SHI-Labs/OneFormer

Different result in different GPU

Opened this issue · 1 comments

First I create a predictor in the following way to get an output:
cfg.MODEL.DEVICE = "cuda:0"
predictor = DefaultPredictor(cfg)
outputs = predictor(image, "panoptic")

Then I want to use the model in second GPU and get the output, I also create it in the following way.
cfg.MODEL.DEVICE = "cuda:1"
predictor = DefaultPredictor(cfg)
outputs = predictor(image, "panoptic")

However, when I compare both of the outputs, the result is totally difference. The output using cuda 0 do the segmentation task very well. But the outputs using cuda 1 cannot recognize any object. I don't know why

Did you figure it out?