 0.8485000000000001 |
FocalLoss & NormFace | 0.8396666666666667 | FocalLoss & NormFace | 0.8504999999999999 |
Notes: Train from scratch and run 20 epochs @ Tesla P100 16G; SEResNet18 @ 20epoch、batchsize=256; SEResNet34 @20epoch、batchsize=128
SEResNet18 | Best acc on LFW | SEResNet34 | Best acc on LFW |
---|---|---|---|
Contrastive(Scratch) | 0.6135(20epoch、batchsize=128) | Contrastive(Scratch) | \ |
Triplet(Scratch) | 0.7968(4epoch、batchsize=64) | Triplet(Scratch) | 0.8265(4epoch、batchsize=256、Quadro RTX8000 48G) |
Contrastive(Finetune) | 0.6371666666666667(20epoch、batchsize=128) | Contrastive(Finetune) | \ |
Triplet(Finetune) | \ | Triplet(Finetune) | \ |