RuntimeError: cuda runtime error (98) : invalid device function at atss_core/csrc/cuda/SigmoidFocalLoss_cuda.cu:139

[hust-kevin]

when I train the model, I meet this error.

PS:
pytorch 1.3, torchvision 0.4 cuda 10.0

[sfzhang15]

@hust-kevin
You can try to remove the build folder and then build the code again.

[hust-kevin]

thanks, I solve this issue by re-build
another question, in your paper said that retinanet ap is 36.3, however in paper and detectron, the ap of retinanet is 35.7

[sfzhang15]

@hust-kevin
Please refer to this.

[hust-kevin]

thanks for reply
in table 7. the ap 39.3 actually is come from retinanet + imprs + atss.
did you do experiment about only apply atss on retinanet ? what's the ap of retinanet + atss

[sfzhang15]

@hust-kevin
Sorry about that. We do not do experiment about only applying atss on retinanet. You can try to do this experiment, we believe it will bring some improvement.

[laisimiao]

@hust-kevin I also have this issue, but can't be solved by re-build?