search

fudan-zvg / SETR

[CVPR 2021] Rethinking Semantic Segmentation from a Sequence-to-Sequence Perspective with Transformers
MIT License
1.05k stars 149 forks source link

Cant achieve the best miou when batchsize=4 #25

Closed cocolord closed 3 years ago

cocolord commented 3 years ago

I cannt achieve the most miou proposed in original paper while i didnt change the hyperparameter except tuning the bs from 8 to 4. How to achieve the best miou? Is that related to batchsize?

lzrobots commented 3 years ago

Yes you need batch size 8. How much did you get with bs=4?

cocolord commented 3 years ago

I tried Naive Deit and i have 77.64 miou after 80000iter.

cocolord commented 3 years ago

Basically 1% miou fall behond. But I dont have 8 cards, Is there any ways to train better result based on four cards?

lzrobots commented 3 years ago

not bad. you can increase the crop size

zbhit123 commented 3 years ago

I have four 2080Ti cards, 12G memory each Can i run this code ? tks !!!@cocolord

cocolord commented 3 years ago

I have four 2080Ti cards, 12G memory each Can i run this code ? tks !!!@cocolord

you can infer with the small vit or deit model