compile error with pytorch1.0.0 nightly about fcos

[tianzhi0549]

@johnlanbor If you use exactly the same environment as maskrcnn-benchmark for FCOS, can you run it normally? We did not change any codes in these c++ files, and therefore I think it should work well in the environment where maskrcnn-benchmark can work.

[qianwangn]

@tianzhi0549 I compile maskrcnn benchmark and FCOS in the same environment, but FCOS cant compile correctly while benchmark can .

when i copy FCOS code to benchmark, and train with this setting: fcos_R_50_FPN_1x.yaml
when test woth 65000 pth can got AP:36
but I got loss(nan) when 79800 iter, is that because the lr? I use 4Gpus , so the lr should be half of 0.01?

[tianzhi0549]

@johnlanbor You don't need to change any configs because of 4 GPUs. The number of GPUs does not change the batch size.

The loss might become NAN occasionally. You can try to continue training from your last checkpoint.