search

microsoft / CSWin-Transformer

CSWin Transformer: A General Vision Transformer Backbone with Cross-Shaped, CVPR 2022
MIT License
539 stars 78 forks source link

how to set param drop-path for base and large model? #4

Closed zimenglan-sysu-512 closed 3 years ago

zimenglan-sysu-512 commented 3 years ago

hi @LightDXY how to set param drop-path for base and large model? and also when finetune on 384x384 input size?

LightDXY commented 3 years ago

Hi, we set the drop path rate=0.5 for the base model and we only train the large model on the IamgeNet22k.

When finetuning on 384x384 images, a drop path rate slightly larger than the pretraining setting should be better.

zimenglan-sysu-512 commented 3 years ago

thanks