search

uber-research / UPSNet

UPSNet: A Unified Panoptic Segmentation Network
Other
649 stars 120 forks source link

Reproducing results #69

Closed gaussiangit closed 5 years ago

gaussiangit commented 5 years ago

I trained coco for more than 160000 iterations, still get PQ around 35 with given config. Training for longer does help a bit but I see that your pretrained weights achieved results with 98000 iterations. Do you know what could be wrong ?

YuwenXiong commented 5 years ago

If you only use 4 gpus (i.e batch size =4) the expected #iter is 360k or longer, the 90k one is trained with 16gpus

gaussiangit commented 5 years ago

Got it. Thank you :)