chexuanqiao / pixelwise_knowledge_distillation_pspnet18

Pixelwise_Knowledge_Distillation_PSPNet18

A PyTorch implementation of PSPNet18 with pixelwise knowledge distillation.

This repo is based on and modified from semseg.

Crop size 481 * 481
With 713 * 713 pretrained PSPNet50 as teacher_net:
Main_loss = Kldiv * (1 - alpha) + cross_entropy * alpha

Config	mIoU/mAcc/aAcc
T1.8 A0.2	0.7130/0.7904/0.9505.
T2 A0.0	0.7119/0.7876/0.9507.
T2 A0.1	0.7116/0.7876/0.9501.
T2 A0.2	0.7129/0.7895/0.9503.
T2 A0.5	0.7095/0.7863/0.9504.
T3 A0.1	0.7023/0.7820/0.9500.
T4 A0.2	0.7072/0.7847/0.9500.
T6 A0.8	0.7045/0.7800/0.9494.
T10 A0.2	0.6981/0.7747/0.9488.
Baseline481	0.6927/0.7639/0.9483.

pixelwise_knowledge_distillation_pspnet18's Issues

Consult some details

Hi author,

Thanks for your intetesting work. It is very useful for me. However, I want to consult some details for it. For example, in you readme, 'T1.8 A0.2' means temperature=1.8 and alpha=0.2 in config? And mIoU/mAcc/aAcc=0.7130/0.7904/0.9505 is based on students PSPNet18 teacher PSPNet50 ? Another more, when I train student PSPNet18 with teacher PSPNet50, it has more GPU occupy than only PSPNet18, it may be close to even more GPU occupy with only PSPNet50. It is reasonable ?

Thank you,
Best regards.