mapillary / seamseg

Seamless Scene Segmentation
BSD 3-Clause "New" or "Revised" License
292 stars 51 forks source link

training with resolution 1024 #28

Open XAVILLA opened 3 years ago

XAVILLA commented 3 years ago

I don't have access to enough computation resource so I'm training on 8gpus with resolution 1024 on cityscapes, and I find it really hard to recover the result that's even nearly compatible to the ones in the paper. Would really appreciate if you can let me know if this is caused by the resolution? And what should I modify in configuration for resolution 1024 to work reasonably well?

XAVILLA commented 3 years ago

I keep getting validation pq around 35-40, is that normal?