The BAS-loss: RuntimeError: CUDA error: device-side assert triggered

[EudicL]

When I change to single category training, the following error is reported:
RuntimeError: CUDA error: device-side assert triggered

[qmm1234]

Hello, have you encountered a situation during training where the loss doesn't converge, and both semantic loss and sb loss remain consistently at zero? If not, could you provide details on how you set the parameters during training?