Segmentation - Incorrect IoU being printed on master process

[siddharth9820]

vision/references/segmentation/utils.py

Line 91 in 96950a5

reduce_across_processes(self.mat)

The expected behavior of this line is to all-reduce self.mat in-place. However, the reduce_across_all_processes method creates a copy of the tensor (

vision/references/segmentation/utils.py

Line 297 in 96950a5

t = torch.tensor(val, device="cuda")

) and runs the allreduce on that copy. Thus self.mat is not updated and the master process prints incorrect IoUs.

I guess this is a potential fix -

self.mat = reduce_across_processes(self.mat).to(torch.int64)

PR - #7916 attempts to fix this.

[NicolasHug]

Fixed in #7916, thanks for the report and for the fix @siddharth9820